# Practicum for "When Intelligence Became Cheap" — Book 2: Business in the Age of AI Agents

Everything the printed chapters of this book send you to, in one file: worksheets, prompts and
agent skills. Put it into ChatGPT Projects, a Claude project, NotebookLM or any assistant that
accepts an attachment, then work through it question by question.

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

Assembled 2026-09-07. Chapter-to-material contract updated 2026-08-12.
Canonical home: https://cheap-intelligence.vercel.app/en — each page below also lives there at its own address.

The living layer is deliberately **not** in this file. Tool maps, market pages and anything else
that carries a check date changes faster than a download: read it at https://cheap-intelligence.vercel.app/en/watch.

Attribute material to Denis Ermilov and the Practicum, with the page address and its check date.
Texts are CC BY-NC-SA 4.0; code and browser tools are MIT.

---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/tom2-workbook -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

---
reading_level: b2-c1-technical
title: "Book 2 Workbook: Business in the Age of AI Agents"
description: A route for managers and business owners, from an honest maturity check to a governed agent team and a change compass.
---

# Book 2 Workbook: Business in the Age of AI Agents

> The end-to-end Practicum route for *Business in the Age of AI Agents* (Book 2 in the
> *When Intelligence Became Cheap* series).
> This route through the contents is for **managers and business owners**. It starts with an honest
> maturity check and moves toward an agent-based team, a market where an agent buys, and a change
> compass. Professionals can use the [Volume 1 Workbook](tom1-workbook.md).
>
> **Your business functions are the axis of this route.** Volume 2 is organized around company
> functions such as sales, support, analytics, HR, and operations. A manager examines functions, not
> one profession. Browse [business cases](../watch/business-cases.md) for examples by function.

---

## The minimum agent governance package

If you do not have time for the full volume, build at least this package for **one** process:

1. [Process scoring](process-scoring.md) to choose a place where an agent can pay off.
2. [Human Review Matrix](human-review-matrix.md) to decide what AI can do alone, what requires a
   review, and what must stay with a person.
3. [Agent Contract](agent-contract.md) to define the role, access, owner, escalation path, and stop threshold.
4. [Evaluation set](eval-set-builder.md) to test the agent against past cases with known outcomes.
5. **Shadow mode:** two weeks of real work with no right to reply to a customer or the market on its own.

This is the minimum governance chain. Without it, an "agent in the process" remains a demo rather
than a managed change.

---

## Teams and business (Chapters 1 to 6)

**Step 1: Rebuild, do not decorate (Chapter 1).** Be honest. Did you "bolt AI onto" the process, or
did you rebuild it?

- Artifact: [Process maturity check](ai-native-maturity-check.md).
- **My maturity level (1 to 6):** `____` · **where we only "bolted it on":** `__________________`

**Step 2: Move from output to outcome (Chapter 2).** Decide what the agent can do and what stays
with a person.

- Artifact: [Human Review Matrix](human-review-matrix.md).
- Live page: [Autonomy boundary](../watch/output-to-outcome-watch.md).
- **Task for evaluation:** `__________________` · **decision owner:** `____________`

**Step 3: A team with agents (Chapter 3).** Define the role, access, review, escalation, and owner.

- Artifact: [Agent Contract](agent-contract.md) + [team-brain starter kit](team-brain-starter.md).

**Step 4: A company of one (Chapter 4).** Put one complete function on agents under supervision.

- Prompt: [Reinvent the company from zero](../prompts/zero-human-company-reframe.md).
- Checklist: [Resilient automation](resilient-automation.md).
- Live page: [The zero-human company](../watch/zero-human-company.md).

**Step 5: An agent-based staff for an SMB (Chapter 5).** Map functions against financial priority
and choose where to start.

- Artifact: [SMB staff map](smb-staff-map.md) + [where to start checklist](smb-start-checklist.md).
- **Function with the fastest cash impact:** `__________________`

**Step 6: When the customer is not human (Chapter 6).** Can the buying agent find and understand you?

- Live pages: [Agent commerce protocols](../watch/agent-protocols.md) ·
  [agentic browsers](../watch/agentic-browsers.md).
- Artifact: [Can an agent see me? checklist](../skills/visible-to-agent/) +
  [AI search visibility worksheet](geo-ai-search-kit.md).
- **Where an agent cannot find or understand me:** `__________________`

## Beyond the person and the company (Chapters 7 to 10)

- **Chapter 7: Children:** [what to teach when answers become cheap](kids-ai-skills-guide.md) ·
  [family 3C audit](family-3c-audit.md) ·
  [AI tutor prompt](ai-tutor-training-mode.md) ·
  [AI skills and tools for children](../watch/kids-skills.md).
- **Chapter 8: The new economy:** [Your own ladder: three materials](own-ladder.md) ·
  [Who owns your ladder?](ladder-ownership-audit.md) ·
  [AI and the labor market](../watch/labor-market.md).
  - **The thinnest material in my ladder (work / relationships / reputation):** `__________________`
- **Chapter 9: Physical AI:** [Platform rent map](platform-rent-map.md) ·
  [Who owns the rails?](../watch/physical-ai-rails.md).
  - **My total platform rent as a share of revenue:** `____%`
- **Chapter 10: 2030 scenarios:** [Which world are you building?](which-world.md) ·
  [No-regret moves](no-regret-moves.md) · [AGI claim check](agi-claim-check.md) ·
  [Signals without fog](../watch/agi-signals.md).
  - **My weakest no-regret move:** `__________________`

## Final step (Chapter 11: The Change Compass)

**Use a compass, not a calendar: four arrows and seven gates.** Bring the conclusions above into one
route. Move when a gate is passed, not because a date arrived.

- **WHERE arrow:** [Process scoring](process-scoring.md), with five questions about where an AI agent can pay off.
- **Worked example:** [IKEA through the Change Compass](ikea-compass-example.md), a public case run through four arrows and seven gates.
- **Route:** [Change gates: Order of Moves](change-gates.md), the printable Volume 2 capstone.
- **PEOPLE arrow:** [Learning without a box-ticking course](training-in-the-loop.md).
- **By role:** [manager route](manager-gates-route.md) · [business route](business-gates-route.md).
- **My process for the first round:** `__________________`

---

## My one-page result

Bring the route conclusions into one table. This is a business profile you can take to a team meeting
or discuss with a partner.

| Question | My conclusion |
|---|---|
| 1. Maturity level + where we only "bolted it on" | `____` / `____________` |
| 2. Task for evaluation + decision owner | `____________` |
| 3. First agent: role, boundaries, owner | `____________` |
| 5. Function with the fastest cash impact | `____________` |
| 6. Where a customer agent cannot find me | `____________` |
| 8. The thinnest material in my ladder | `____________` |
| 1. Is this a rebuild or decoration? | `____` |
| 2. What do we give to the agent, and what stays with a person? | `____` |
| 3. First agent: role, access, review, escalation | `____` |
| 4. Which function can become a supervised "company of one"? | `____` |
| 5. Function with the fastest cash impact for an SMB | `____` |
| 6. Where can a buying agent not see us? | `____` |
| 7. What must people practice instead of replacing it with an answer? | `____` |
| 8. Who owns our value ladder? | `____` |
| 9. Platform rent as a share of revenue | `____` |
| 10. No-regret move for 2030 | `____` |
| 11. First change round: process, gate, owner | `____` |
| 10. Weakest no-regret move | `____________` |
| 11. First-round process + nearest gate | `____________` |

Use this prompt for a second opinion:

```text
Here is my business profile from the Volume 2 Workbook for When Intelligence Became Cheap. It covers
maturity, evaluation, the first agent, the function with the fastest cash impact, agent visibility,
the value ladder, rent, no-regret moves, and the first change round. Check it for honesty. Where do
the conclusions conflict? Where am I overestimating our readiness? Which gate am I trying to skip?
Do not suggest "put AI everywhere." Ask three questions that will help me improve the first round.
[paste the table]
```

See a completed route in the [small-business owner example](../cases/persona-vladelec-msp.md) and the
[support manager example](../cases/persona-rukovoditel-podderzhki.md).


---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/ai-native-maturity-check -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# Process maturity check: surface polish or rebuild?

> Practicum for **Volume 2, Chapter 1**, "AI Is Not a Plugin. It Is a Rebuild." Time: 35-45 minutes.
> Warning: choose **one process**, not the whole company. It could be a contract, an application, customer support, reporting,
> lead generation, purchasing, or hiring. This check is a mirror for a useful conversation, not a consulting score.
> Interactive version: [process maturity check](../skills/ai-native-maturity-check/).

AI creates value when it changes the process, not when a team bolts it onto the old way of working.
This worksheet helps you see where one process stands today and choose the next move.

---

## A. Choose a process

- **Process:** `________________________`
- **Where AI is already used:** `________________________`
- **Outcome that should change** (money, speed, risk, or customer result): `________________`
- **Where you can see money, speed, or risk today:** `________________________`

## B. Six signs of a real rebuild

Check what is **already true**, not what you plan to do. Each sign describes the process, not a slogan.

- [ ] **The work map has changed.** The steps are different rather than "the same work, faster."
- [ ] **The roles have changed.** People do different work than they did a year ago, with more framing and review.
- [ ] **The metrics have changed.** You track the result or cycle, not the amount of work produced.
- [ ] **Quality review is built in.** It is part of the flow, not something done "when there is time."
- [ ] **A person remains at the critical point.** A person decides when the cost of an error is high.
- [ ] **There is a measurable result.** It affects the customer or the money instead of ending at an "AI implementation."

Number of checked boxes: `___ / 6`.

## C. Level 0-5

Where is the process **now**, not where you want it to be?

| Level | What it means |
|---|---|
| 0 | AI is not used |
| 1 | Chaos: people experiment by instinct, without a system |
| 2 | Personal productivity: a few people work faster, but the process is unchanged |
| 3 | Team process: the team has agreed where and how to use AI, and review is in place |
| 4 | Operating model: the process has been rebuilt around AI, with new roles and metrics |
| 5 | AI-native: the old form of the process no longer exists, and the result is measurable |

My level for this process: `___`.

## D. Diagnosis

- **0-2 boxes / level 1-2: surface polish.** AI has been bolted on, so you do the old work faster. That is a normal starting point, but the source of value has not changed yet.
- **3-4 boxes / level 3: transition.** Roles and review have started to change. The question is whether you finish the rebuild or fall back to simple acceleration.
- **5-6 boxes / level 4-5: a system rebuild.** The process is genuinely different. Your next task is to scale it while keeping a person at the critical point.

My diagnosis: `________________`.

## E. One next move

Do not write a one-year plan. Choose **one** move that takes the process from surface polish toward a rebuild.

- What changes in the **work map:** `________________`
- Which **role** owns the result: `________________`
- Which **outcome metric** replaces volume: `________________`
- Where the **review** happens, and who reviews it instead of another AI: `________________`
- Where a person must be able to **stop** the process: `________________`

---

## Next

- If you lead the team or company, test the business with a stricter frame:
  [Reimagine the company from zero](../prompts/zero-human-company-reframe.md).
- If you are at level 2-3, decide what AI may do and what a person should keep:
  [Human Review Matrix](human-review-matrix.md) (Chapter 2).
- Follow the full route in the [Volume 2 Workbook](tom2-workbook.md).

> Warning: this is a diagnosis of one process, not a company rating. The score should prompt thought, not pass judgment.


---

<!-- Page: https://cheap-intelligence.vercel.app/en/prompts/zero-human-company-reframe -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# Prompt: rebuild a company from zero (Zero Human Company reframe)

> A working tool for **Volume 2, Chapter 1** ("AI is not a plugin. It is reconstruction") and
> **Chapter 4** ("The Company of One"). Chapter 1 ends with a question: are you attaching AI to an old
> process, or **changing the process itself**? This prompt takes the second option seriously. It asks
> you to view the company as **one cognitive function** rather than a hierarchy of people, then rebuild
> it for cheap intelligence. **Date: 2026-06.**
>
> **An honest frame, without hype.** "Zero Human Company" is a fashionable label. The goal here is not
> a company without people. The goal is to see what part of the business is *intermediation* that may
> disappear, and what remains human: distribution, trust, and accountability. You remain the **owner
> of the outcome**. This is a thinking tool, not a plan for firing a team. Check the AI's answers and
> make the decisions yourself.
>
> For the general method, see [Prompts: method and review](index.md).

## How to use it

1. Open a capable model from this [list](../start/01-vybrat-neyroset.md).
2. Paste the full prompt and complete the "My company context" section.
3. The model will ask five follow-up questions first. Answer them honestly and avoid general phrases.
4. Treat the analysis as a critic's opinion, not a verdict. Check the reasoning and argue with it.

---

```text
# ROLE
You are the founder's strategy partner and critic.
You are not a McKinsey consultant, motivational coach, or AI evangelist.
Guide me through rebuilding the company with the Zero Human Company principle:
the company is one cognitive function, not a hierarchy of 200 biological brains.

# MY COMPANY CONTEXT
- What we do: [describe it in 2-3 sentences]
- Team size: [people]
- Revenue / stage: [RUB millions, growth, profit]
- Main product or service: [what]
- Our customer: [segment, industry]
- What currently acts as "intermediation" between parties: [answer honestly]
- Where the pain is greatest now: [symptom, not cause]

# WHAT I WANT FROM YOU
1. IDENTIFY MY COMPANY'S COGNITIVE FUNCTION.
   What do we really do after removing the people, departments, and procedures from the description?
   Write one plain sentence without grand language.
2. FIND THE INTERMEDIATION.
   Where does my business model make money by connecting two people, systems, or processes?
   Name it directly. This is the part that may disappear within 2-5 years.
3. REBUILD THE COMPANY FROM ZERO.
   Do not digitize the current setup. Rebuild it.
   Give me 5-7 basic prompts that could form the company's system foundation if we started today.
4. DESIGN THE AGENT ARCHITECTURE.
   Do not "automate Maria's and Peter's tasks."
   Explain which agents live inside the cognitive function, how they replicate x2,
   and which 3 parallel scenarios I can start tomorrow.
5. MAP DECLINE AND GROWTH.
   Which current roles and departments will no longer be needed in 2 years?
   Which new roles will appear? Who on the team is most likely to succeed in the new architecture,
   even if that person is not the obvious choice today?
6. MY ONE-MONTH WINDOW.
   What opportunity should I act on in my market right now?
   Look for a window of 4-8 weeks, not a year.
7. 60 MINUTES TODAY.
   What can I do in one hour without leaving my desk to start the process?
   Be specific. Do not tell me to "think about it."

# RULES
- No filler. Do not use 5C, SWOT, McKinsey frameworks, "key drivers," or "synergies."
- If my context is weak, ask questions instead of guessing.
- If my answer is vague, say: "This says nothing. Rewrite it."
- If my business is intermediation, do not soften the answer. Say: "You have 2 years."
- Do not flatter me. I pay for criticism, not praise.
- Think like a founder with 100 agents on the team, not a corporate analyst.

# START
Before you answer, repeat your understanding of my business inputs
and ask 5 follow-up questions that you need for a useful analysis.
```

---

## After the analysis, run three checks from the book

- **Intermediation is not distribution.** AI weakens intermediation when a business only connects A
  and B. Access to customers, trust, and accountability for the outcome become more valuable. Do not
  cut your customer relationship when you mean to cut intermediation. See Volume 2, Chapters 1 and 6.
- **Make the function autonomous, not the whole company.** Chapter 4 gives a more honest boundary. A
  function can run on agents under supervision. An entire company cannot safely run on autopilot.
  Where will you keep a person at the critical point?
- **Outcome, not output.** The new architecture should change decision speed, risk, or money rather
  than the amount produced. If the metric stays the same, the redesign is still cosmetic.

> The [What to Watch](../watch/skills-now.md) section covers the "Zero Human Company" trend, platforms,
> and recent cases that age quickly. This page keeps only the thinking tool that should last longer.


---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/human-review-matrix -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# What AI should do and what people should keep (Human Review Matrix)

> Practicum for **Volume 2, Chapter 2**, "More Output Does Not Mean More Value."
> Time: 15-20 minutes. One process on one screen, not a 10-page AI policy.

An outcome does not come from producing more output. It comes from a process with the right review and the right metric.
This worksheet turns the chapter's main advice into an exercise: take one process and decide what the machine may do,
what needs review, and what a person should keep.

**How to fill it in.** List 5-7 tasks from **one** process as rows. Work through the columns for each task.
Record the decision in the "Mode" column.

A minimum governance packet is the small set of documents that defines control and ownership. This matrix is the second worksheet. First choose a process with
[process scoring](process-scoring.md). Use this matrix to assign a mode to each task. Then define the agent with an
[Agent Contract](agent-contract.md) and test it on an [evaluation set](eval-set-builder.md).

| Task | Cost of error | Reversible? | Can it be described as a process? | Can you check it? | Mode | Control layer | Flow metric |
|---|---|---|---|---|---|---|---|
| `________` | low / medium / high | yes / no | yes / no | yes / no | AI alone / review required / person only |  |  |
| `________` | | | | | | | |
| `________` | | | | | | | |
| `________` | | | | | | | |
| `________` | | | | | | | |

**What the columns mean:**

- **Cost of error:** what happens if AI gets it wrong and nobody notices.
- **Reversible:** whether you can undo the action if the wrong result goes out.
- **Can it be described as a process?** This is the chapter's filter. If the answer is no, the task is not ready for a machine.
- **Can you check it?** Whether you have a quick way to tell if the result is right, such as a reference answer, rule, or eval.
- **Mode:** the final decision for the task.
- **Control layer** for "review required": visibility, guardrails, access limits, or a person at the critical point.
- **Flow metric:** one end-to-end measure instead of output volume, such as time to result or the share completed without rework.

## Default logic

- High cost of error plus poor reversibility means **person only**, or a person at the critical point.
- If you cannot describe the task as a process, **do not hand it over yet**. Understand it first.
- If you cannot check the result, **build the check first**. Otherwise, you have a demo rather than a result.

## Output

A completed matrix for one process and **one** end-to-end metric that you track instead of volume.
Return in three months and see what has moved from "person only" to "review required."

---

## Next

- For tasks marked "AI alone" or "review required," define the agent with an
  [Agent Contract](agent-contract.md) (Chapter 3).
- Do not send the first version straight to production. Build an [evaluation set](eval-set-builder.md) and run it in shadow mode, where it drafts or observes without taking live action.
- Reviewing output is a scarce skill. Train it with [Skills Watch](../watch/skills-that-pay.md) (Volume 1, Chapter 7).
- Follow the full route in the [Volume 2 Workbook](tom2-workbook.md).

> The chapter covers flow metrics. This worksheet decides who does what. ⚠️ Check the output outside the model.


---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/agent-contract -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# Agent Contract

> Practicum for **Volume 2, Chapter 3**, "A Team with Agents." Time: 15-20 minutes per agent.
> Use one sheet for one agent. Copy the template and fill it in for your case.

An agent is not a "magic employee." It is a role with clear boundaries, review, and a **human owner**.
The contract turns "let's add AI" into an agreement the team can understand and enforce.

A minimum governance packet is the small set of documents that defines control and ownership. In that packet, the contract follows the [Human Review Matrix](human-review-matrix.md).
The matrix decides which tasks can be handed over. The contract defines **one** agent, its permissions, and its brakes.
Do not scale right away. Test the agent on an [evaluation set](eval-set-builder.md) and in shadow mode first, where it drafts or observes without taking live action.

## Template: 10 fields

| Field | What to enter |
|---|---|
| **Role** | the agent's name and function |
| **Mission** | one sentence explaining why it exists |
| **Inputs** | the data and formats it accepts |
| **Tools** | the systems, APIs, and files it may access |
| **Memory** | what it remembers between sessions, and what must **never** enter memory |
| **Boundaries** | what the agent **never** does without approval |
| **Escalation** | the conditions that send the task to a person |
| **Review** | how and when a person checks the result, rather than one AI checking another |
| **Metrics** | 1-2 numbers that show whether it works |
| **Owner** | the name of the person accountable for it |

> Default rule: without **Boundaries + Escalation + Owner**, this is a wish, not a contract. All three are required from day one.

### A measurable threshold gate: KPI -> escalation -> stop

A contract metric is not something to report at the end. It is a **guardrail**, a preset threshold that makes the agent stop and call a person.
Connect three fields in one rule:

| What you define | Example |
|---|---|
| **Signal metric** | share of answers accepted without edits; unusual numbers; escalation rate; daily cost |
| **Threshold**, set in advance | "below 85% accepted without edits"; "difference > 10%"; "> N requests in a row"; "above the ₽/day limit" |
| **Action at the threshold** | pause and human review; return to manual work; owner-triggered shutdown |

> Give the agent the smallest set of permissions it needs. This is called **least privilege**. A measurable threshold turns Metrics and Escalation into a working brake. Set the threshold **before** launch, not after an incident.

## Example 1: first-line support agent

- **Role:** sorts incoming requests and answers common questions. **Mission:** remove routine work from the support team.
  **Inputs:** request text and answer base. **Tools:** read-only access to the knowledge base.
- **Boundaries:** does not promise compensation or touch billing. **Escalation:** complaints, refunds, or conflict go to a person.
  **Review:** a person audits 10 conversations a day. **Metrics:** share resolved without a person and number of escalations.
  **Owner:** head of support.

## Example 2: reporting analyst agent

- **Role:** prepares the weekly sales summary. **Inputs:** CRM exports. **Tools:** read-only data access.
  **Boundaries:** does not send the report. **Escalation:** an anomaly > X% is flagged for a person.
  **Review:** the owner checks the totals against the source before sending. **Metrics:** time to finished report and share of figures accepted without edits.
  **Owner:** team analyst.

## Pre-flight check

- [ ] The three required fields are complete: Boundaries, Escalation, and Owner.
- [ ] Access follows **least privilege** and covers only what the mission requires.
- [ ] Memory contains no secrets or personal data that do not belong there.
- [ ] A human review method is built into the flow.
- [ ] The team has agreed on the signal that **shuts the agent down**.
- [ ] A **measurable threshold gate** is set: metric + number + action (pause / rollback / stop).

---

## Next

- Build shared context for the team and its agents with the [team-brain starter kit](team-brain-starter.md).
- Decide what the agent may do and what a person should keep with the [Human Review Matrix](human-review-matrix.md) (Chapter 2).
- Review third-party skills before they enter your system: [how to check them](../watch/skill-banks.md).
- Follow the full route in the [Volume 2 Workbook](tom2-workbook.md).


---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/team-brain-starter -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# Team-brain starter kit: a shared team brain in 14 days

> Practicum for **Volume 2, Chapter 3**, "A Team with Agents." For a manager or founder.
> This follows the [Agent Contract](agent-contract.md). Once agents have boundaries, give them shared context. Without it, each agent fails in its own way.

Agents and new team members can only work with the context they receive. A "team-brain" is a small shared knowledge repository.
It is not a year-long corporate wiki project. It is a handful of files that an agent reads first.

## Minimum structure

```
team-brain/
  product/        - what we make, who it is for, and the main metrics
  customers/      - who the customer is and what we know about them
  decisions/      - decision log (decisions.md)
  glossary/       - team terms (the agent reads these first)
  skills/         - reusable instructions and prompts
```

You do not need more folders. Five active files beat thirty empty ones.

## Decision log (`decisions/decisions.md`)

Use one line per decision so agents and new colleagues do not have to ask why it was made:

```
Date | Decision | Context (why) | Alternatives rejected | Owner | Review date
```

## Memory: three states

New knowledge moves through three states so the shared brain does not fill with junk:

- **staged:** a candidate has been recorded but is not yet shared, like a memory draft;
- **review:** a person has checked that it is true and reusable;
- **prompted:** it is now part of the context that the agent actually reads.

Nothing moves from *staged* to *prompted* without a person.

## Agent evaluation template

Before you trust an agent with a task, run it **with and without the context** and compare the results.

| Task | Inputs | What counts as a good answer | Score without team-brain | Score with team-brain | Decision |
|---|---|---|---|---|---|
| `____` | `____` | `____` | `_/5` | `_/5` | `____` |

If there is no difference, the context is not working, or the task does not depend on context.

## 14-day plan

- **Days 1-3.** Find one bottleneck where the team loses time by explaining the same thing again.
- **Days 4-7.** Collect the minimum knowledge for it in `product/`, `customers/`, and the first entries in `decisions/`.
- **Days 8-14.** Launch **one** agent with an [Agent Contract](agent-contract.md). Run the evaluation with and without context. Keep it only if the difference is measurable.

---

## Next

- Give each agent its own [Agent Contract](agent-contract.md) with a role, boundaries, and owner.
- Decide what the agent may do and what a person should keep with the [Human Review Matrix](human-review-matrix.md).
- Follow the full route in the [Volume 2 Workbook](tom2-workbook.md).

> The goal is a few files that become useful quickly, not a corporate wiki project.


---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/resilient-automation -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# Resilient automation checklist

Worksheet for Volume 2, Chapter 4, "The One-Person Company."

Use this checklist to test one agent-run function before it touches customers, money, documents, or reputation.
This is not a guide to automating everything. It tells you what must be visible, testable, and easy to stop.

## 1. Function

Choose one function, not the whole company.

| Question | Your answer |
|---|---|
| What repeatable result should this function produce? |  |
| What goes in: emails, applications, documents, deals, or data? |  |
| What comes out: an answer, record, invoice, report, or recommendation? |  |
| Where does the cost of an error become high? |  |
| What must the agent never do without a person? |  |

If you cannot describe the output in one sentence, it is too early to automate the function.

## 2. Tests

Collect at least ten test cases.

| Type of case | What to include |
|---|---|
| Normal case | Work the function handles every day |
| Difficult customer | Large order, unusual promise, or sensitive tone |
| Legal risk | Contract, invoice, personal data, or disputed wording |
| Bad input | Missing data, contradiction, or an attachment in the wrong format |
| Red zone | A case where the agent must stop |

A test works only when you decide in advance which answers are acceptable.

## 3. Logs

The agent must leave a record of its actions.

- What the agent received.
- Which sources or tools it used.
- What it decided on its own.
- Where it was uncertain.
- Which rules fired.
- Why it stopped or continued.

If you cannot reconstruct the path after an error, you do not have a function. You have a black box.

## 4. Warning signals

Write down what should raise a flag.

| Signal | Action |
|---|---|
| A new promise is made to a customer | Stop and show it to a person |
| A price, discount, or deadline changes | Stop and show it to a person |
| A document is about to leave the company | Require human review |
| The agent is uncertain or finds a contradiction | Escalate instead of guessing |
| The same error happens again | Stop the flow and repair the system |

A warning signal without an owner is useless. Name the person who will see each signal.

## 5. Permissions

Give the agent the least access it needs.

| Access question | Decision |
|---|---|
| Is access read-only, or may the agent write? |  |
| May the agent send a message to a customer without approval? |  |
| May it change the price, deadline, or deal status? |  |
| Is there a money or risk limit? |  |
| When does access turn off automatically? |  |

The chapter's rule is simple: give the agent exactly the access the task requires, and nothing more.

## 6. Repair

Decide what happens after a failure before one occurs.

1. Who sees the incident?
2. Who stops the flow?
3. Who repairs the rule, prompt, integration, or data?
4. Who tells the customer if the error reached them?
5. How does the new case enter the test set?

Without repair, automation builds up debt. With repair, it becomes a system.

## Final decision

| Check | Yes / no |
|---|---|
| The function is described separately from the company |  |
| Test cases exist |  |
| Actions are logged |  |
| Warning signals exist |  |
| Permissions are minimal |  |
| The person who repairs a failure is known |  |
| A person remains at the critical points |  |

If three or more answers are "no," the function is not ready for live work. Start with one pain point, one agent, and one stop rule.


---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/smb-staff-map -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# Small-business agent staff map

> Practicum for **Volume 2, Chapter 5**, "Small Business: AI as an Affordable Staff." Time: 20-30 minutes.
> For a small or midsize business owner. The result is **not** "automate everything." It is 1-2 cells where you can start.

AI can give a small business the kind of support staff it could not afford before. Build that staff around **customer pain and money**, not around a tool.
This worksheet helps you set priorities.

## Step 1. Break the business into functions

Check the functions that exist in your business:

- [ ] Sales / inbound leads
- [ ] Customer support
- [ ] Marketing and content
- [ ] Documents / contracts
- [ ] Analytics and reporting
- [ ] Operations / logistics
- [ ] Finance / payments

## Step 2. Score each function on four dimensions

Give each function a score from 0 to 3. This is a rough filter for an owner, not an academic model. Focus on money, costs, and the customer.

- **New revenue:** the function helps you sell more, sell something new, or move a lead to payment faster.
- **Costs:** the function reduces manual work, errors, waiting, or contractor costs.
- **Customer / trust:** the function improves response time, the quality of the conversation, repeat business, or direct contact.
- **Repeatability / testability:** the work repeats, can be described, checked, and handed to a person in time.

| Function | New revenue (0-3) | Costs (0-3) | Customer / trust (0-3) | Repeatability / testability (0-3) | Total |
|---|---:|---:|---:|---:|---:|
| `____________` | | | | | |
| `____________` | | | | | |
| `____________` | | | | | |
| `____________` | | | | | |

## Step 3. Choose 1-2 starting cells

- **Total 8-12 + repeatability 2-3:** the first candidate. An agent staff can produce a result here fastest.
- **High new revenue / customer score, but repeatability 0-1:** use a person with an AI assistant, not autopilot.
- **High repeatability, but low revenue / customer score:** it can wait, whatever vendors promise.
- **High cost score + high repeatability:** a good training entry point, especially if you have outsourced this work before.

**My starting function:** `________________`  |  **Why this one:** `________________`

> This directly applies two ideas from the chapter: the three levers of value and "start with the hole, not the drill."
> This is not a tool list. Find services for your chosen function in the [Tool Map](../watch/tools-map.md).

---

## Next

- Once you choose a function, use the [safe-start checklist](smb-start-checklist.md).
- To run one function with supervised agents, use the [Resilient Automation Checklist](resilient-automation.md) (Chapter 4).
- Follow the full route in the [Volume 2 Workbook](tom2-workbook.md).

> Warning: start with one cell, not a plan to "digitize the whole business."


---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/smb-start-checklist -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# Safe-start checklist: an AI staff without unnecessary risk

> Practicum for **Volume 2, Chapter 5**, "Small Business: AI as an Affordable Staff." Time: 10-15 minutes.
> This follows the [Small-Business Agent Staff Map](smb-staff-map.md). Once you choose one function, use this worksheet to start without breaking the business.

The entry point stays cheap and reversible when you begin with **one pain point tied to money** and keep a person in the loop.
Use these five steps.

## Step 1. Name one pain point tied to money

Do not write "implement AI." Name a specific problem that costs money or time every week.

- **Pain point:** `________________`  |  **Weekly cost** in hours or money: `________________`

## Step 2. Check whether you have outsourced it before

Work you have already given to a contractor is often a good candidate. You have described it as a service, so it is easier to hand to an agent.

- Outsourced before? `yes / no`  |  If yes, **what exactly:** `________________`

## Step 3. Choose assistant or autopilot

- **Assistant:** AI prepares the work; a person decides and sends it. Use this whenever an error has a meaningful cost.
- **Autopilot:** AI acts on its own. Use it only for cheap, reversible, testable routine work.
- **My choice for this function:** `________________`

## Step 4. Set the control points

Before launch, write down what the agent **does not do** without you.

- Does not touch: `________________` (money / contracts / customer messages, for example)
- Escalates to a person when: `________________`
- How I check the result: `________________`

## Step 5. Set a time limit

Do not let "setting up AI" consume a month.

- **Time allowed for the test:** `____` (for example, 2 weeks)
- **Signal that it works, or that I should stop:** `________________`

---

## Output

One function, one mode (assistant or autopilot), clear control points, and a test period. This is not a "digitized business." It is an honest first step.

## Next

- Set function priorities with the [Small-Business Agent Staff Map](smb-staff-map.md).
- Add tests, logs, and warning signals with the [Resilient Automation Checklist](resilient-automation.md).
- Follow the full route in the [Volume 2 Workbook](tom2-workbook.md).

> Warning: one pain point, one control loop, and an honest deadline. Do not try to automate everything.


---

<!-- Page: https://cheap-intelligence.vercel.app/en/skills/visible-to-agent/index -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# Skill: can an agent see me

> A zero-install tool for **Volume 2, Chapter 6**, "When the Customer Is Not Human." It runs in your
> browser, calculates on your device, and sends no data anywhere. Code license: MIT.
> This is a manual visibility check for an agent buyer.

When an agent chooses and pays, louder advertising matters less than **being clear to a machine**. This
checklist tests three levers: **findability, evaluability, and actionability**, then gives you one move.

<a class="md-button md-button--primary" href="tool.html">Open the tool</a>

## What it does

- You answer short yes or no questions about the three levers: can an agent find you, understand you
  quickly, and complete the purchase without negotiation?
- The tool scores each lever, marks the **weakest one**, and suggests one next move.
- It returns a result that you can copy and give to your AI assistant.
- After scoring, a **Report** section explains all three levers and gives three next steps, starting
  with the weakest. You can copy or print the report for your team or manager.

## How to use it

1. Open the tool and mark what you already have.
2. Copy the result.
3. Tell AI: "Here is my findability, evaluability, and actionability analysis. Suggest one move for the
   weakest lever and what I should check." Ask for options, not a finished decision.

## Discipline

This is a map, not a verdict. Visibility to agents is a new form of distribution, but you remain
responsible for the transaction. For more depth, see the
[AI Search Visibility Kit](../../playbooks/geo-ai-search-kit.md).

## Chat version if you cannot use the browser tool

```text
Help me assess whether an agent buyer can see and use my product or service, based on *When
Intelligence Became Cheap*, Volume 2, Chapter 6. Analyze three levers:
- Findability: can the agent find me where AI systems get answers?
- Evaluability: can it quickly understand my offer from a clear description, price, and examples?
- Actionability: can it complete the next step without negotiation through a clear action, contact,
  or payment method?

Here is my product and how it appears now: [describe it].
Name the weakest lever and one specific move for the next two weeks. Tell me what to verify so we do
not invent an answer. Avoid hype.
```

---

Source worksheet: [AI Search Visibility Kit](../../playbooks/geo-ai-search-kit.md). Other tools:
[Prompts and Skills](../).


---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/geo-ai-search-kit -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# AI search visibility kit: how AI finds and cites you

> Practicum for **Volume 2, Chapter 6**, "When the Customer Is Not Human." Time: 30 to 40 minutes.
> This is a manual worksheet for a first measurement. The details change quickly, so check the date.

Search is turning into an **answer**. A person may ask ChatGPT or another assistant what to buy or whom to hire, then receive a shortlist without opening ten links. A high search rank or an SEO trick is not enough. You need to appear in sources the AI trusts and can cite. "We have a website" becomes "the same checkable facts about us appear across sources a machine can read."

> Gartner forecasts declining traffic from traditional search, with estimates of 25% less search volume by 2026 and 50% less organic search traffic by 2028. The KDD 2024 paper "GEO" by Aggarwal et al. found that source citations, quotations, and statistics produced the largest gains in visibility inside generative answers, about 30% to 40% for the best-performing methods. Treat these figures as reference points. Primary sources appear below.

---

## Step 1. Map AI answers

Write five to seven real questions your customers ask. Run each question through three or four AI systems, such as ChatGPT, Gemini, Perplexity, or a locally popular assistant. Record what appears in the answer:

| Customer question | Brands in the answer | Sources cited | Tone about you |
|---|---|---|---|
| `________` | | | positive / neutral / negative / absent |
| `________` | | | |
| `________` | | | |

> If you do not appear, the answer has left you out. If you appear in a weak or negative context, trust can fall before the customer reaches your site.

## Step 2. Check the three conditions for being chosen

The chapter gives you three levers. Mark the weak ones:

- [ ] **Findability:** Can the agent find you in the places it uses to build an answer?
- [ ] **Evaluability:** Can the agent quickly understand your value from a clear description, price, and examples? The technical term is *token-to-value*: how much useful information the agent gets from a small amount of text.
- [ ] **Actionability:** Can the agent take the next step without a negotiation, using a clear contact, checkout, or payment path?

Weakest lever: `________________`. Start there.

## Step 3. Measure AI search visibility

Do not track only a Google position. Track your presence inside answers:

| Metric | What it means |
|---|---|
| Share of AI mentions | how often you appear in answers to your test questions |
| Position / order | whether the answer names you first or near the end |
| Sentiment | whether the description is positive, neutral, or negative |
| Source share | which sources the system cites about you and whether it trusts them |
| Co-mentions | which competitors or categories appear next to you |
| NAP consistency | whether your name, address, and contact details match across sources |

> **Remember: API does not equal UI.** A model response from an API is not the same as the answer a person sees in a product interface. Personalization, locale, and device can change the result. AI search measurement is probabilistic, not exact. Run each test more than once, then record the date and model.

## Step 4. Find the sources that matter in your region

You earn citations in places the AI trusts. Your own site may not be the strongest source. Industry publications, professional communities, marketplace listings, local directories, and review platforms may matter more. Build a current list for your market. Your goal is to place a checkable fact or quotation there, not an advertisement.

## Step 5. Make a 14-day plan

- **Days 1 to 3.** Map AI answers from Step 1. Find the questions where you are absent or described poorly.
- **Days 4 to 9.** Fix **one** weak lever from Step 2. Add a machine-readable description, price, or example. Prepare two or three specific facts, case results, or quotations for the sources from Step 4.
- **Days 10 to 14.** Publish that material where AI systems get answers. Repeat the map and compare your share of mentions and sentiment. Record the date and model.

Output: one repaired lever and a baseline for another measurement next quarter.

---

## Next

- For a quick self-check, use [Can an agent see me?](../skills/visible-to-agent/).
- For agent payments and protocols, see [agent commerce protocols](../watch/agent-protocols.md). For the software a customer agent uses, see [agentic browsers](../watch/agentic-browsers.md).
- For the full route, open the [Volume 2 workbook](tom2-workbook.md).

## Sources (verified 2026-07)

- **Aggarwal et al., "GEO: Generative Engine Optimization"** (KDD 2024): methods that increase citations in generative answers. <https://arxiv.org/abs/2311.09735>
- **Gartner:** forecasts for declining traditional search traffic. <https://www.gartner.com/en/newsroom>

> ⚠️ AI search measurement remains probabilistic because API results do not equal UI results.


---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/kids-ai-skills-guide -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# What to teach when answers get cheap

> Practicum for Volume 2, Chapter 7, "Children: What to Teach and Where to Guide Them."
> Time: 10-15 minutes. This is not a list of future careers. It is a map of abilities worth developing at home over many years.

The chapter's main question is not "What job will my child have?" It is "What will grow stronger in my child while machines make ready-made answers cheaper?"
Use this worksheet as a family filter. Decide what to practice without AI, where AI can act as a training partner, and where it has started to remove the effort.

---

## 1. Five abilities that become more valuable

| What becomes more valuable | What it means at home | How to use AI without losing the skill |
|---|---|---|
| **Questions** | The child notices what is missing, what does not fit, and what nobody has asked. | Ask AI to hold back the answer and ask questions about the child's plan first. |
| **Taste and judgment** | The child can tell strong work from weak work and explain why. | Compare two options, one made by the child and one by the machine. Let the child choose and defend the choice. |
| **Agency** | The child sets a goal, starts without permission, and changes the plan after a setback. | AI may split a project into steps, but the child keeps the goal, choices, and responsibility. |
| **Human connection** | The child listens to another person, handles disagreement, and asks for help. | Use AI to rehearse a conversation, but never as a replacement for the conversation. |
| **Learning how to learn** | The child completes the cycle: understand, apply, make a mistake, repair it, and try again. | AI may provide exercises, hints, and check questions, but it should not complete the cycle for the child. |

If an activity develops none of these abilities, ask what it really is. It may be healthy rest, or it may be another shortcut to finished work.

## 2. What you can safely give to AI

Give AI work that is not the main exercise for the skill.

| AI may do this | Do not hand this over completely |
|---|---|
| find options, examples, and sources | form the child's own question |
| explain a topic in different ways | make the first attempt to understand it |
| check a draft for errors | decide what counts as a good result |
| create practice problems | do the practice |
| plan project steps | choose the goal, make the bet, and accept responsibility |

The rule is simple. The mode works if AI helps the child think, choose, or check more skillfully.
If the child only brings back a finished answer faster, the practice has been skipped.

## 3. Family mini-audit

Choose one activity: school, a club, sports, a project, a game, or family work.

| Question | Answer |
|---|---|
| Where does the child ask their own questions? | |
| Where do they compare quality and explain a choice? | |
| Where do they have a real choice with consequences? | |
| Where is there a real person whose craft they can observe? | |
| Where do they complete the learning cycle instead of receiving hints only? | |

If more than two rows are blank, do not rush to find a fashionable new class. Add one useful step first:
a real project, a human mentor, a public record of the result, or a "hint, not answer" rule.

## 4. One move for this week

Choose one.

| Skill that needs work | Move for the week |
|---|---|
| Questions | At dinner three times this week, ask, "What surprised you today?" Do not answer for the child. |
| Taste and judgment | Review one film, text, object, piece of code, or dish together: what works, what does not, and how it could improve. |
| Agency | Give the child a small project with a real stake: someone outside the family will see the result. |
| Human connection | Plan one screen-free conversation without judgment, where the adult listens first. |
| Learning how to learn | Ask AI to create a practice session, but do the final check without AI. |

After a week, do not ask whether it went perfectly. Ask where the child did the work alone.

---

> Related: [Volume 2 Workbook](tom2-workbook.md) |
> [Family 3C Audit](family-3c-audit.md) |
> [AI Tutor in Training Mode](ai-tutor-training-mode.md) |
> [Children's AI Skills and Tools](../watch/kids-skills.md) |
> [Prompt Bank: Friction Instead of a Finished Answer](training-not-answers.md).


---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/family-3c-audit -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# Family 3C audit: where your child builds mastery

> Practicum for Volume 2, Chapter 7, "Children: What to Teach and Where to Guide Them."
> Based on Matt Beane's *The Skill Code*: challenge, complexity, connection.

This audit looks past grades and schedules. It checks the environment where your child learns. Do they face a real challenge, see the full complexity of the work, and spend time with a skilled person? Without these three conditions, AI can turn learning into the quick submission of a ready-made answer.

## 1. Choose one activity

Do not audit your child's whole life at once. Pick one activity: school, soccer, music, drawing, robotics, a family project, a part-time job, or helping with a business.

## 2. Challenge: is the task demanding in a healthy way?

| Question | Yes / no | What to change |
|---|---|---|
| Is the task slightly above your child's current level? |  |  |
| Can your child make and discuss a mistake without immediate punishment? |  |  |
| Does AI help with the next step without taking over the whole task? |  |  |

## 3. Complexity: does your child see the whole job?

| Question | Yes / no | What to change |
|---|---|---|
| Does your child understand who needs the result and why? |  |  |
| Do they see the path from an idea to a finished artifact, rather than one small piece? |  |  |
| Does the project leave a visible result, such as a page, album, performance, object, or review? |  |  |

## 4. Connection: is there a skilled person nearby?

| Question | Yes / no | What to change |
|---|---|---|
| Is there an adult whose way of thinking and working your child can observe? |  |  |
| Does this person give specific feedback rather than only a grade? |  |  |
| Does your child spend an hour next to real work, rather than only watching a video lesson or using a chat? |  |  |

## 5. Make one move this quarter

Find the emptiest column and add one step:

| What is missing | What to add |
|---|---|
| Challenge | Give your child a project with a real stake. Someone else will see the result. |
| Complexity | Ask your child to explain the full path: the goal, steps, mistakes, what AI did, and what they did. |
| Connection | Find a skilled person they can work beside: a coach, club leader, colleague, neighbor, or relative with a craft. |

Come back to this sheet in three months. Do not ask, "Did it get easier?" Ask, "Where did my child do the work?"

> Related: [guide to children's skills](kids-ai-skills-guide.md) and
> [AI tutor in training mode](ai-tutor-training-mode.md).


---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/ai-tutor-training-mode -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# AI tutor in training mode

> Practicum for Volume 2, Chapter 7, "Children: What to Teach and Where to Guide Them."
> Goal: set up an AI tutor that leaves the effort to the child instead of doing the work for them.

Copy the text below into the family's AI chat before a learning task.

```text
You help a child learn, but you do not solve the task for them.

Rules:
1. Do not give the finished answer right away.
2. First ask what the child already understands and where they are stuck.
3. Give only one hint at a time.
4. After each hint, ask the child to take the next step alone.
5. If the child makes a mistake, explain it with a question or a small example. Do not rewrite the whole piece of work.
6. At the end, ask 3 check questions to see whether the child keeps the skill without your help.
7. Separately state what the child did, what AI suggested, and what still needs practice.

Tone: calm and kind. Do not judge the child as a person.
```

## How to use it

1. The child makes the first attempt alone: a plan, solution, draft, or hypothesis.
2. Then the child turns on AI with the rule above.
3. AI asks questions and gives hints while the child moves the work forward.
4. The final check happens without AI for 5-10 minutes, in writing, aloud, or on paper.

## Quick prompt swaps

| Instead of | Ask for this |
|---|---|
| "Write my essay" | "Check my outline and ask 5 questions that will make the idea stronger." |
| "Solve the problem" | "Give me the first hint, not the solution. Then check my next step." |
| "Make the presentation" | "Help me collect facts, but I will choose and defend the slide structure." |
| "Rewrite this beautifully" | "Show where the idea became more precise and where the wording only became smoother." |

## A parent's check question

After the task, ask: **What can the child now do without AI?**

If there is no answer, the tool acted as a crutch rather than a training partner.

> Related: [Family 3C Audit](family-3c-audit.md) |
> [Prompt Bank: Friction Instead of a Finished Answer](training-not-answers.md).


---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/own-ladder -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# Build your own ladder: three materials

> Practicum for **Volume 2, Chapter 8**, "The New Economy of Value."
> Time: one evening. For professionals, managers, and business owners. Each of you stands on a different ladder.

The old career ladder was public: degree -> entry -> seniority -> promotion. Everyone knew the rules.
The new ladder has a narrow first step and a shortage of middle positions. A platform owner writes the rules and can change them without warning.
This does not mean everything is lost. It means resilience must come from materials that belong to you, not to the ladder.
There are three. Use this worksheet to audit them and choose one move per quarter for each.

---

## Material 1. Proof of work

When a machine reads the resume and everyone has credentials, **finished work** carries more weight.
Show projects with visible results, cases that follow "I took it, finished it, and this changed," and artifacts people can inspect.
Rule: every six months, add one result to your portfolio that you can explain in two minutes with a number, a name, or a consequence.

| Question | Your answer | Risk signal |
|---|---|---|
| Which 3 results can I show right now? |  | I can show only job titles and credentials |
| When did the latest result enter my portfolio? |  | More than six months ago |
| Can I explain any one of them in 2 minutes with a number? |  | I can only say "participated" or "was responsible" |

**Move for this quarter:** `_____________________________________________________`

## Material 2. Direct relationships

People who will call you without a platform offer the strongest protection from the platform.
They may be former colleagues, satisfied customers, or members of a professional community.
This is your own job market, with no commission, platform rating, or one-sided rule change.
Direct relationships cannot scale with a button. That is why they become more valuable.

| Question | Your answer | Risk signal |
|---|---|---|
| Who would invite me to a job or project without a platform? |  | Nobody specific, only cold applications |
| Do I have a public record that makes people contact me first? |  | No. A direct message feels like asking for a favor |
| When did I last go where people in my field meet in person? |  | My whole funnel is online, with no in-person points |

**Move for this quarter:** `_____________________________________________________`

> Do not confuse this with aggressive networking. People help when you approach them about real work and intend to contribute, not take their attention.

## Material 3. Portable reputation

A platform rating is capital held in someone else's bank. Change platforms and you start again.
A portable reputation is capital you control: a name people search for and find, public reviews of your work, writing, talks, and contributions to a community.
That record stays with you when you change employers, platforms, or even careers.
A quiet name known by twenty relevant people is stronger than a loud name that means nothing to everyone.

| Question | Your answer | Risk signal |
|---|---|---|
| What do people find when they search for my name? |  | Nothing, or only profiles owned by platforms |
| Which part of my reputation moves with me to another platform? |  | Everything depends on platform accounts and ratings |
| What makes my story authentic and hard to generate? |  | A course created my "personal brand," but there are no decisions behind it |

**Move for this quarter:** `_____________________________________________________`

---

## Final check: where is my weakest material?

- **Weakest of the three today:** `________________________________________`
- **One main move for the next quarter:** `_____________________________`

## Next

Check which platforms hold your visibility and rating with [Who Owns Your Ladder?](ladder-ownership-audit.md).
See the current labor market in [AI and the Labor Market](../watch/labor-market.md).
Use the same three materials in your decision portfolio with [No-Regret Moves](no-regret-moves.md) (Chapter 10).

Back: [Volume 2 Workbook](tom2-workbook.md).


---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/ladder-ownership-audit -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# Who owns your ladder?

> Practicum for **Volume 2, Chapter 8**, "The New Economy of Value."
> Time: 20-30 minutes. Map your dependence on the platforms that bring you work and customers.

A tradesperson spent five years building a rating: three hundred reviews and almost five stars.
One morning, the app brought half as many jobs. The person's work had not become worse. The platform had changed its ranking.
The skill remained, but visibility, reputation, and customer access sat inside somebody else's system, where the rules changed without warning.
This audit shows **which parts of your career infrastructure do not belong to you** and what happens if the owner changes the rules.

## Step 1. List your platforms

List every place that gives you work, orders, customers, or reputation: freelance sites, marketplaces,
professional networks, aggregators, and social platforms.

`___________________________  ___________________________  ___________________________`

## Step 2. Ask three questions about each one

| Platform | Where does my **visibility** live? | Who owns the **rating**, and can it move? | What happens after a **rule change**? |
|---|---|---|---|
|  |  |  |  |
|  |  |  |  |
|  |  |  |  |

- **Visibility:** can people find you without the platform, or only through its search and ranking?
- **Rating:** five stars and three hundred reviews are capital in somebody else's bank. They do not move when you leave.
- **Rule change:** what happens to your flow if the platform raises its fee, changes ranking, or limits access tomorrow? Do you have another route?

## Step 3. Rate your dependence

| Sign | Check if true for you |
|---|---|
| More than half my income or orders come through one platform | [ ] |
| My rating and reputation will not move if I leave | [ ] |
| Customers know me as a line in search results, not by name | [ ] |
| I have no backup channel if the rules change tomorrow | [ ] |

Three or four checked boxes mean the platform has a strong hold on you. Make the move below.

## Step 4. One move toward your own ladder

You may not need or be able to leave platforms. Move at least one asset from rented space into something you control:
direct contact with some customers, your own channel, or a portable name.

- **Asset I will move into my ownership this quarter:** `_______________________`
- **How I will do it** through direct contact, my own channel, or a public record: `______________`

---

## Next

Build that support with [Build Your Own Ladder: Three Materials](own-ladder.md).
The same rent mechanism affects businesses and computing infrastructure. Map it with the [Platform Rent Map](platform-rent-map.md) (Chapter 9).
See current market conditions in [AI and the Labor Market](../watch/labor-market.md).

Back: [Volume 2 Workbook](tom2-workbook.md).


---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/platform-rent-map -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# Platform rent map

> Practicum for **Volume 2, Chapter 9**, "Physical AI and Who Owns the Rails."
> Time: 30-40 minutes. For professionals, managers, and business owners. Each stands on different rails.

Platform rent is a regular payment to an infrastructure owner for access to customers, tools, or computing.
Examples include a marketplace commission, payment processing fee, model subscription, or the ad platform's share of customer acquisition cost.
Each payment may look reasonable on its own. The problem is paying **without looking** at the total or asking what happens if the owner raises the price.
This worksheet turns invisible rent into one number you can manage.

## Step 1. List every toll

Use the online store in the chapter as a guide: marketplace commission, payment processing, marketplace ads,
service subscriptions, delivery and returns, and per-request agent costs. Each looks small alone. Together, they are your margin.

| Toll | Who receives it | Monthly amount | % of revenue |
|---|---|---:|---:|
|  |  |  |  |
|  |  |  |  |
|  |  |  |  |
|  |  |  |  |
| **Total platform rent** |  | **₽ _______** | **____ %** |

> The point is not to leave every platform. It is to see your total rent as one number for the first time.

## Step 2. Find your position on the rails

On any rails, from a marketplace to a company AI subscription, you occupy one of three positions.
Mark your position as a professional and as a business.

| Position | How to recognize it | Move up |
|---|---|---|
| **Owns a junction** | Other people pay to access something you own: direct customer relationships, unique data, expertise with demand, or your own distribution | Deepen the moat through trust, context, and responsibility (Volume 1, Chapters 5-7) |
| **Builds on someone else's rails with eyes open** | You pay rent, know the total, keep a backup route, and own at least one junction | Move one asset from a platform into your ownership each year |
| **Pays rent without looking** | You do not know the total platform cost or what to do if prices double and access narrows | Start with this worksheet: list, add, and ask what you can replace or own |

**My position now:** `________________________`  **Move up that I will make:** `___________`

## Step 3. Check all three roles

- **Professional:** whose platforms stand between you and your value, and what do you own besides an account?
  This is your [own ladder](own-ladder.md): finished work, people, and a name. `_______________`
- **Manager:** whose models, clouds, and data support the process, and what happens to unit economics, the profit or cost of one unit,
  **if computing becomes three times more expensive** and access terms change? `_______________`
- **Business owner:** what is total rent as a percentage of revenue, and which **one asset** will you move into your ownership this year? `_______________`

## Step 4. Run the "what if the price rises?" test

Use rough but useful arithmetic. If the AI bill is 5% of unit cost, a threefold increase hurts but may not break the model.
If half the process depends on one API, a price increase affects the design of the business, far beyond one expense line.

- **AI and platform share of unit cost:** `____ %`
- **What breaks if prices become 3x higher tomorrow or access closes:** `_____________________`
- **My backup route** (multiple providers, portable data and prompts, open models, or local inference when justified): `_________________________________________`

> You probably do not need to build your own rails. For most people, the right answer is to use someone else's rails with open eyes.
> Know the toll, map the detour before the bridge closes, and own one junction that makes you useful on any route.

---

## Next

The same mechanism affects careers: [Who Owns Your Ladder?](ladder-ownership-audit.md) (Chapter 8).
See the current map of rails and infrastructure spending in [Who Owns the Rails?](../watch/physical-ai-rails.md).
Treat a backup route as a no-regret move with [No-Regret Moves](no-regret-moves.md) (Chapter 10).

Back: [Volume 2 Workbook](tom2-workbook.md).


---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/which-world -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# Which world are you building?

> Practicum for **Volume 2, Chapter 10**, "2030: Four Futures."
> Time: 15 minutes. This is not fortune-telling. It reflects your decisions. A scenario is not weather, and every decision is a vote for one world.

The chapter describes four possible worlds in 2030 as choices rather than forecasts.
The world that appears in your field will come from many decisions, including yours.
This check shows **which world your current decisions support**.

**The four worlds**, based on who receives the gains and what happens to the value of human contribution:

- **Platforms:** everything is convenient, but almost nothing is yours. Rail owners receive the gains.
- **Divide:** people keep the same qualifications, but the ladder disappears and the gains go to a narrow group.
- **Multiplication:** the machine prepares; the person decides and becomes more valuable.
- **Human premium:** anything can be generated, so what cannot be faked becomes more valuable.

---

## Choose the answer closest to your situation

| # | Question about your decision | Your choice |
|---|---|---|
| 1 | **Which channels bring sales or work?** A) only platforms and aggregators; B) mostly platforms, few direct contacts; C) platforms plus a growing direct base; D) mostly direct relationships and my own name | |
| 2 | **Do you have direct access to customers or an audience?** A) no, everything is in someone else's systems; B) I have contacts but do not use them; C) I maintain my own channel or list; D) customers know me by name | |
| 3 | **Where do the gains from AI go in your team or business?** A) I cut labor costs and share nothing; B) I have not measured it; C) freed time goes into people's growth; D) people become more valuable because they decide and take responsibility | |
| 4 | **What are you doing with your reputation?** A) keeping a platform rating; B) nothing; C) building a portable name; D) my signature on the result is the product | |
| 5 | **Your rails, such as models, clouds, and platforms:** A) one provider, dependence not measured; B) one provider, but I am thinking about it; C) several plus a backup route; D) I own a junction of value | |
| 6 | **What do you tell children or junior colleagues?** A) "learn a safe profession"; B) I have not thought about it; C) "train your thinking and build a portfolio"; D) "make what cannot be generated" | |
| 7 | **How do you respond to cheap output?** A) make more of the same for less; B) do what everyone else does; C) take responsibility for review and the decision; D) own the full result | |
| 8 | **Your bet for the next year:** A) a specific tool or technology; B) no clear bet; C) assets that survive a change of scenario; D) a human junction: judgment and trust | |

## Count the answers

See which letter appears most often.

- **Mostly A:** your decisions support the **Platform World**. It is convenient now, but you are giving away the junction of value. Make one move toward something you own.
- **Mostly B:** you are **moving by inertia**. The surest way to lose is to make no choice. Start with [No-Regret Moves](no-regret-moves.md).
- **Mostly C:** you are building the **Multiplication World**. Machines increase capacity while people grow. Keep going deeper.
- **Mostly D:** you are building the **Human Premium World** by investing in what cannot be faked. Check that you also have a backup route on the rails.

**The world my decisions support:** `___________________________`

---

## Next

Turn the result into action with [No-Regret Moves](no-regret-moves.md), a portfolio that works in all four worlds.
Separate a real shift from hype with [Checking AGI Claims](agi-claim-check.md), current [Signals Without Fog](../watch/agi-signals.md),
and [AI and the Labor Market](../watch/labor-market.md).

Back: [Volume 2 Workbook](tom2-workbook.md).


---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/no-regret-moves -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# No-regret moves

> Practicum for **Volume 2, Chapter 10**, "2030: Four Futures."
> Time: 20-30 minutes. Build a five-move portfolio that helps in any of the four worlds.

Strategists use the term *no-regret moves* for actions that pay off across different futures.
They make these moves before the future becomes clear. An investor might call this a portfolio instead of a single bet.
The book's argument leads to a portfolio of five moves. None says "learn tool X."
A tool is a bet on one scenario. This portfolio holds assets that can survive a change of scenario.

## Five moves: where I stand and what comes next

| # | No-regret move | Where I am now (1-5) | Next step |
|---|---|---|---|
| 1 | **I own the outcome, not the output.** I answer for the result, not the volume produced |  |  |
| 2 | **I keep direct access** to at least some customers or audience through a list that an algorithm change cannot take away |  |  |
| 3 | **I build a portable reputation, my own ladder:** finished work, people, and a name |  |  |
| 4 | **I strengthen the human junction:** judgment, trust, and responsibility become more valuable in every world |  |  |
| 5 | **I keep a backup route on the rails:** several providers plus portable data and prompts |  |  |

Rate each move from 1 to 5. A score of 1 means you do not work on it; 5 means it is a strength.
**My weakest move:** `_____________`

## How each move works in the four worlds

| Move | Platforms | Divide | Multiplication | Human premium |
|---|---|---|---|---|
| Own the outcome | a junction others pay to access | bargaining power | machines multiply my work | signature becomes the product |
| Direct access | insurance | support | accelerator | accelerator |
| My own ladder | an outside system cannot optimize it away | the only support | leverage | capital |
| Human junction | scarce | scarce | scarce | highest value |
| Backup route | protection from price changes | protection from disruption | sound engineering | sound engineering |

## Use the portfolio to test major decisions

The same five moves can test any major choice: an employer, sales channel, annual strategy, or a child's education.
Ask: **In how many of the four worlds does this option work?**
An offer with more pay but routine middle-layer work may win in one world.
A smaller offer with customer contact, a portable name, and new problems may win in all four.

**Three upcoming decisions I will test with "How many of the 4 worlds does it work in?":**

| Decision | Option A: how many worlds? | Option B: how many worlds? | Choice |
|---|---|---|---|
|  |  |  |  |
|  |  |  |  |
|  |  |  |  |

> This question cannot guarantee the right answer. It stops you from placing everything on one forecast without admitting it.

---

## Next

See which world your decisions support in [Which World Are You Building?](which-world.md).
Build move 3 with [Build Your Own Ladder: Three Materials](own-ladder.md) (Chapter 8).
Build move 5 with the [Platform Rent Map](platform-rent-map.md) (Chapter 9).

Back: [Volume 2 Workbook](tom2-workbook.md).


---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/agi-claim-check -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# AGI claim check: five questions instead of "Is this AGI?"

> Worksheet for Volume 2, Chapter 10, "2030: Four Futures."
> Time: 20-30 minutes for one claim.

Use this worksheet when you see a promise such as "we have AGI," "the model will replace a department,"
"the agent will do the job alone," or "people no longer need to learn."

Do not argue about the label. Break the claim into five testable questions.

## 1. What is the task?

Choose one task, not a profession or a whole company.

| Question | Answer |
|---|---|
| What result should the task produce? |  |
| What are the inputs? |  |
| What counts as a successful output? |  |
| How long does it take a person who has the right context? |  |
| Where does the task get messy: people, contradictions, missing data, politics, or trust? |  |

If you cannot describe the task in one paragraph, you cannot test the claim yet.

## 2. What does an error cost?

| Cost of error | Example | Your case |
|---|---|---|
| Low | draft, idea, internal note |  |
| Medium | customer email, calculation, public text |  |
| High | money, contract, health, safety, dismissal, reputation |  |

The higher the cost, the less suitable the "AI will figure it out" mode becomes.

## 3. How reliable is it?

Build a small test set of 20-30 real examples.

| Type of example | What to include | Result |
|---|---|---|
| Normal | what happens every day |  |
| Difficult | heavy context, ambiguity, or a long document |  |
| Bad input | missing data, contradictions, or the wrong format |  |
| Red zone | a case where AI must stop |  |

Do not record only pass or fail. Record the error type: invented a fact, missed a risk, took the wrong action,
failed to stop, or failed to explain its path.

## 4. Who controls the action?

| Control | Yes / no |
|---|---|
| A person owns the process |  |
| A log records inputs, sources, actions, and uncertainty |  |
| A stop rule exists |  |
| Access and permissions are limited |  |
| An action can be rolled back quickly |  |
| Each error returns to the test set |  |

Without an owner and a stop rule, this is not autonomy. It is ownerless risk.

## 5. Which human ability becomes stronger?

A useful AI system should strengthen a human ability, not only produce more output.

| Ability | Stronger? | How to tell |
|---|---|---|
| Judgment |  | the person sees options and consequences more clearly |
| Review |  | the person finds errors and weak points faster |
| Learning |  | the person understands the path instead of copying the answer |
| Responsibility |  | it is clear who signs off on the result |
| Independence |  | the person builds the workflow instead of waiting for instructions |

If the claim makes the person more passive, that is a warning sign even when the demo looks good.

## Final decision

| Question | Decision |
|---|---|
| Is the task clear? | yes / no |
| Is the cost of error acceptable? | yes / no |
| Did it pass the test set? | yes / no |
| Is control sufficient? | yes / no |
| Does the person become stronger? | yes / no |

### What to do next

- 5 yes answers: you may increase autonomy for this task.
- 3-4 yes answers: use it as an assistant or draft tool, not as an independent agent.
- 0-2 yes answers: the claim is still marketing. Define the task and build the test first.

## Formula

Do not ask, "Is this AGI?"
Ask: **Which task does the system solve, how reliably, at what cost of error, under whose control, and which human ability does it strengthen?**


---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/process-scoring -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# Process scoring: where an AI agent will pay off

> Practicum for **Volume 2, Chapter 11**, "The Change Compass" (WHERE arrow).
> Time: half an evening. Practicum address: `https://cheap-intelligence.vercel.app/`.

Do not try to "add AI everywhere." Choose the processes where an agent can produce a real return.
Apply the same questions to every candidate instead of following personal preference or a fashionable tool.
There are five questions. The threshold is simple: **fewer than three yes answers means leave it alone for now.**

---

## Step 1. List candidate processes

Map work, not departments. List 10-20 repeated processes with enough volume and pain.
Do not write "marketing." Write "answer inbound leads," "prepare proposals," or "match delivery notes."

`___________________________  ___________________________  ___________________________`
`___________________________  ___________________________  ___________________________`

## Step 2. Ask five questions about each process

| # | Question | A yes looks like this |
|---|---|---|
| 1 | **Does the process repeat?** | the same steps, rather than starting from zero every time |
| 2 | **Is the volume meaningful?** | tens or hundreds of cases a month, not five |
| 3 | **Can you check the result?** | a checklist, example, or number quickly shows whether it is good |
| 4 | **Is the cost of an error manageable?** | an error means rework, not a lost contract, reputation, or fine |
| 5 | **Do data and context exist?** | rules, examples, and history live somewhere beyond one person's head |

Use one more filter outside the score: **if you cannot write instructions for the process, it does not suit an agent.**
The agent is not the problem. You do not yet know what you want it to do.

## Step 3. Fill in the scoring table

| Process | Q1 | Q2 | Q3 | Q4 | Q5 | Total yes | Can I write instructions? | Decision |
|---|---|---|---|---|---|---|---|---|
|  |  |  |  |  |  |  |  |  |
|  |  |  |  |  |  |  |  |  |
|  |  |  |  |  |  |  |  |  |
|  |  |  |  |  |  |  |  |  |
|  |  |  |  |  |  |  |  |  |

**How to read the total:**

- **5 yes answers:** start here.
- **3-4 yes answers:** candidate for the second round.
- **Fewer than 3 yes answers:** do not touch it now. That is a result, not a defeat. Money you do not waste on the wrong implementation stays in the business.

## Step 4. "No" is also a result

Add a separate column for low-scoring processes. "No" will be the right answer for about half your processes.
Record **why**. This protects you from other people's excitement and from a vendor selling an implementation.

| Process | Why the answer is "no" now | What must change for it to become "yes" |
|---|---|---|
| low volume: | | |
| human relationships (human premium): | | |
| result cannot be checked: | | |
| high cost of error: | | |

> A useful reference from the chapter is Heathrow / Hallie. Repeated passenger questions earn five yes answers, so an agent fits.
> An unusual conflict or a difficult personal situation is not the first candidate. There, the agent prepares context for a person.

Before filling in your table, review the full public example:
**[IKEA Through the Change Compass](ikea-compass-example.md)**.

---

## Next

Once you choose a five-point process, map the route with the [Order of Moves Gate Sheet](change-gates.md).
Use the [Human Review Matrix](human-review-matrix.md) to decide what the agent may do and what a person should keep.
Check whether this is surface polish or a rebuild with the [Process Maturity Check](ai-native-maturity-check.md).

Back: [Volume 2 Workbook](tom2-workbook.md).


---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/ikea-compass-example -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# Example: IKEA through the Change Compass

> Practicum for **Volume 2, Chapter 11**, "The Change Compass."
> This is a teaching review of a public case, not a recommendation to copy IKEA word for word.

This worksheet shows what a completed compass looks like in a case you can check.
See how the four arrows apply to IKEA / Ingka Group, then return to your own processes.

**Case source:** Ingka Group, "AI and Remote Selling bring IKEA design expertise to the many"
(29 June 2023). The source says that from FY21 through 2023, the Billie chatbot resolved approximately 47% of
customer enquiries across 3.2 million interactions, with nearly EUR 13 million in savings. It also says 8,500
call-centre co-workers were reskilled in remote interior design, digital retail sales, relationship building, and
complex problem-solving. Sales through remote customer meeting points reached EUR 1.3 billion in FY22,
or 3.3% of total sales.

**Caveat:** this is a public corporate case, not an independent study. Do not write that the chatbot
"created EUR 1.3 billion." A more accurate reading is that IKEA automated a flow of simple requests,
retrained some employees, and developed remote selling and remote customer meeting points into a meaningful sales channel.

---

## 1. WHERE arrow: what the agent should do

IKEA did not give the agent full responsibility for "customer care." The case shows two types of work.

| Work | Score | Decision |
|---|---:|---|
| Simple repeated customer questions: status, product information, basic help | 5/5 | Give to Billie |
| Interior advice, solution design, relationships, and difficult unusual questions | 1-2/5 | Keep with a person and support them with AI |

Why the first flow passes the score:

| Question | Answer from the case |
|---|---|
| Does the process repeat? | Yes. Customer enquiries arrive continuously and follow repeated patterns |
| Is the volume meaningful? | Yes. The chatbot solved 3.2 million interactions |
| Can you check the result? | Yes. The question was resolved or not, the customer received the needed answer, and escalation exists |
| Is the cost of an error manageable? | Yes, if the agent is limited to simple questions and hands difficult cases to a person |
| Do data and context exist? | Yes. IKEA has product information, customer questions, a knowledge base, and 80 years of life-at-home knowledge |

**Lesson for your business:** do not search for a function to replace. Find a task flow inside the function.
The right unit may be the first 40-50% of repeated questions, not "customer support" as a whole.

## 2. PEOPLE arrow: how roles change

The strongest part of the case is what happened to people after Billie took simple questions.
IKEA says 8,500 call-centre co-workers were reskilled for remote customer meetings around interior design competence,
digital retail sales, relationship building, and complex problem-solving.

This was not an AI course completed for a check mark. The role changed.

| Before | After |
|---|---|
| Answer a stream of similar questions | Help the customer build a solution for the home |
| Reduce pressure on a queue | Create value in remote selling |
| Work as a cost centre | Work closer to a revenue channel |

**Check your own case:** if people returned to the old process after the implementation, you added a tool.
If their role changed, you started a real change.

## 3. EFFECT arrow: what to measure

The case gives two groups of metrics. Do not mix them.

| What is measured | Public figure | Honest reading |
|---|---:|---|
| Effect of automating the simple flow | nearly EUR 13 million in savings | Lower operating load |
| Work moved to self-service / bot | 47% customer enquiries, 3.2 million interactions | Scale of the repeated process |
| New or stronger sales channel | EUR 1.3 billion remote customer meeting points sales in FY22 | Sales tied to remote selling; the chatbot did not "create" the full amount |
| Share of the channel | 3.3% total sales | Evidence that the change became a business loop rather than a pilot |

Ask one question about your process: **Which metric changes if the agent is genuinely useful?**
Without a "before" metric, the IKEA case becomes a nice story rather than a working example.

## 4. CONTROL arrow: what is known and unknown

The public source says Billie handles simple customer enquiries while people take more complex and value-adding interactions.
It does not explain the full control design.

| Control question | What the public case supports |
|---|---|
| Is there an autonomy boundary? | Yes. The chatbot handles simpler work and co-workers handle harder work |
| Is there a named owner of the loop? | Not disclosed |
| How does escalation work? | The split between "simpler enquiries" and complex problem-solving shows the logic, but details are not disclosed |
| How is answer quality measured? | Not disclosed |
| What happens after a failure? | Not disclosed |

Apply a stricter standard to your own project. A public case may give you an idea, but your loop needs an owner,
stop rules, and an action record.

## 5. The seven gates in the IKEA case

| Gate | How it appears in the case | What to take into your work |
|---|---|---|
| 1. Work map | Customer enquiries are separated from advisory work | Start with a task map, not a department |
| 2. Scoring | Simple enquiries pass five questions; design consultation does not | Automate a flow, not a profession |
| 3. Baseline | Interaction volume, share of resolved enquiries, and savings are reported | Record "before" before launch |
| 4. Agent contract | The public case shows a boundary around simple enquiries, but not the contract details | Write down your boundaries instead of holding them in your head |
| 5. People learn in the loop | 8,500 employees were reskilled for new roles | Training should change the work, not only add knowledge |
| 6. Effect measured | Savings and remote customer meeting points sales are separate | Do not mix savings with growth |
| 7. Second round | IKEA plans to grow remote selling further | Scale after the first loop produces numbers |

## 6. Apply it to your business

Complete a short version for your case.

| Question | My answer |
|---|---|
| Which simple repeated flow looks like Billie in our business? | |
| Which more valuable human work will this free up? | |
| Which 2-3 "before" metrics must we record before the pilot? | |
| Which autonomy boundary will we put in writing? | |
| Who needs training for a new role rather than "training in AI"? | |

If you cannot answer the second question, you are probably building a cost-cutting project.
That may still be useful, but it is a different conversation. The Change Compass works best when automation of a simple flow
frees people for more valuable work.

---

Next: [Process Scoring](process-scoring.md) | [Gate Sheet](change-gates.md) |
[Training Inside the Work Loop](training-in-the-loop.md).


---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/change-gates -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# Order of moves: gate sheet

> Practicum for **Volume 2, Chapter 11**, "The Change Compass" (first-round route).
> This is the printed capstone for Volume 2. Print it and keep it visible.

A gate is a control point, a question that needs an honest yes before you take the next step.
The rule is **not "by Friday," but "when this is true."** A date asks how much time has passed. A gate asks what has become true.
One business may pass seven gates in a quarter; another may need two. A calendar ignores that difference, while the compass allows for it.
Order matters more than speed. **Every skipped gate returns later at the most expensive point.**

---

## Seven gates in the first round

| # | Gate: what must become true | Honest yes? | Date passed |
|---|---|---|---|
| 1 | **The work map exists.** I can name 15 repeated processes with volume and pain | [ ] | |
| 2 | **Scoring is complete and ONE process is selected.** It has five yes answers and written instructions | [ ] | |
| 3 | **The baseline is recorded.** Four "before" numbers are written down | [ ] | |
| 4 | **The agent contract exists and an owner is named.** There is a contract page and a person's name | [ ] | |
| 5 | **The pilot runs and people learn from it.** The loop has run for a month, 4 weekly review hours happened, and a champion is visible | [ ] | |
| 6 | **The full-cycle effect is measured.** The result is compared with the baseline and tied to one of three levers. Decision: scale / repair / close honestly | [ ] | |
| 7 | **Second round.** The next scored process moves faster because context exists, people know how, and the team has practiced saying no | [ ] | |

**Rules for passing the gates:**

- **One process, not three.** Three first loops running at once divide the owner's attention and rarely survive.
- **Gate 3 cannot move.** Without "before," there is no "after." Record the baseline before launch or accept that it does not exist.
- **Gate 6 is the honesty point.** Closing a pilot with numbers is a sound result. An endless pilot without numbers is the worst result. Set a review date and allow three honest outcomes from the start.
- **My first-round process:** `_________________________________________________`
- **My four "before" numbers for Gate 3:** `___`  `___`  `___`  `___`

Before you fill in the sheet, see how the gates work in a public example:
**[IKEA Through the Change Compass](ikea-compass-example.md)**.

## If you manage a team inside a company

The route stays the same, but the scope is smaller. Your map covers team processes, and your mandate covers one loop.
Your advantage is speed: you do not need the board to pass the first five gates.
A manager who brings leadership a working loop and measured effect has a different conversation from someone presenting slides about AI's potential.

- Detailed routes: [Manager Route Through the Gates](manager-gates-route.md) and
  [Business Route Through the Gates](business-gates-route.md).

## Exit check: did you only add surface polish?

Before you call the round complete, ask whether the work map, roles, or measurements changed.
If none changed, you added surface polish. Scoring will not catch this. The final question will.
Use the [Process Maturity Check](ai-native-maturity-check.md).

---

## Next

If you have not scored processes, start with [Process Scoring](process-scoring.md).
For the PEOPLE arrow, Gate 5, use [Training Inside the Work Loop](training-in-the-loop.md).
For the CONTROL arrow, Gate 4, use the [Agent Contract](agent-contract.md) and [Human Review Matrix](human-review-matrix.md).

> Adding AI is easy. Making the change real is harder. Seven gates and one honest metric separate a rebuild from a gesture.
> Back: [Volume 2 Workbook](tom2-workbook.md).


---

<!-- Page: https://cheap-intelligence.vercel.app/en/playbooks/training-in-the-loop -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

# Training inside the work loop

> Practicum for **Volume 2, Chapter 11**, "The Change Compass" (PEOPLE arrow).
> Time: 20 minutes for diagnosis and a first-month plan.

Everyone finished the course, but nobody uses it. This is one of the seven ways a change dies.
Training works **next to a live work loop**, not away from it. One hour a week reviewing a real task beats a week of lectures.
Use this worksheet to tell live learning from a completion box and plan the first month.

---

## Signs of live learning and box-checking

| Live learning | Box-checking |
|---|---|
| People learn on **real tasks** from their own process | People use abstract course examples |
| A **one-hour weekly review** happens beside the working loop | There is one training session, followed by "figure it out yourselves" |
| A **champion has a mandate** to act and carry the work | Nobody owns it, and everyone does a little |
| A person has a **new role** after training, such as reviewer or loop owner | Roles stay the same, so training changes nothing |
| Each employee understands **what they personally gain** | "Management said so," and resistance goes underground |

What is still missing: `______________________________________________`

## Resistance is a rational calculation

Employee resistance is not simple stubbornness. It is a reasonable response to the unanswered question, "What happens to me?"
Clarity works better than pressure. Give each important person a short, honest answer.

| Person | What they may lose (their fear) | What they gain (honestly) |
|---|---|---|
|  |  |  |
|  |  |  |

> A warning from the chapter: do not turn the team into debtors on a cognitive loan.
> If people stop thinking and only press a button, you have trained dependence rather than a team.
> Training should leave a person stronger, not helpless without the agent.

## First-month mini-plan (Gate 5)

This connects to the [gate sheet](change-gates.md). Gate 5 closes after the loop has run for a month and four weekly review hours have happened.

- **Process loop where we learn:** `_______________________________`
- **Champion, with name and mandate:** `_______________________________________________`
- **Day and time of the weekly review:** `___________________________________`
- **Week 1 review:** `_______________________________________________`
- **Week 2 review:** `_______________________________________________`
- **Week 3 review:** `_______________________________________________`
- **Week 4 review:** `_______________________________________________`
- **New role that appeared in the team:** `___________________`

---

## Next

Follow the full route with the [Order of Moves Gate Sheet](change-gates.md).
Place the person in the loop with the [Human Review Matrix](human-review-matrix.md).
Define the agent's permissions and boundaries with the [Agent Contract](agent-contract.md).

Back: [Volume 2 Workbook](tom2-workbook.md).


---

<!-- Page: https://cheap-intelligence.vercel.app/en/skills/agent-contract/SKILL -->

> **Draft translation, not author-reviewed.** The Russian edition is the source of truth.

---
name: agent-contract
description: >-
  Builds a contract for an AI agent across ten fields — role, mission, inputs, tools, memory,
  boundaries, escalation, review, metrics, owner — and turns the metric into a measurable threshold
  at which the agent stops itself. From chapter 3 of "Business in the Age of AI Agents".
  Use when the user is about to put an agent into a team process and needs its limits, escalation
  path, owner and stop threshold written down. Triggers "agent contract", "agent boundaries",
  "who owns this agent", "where should the agent stop", "we are putting AI into this process".
license: MIT
---

# The agent contract (Practicum for "When Intelligence Became Cheap")

📋 A ready-made skill for an AI assistant. Drop this file into your agent's skills folder — or copy
the block below into any chat and describe the one agent you are about to launch.

One agent, one contract. The worksheet, the template, two worked examples and the pre-flight list:
[The agent contract](../../playbooks/agent-contract.md). Before the contract, work through the
[Human Review Matrix](../../playbooks/human-review-matrix.md): the matrix decides which tasks can be
handed over at all, the contract pins down a single agent.

```text
Help me write a contract for an AI agent, using the model from "Business in the Age of AI Agents"
(chapter 3). An agent on a team is not a magic employee: it is a role with explicit boundaries, a
review step and a human owner.

Collect from me if I have not given it:
1. What the agent is and which process it will work in.
2. What data and systems it will have access to.
3. Who on the team is accountable for it, by name.

What to do — fill in ten fields, asking me where the information is missing:
Role · Mission (one sentence) · Inputs · Tools · Memory (and what must NOT go into it) ·
Boundaries (what the agent never does without confirmation) · Escalation (the condition under which
it hands over to a human) · Review (how and when the result is checked) · Metrics (one or two
numbers) · Owner (a name).

Then tie metric, escalation and stop into a single guardrail:
— signal metric (share of answers needing no edits · anomaly in the numbers · escalation rate ·
  cost per day);
— threshold: a specific number, named BEFORE launch;
— action at the threshold: pause and human review · roll back to manual · switch off.

Discipline (hold to this):
- Without Boundaries, Escalation and Owner this is not a contract but a wish. If I do not supply
  them, do not fill them in with plausible text — say plainly that the contract is not ready.
- The owner is a named person. "The team" and "the department" are not accepted.
- Review is not delegated to AI: name where a human looks at the result.
- Access is least privilege: only what the mission needs.
- The threshold is named before launch, not after an incident.
- Do not scale straight after the contract: next come the eval set and shadow mode.
- Do not invent system names, integrations or numbers on my behalf.

Answer format:
1. Contract table: field | value.
2. The guardrail: metric | threshold | action.
3. Pre-flight: the six-point checklist, marking what is closed and what is not.
4. The weakest point here, in one honest sentence.
```