Search the Practicum

Type two letters · ↑↓ select · Enter open · Esc close

Type at least two letters.

Business cases: where AI agents already produce results

This is a live Practicum page for Volume 2. Verified: 2026-08-18. This field changes every month, so check the date. It is a catalog of practical uses by business function: use case, region, effect, implementation effort, and source.

Discipline: this is a map of applications, not an advertisement or a promise. Numbers become outdated. Vendor statements are marked "vendor claim" and are not independent facts. Information that cannot be verified has a clear label such as "claim," "not disclosed," or "not independently confirmed." Sources are at the end of the page. Another company's case is not your skill. Use the pattern, including how the task was set and checked, rather than copying its number. Do not trust. Verify.

This page shows where AI pays off in specific business functions, then separates those cases from work that still needs a person. Before you copy a case, run your process through the five-question score: repetition, volume, verifiability, affordable errors, and available data.

How to use the catalog

  1. Find your function below. The columns show the use case, region, effect as data or a claim, implementation effort, and source.
  2. Effort is a rough implementation estimate. Low means a standard chat or draft. Medium means a process with checks. High means process redesign, autonomy, or a high cost of error.
  3. Take the pattern, not the tool. Test it on your work with the small-business staff map, then describe the first agent in an agent contract.
  4. Keep the boundary clear. In some cases the agent does the work. In others it only prepares a person for work that should not be delegated.

General context, based on research rather than a vendor: McKinsey's Nov 2025 "The state of AI" report found that 23% of organizations were scaling an agent system in at least one function, and no more than 10% were doing so in any individual function. IT and support were the most common functions. Gartner predicted in Jun 2025 that more than 40% of agent projects would be canceled by 2027 because of cost, unclear value, and weak controls. Both sources appear below.

Sales and marketing

What the agent does. It prepares proposals and emails, qualifies leads, builds customer and market files, handles the first content and newsletter pass, and monitors brand mentions.

Use case Region Effect Effort Source
Content, proposals, customer files, monitoring Russia Marketing and sales are among the functions with the highest GenAI adoption Low Yakov and Partners / Yandex, 2024 to 2025
Retail personalization, such as the Chitai-Gorod case on Yandex AI Studio Russia Vendor reports about a 7.7% rise in conversion, based on a presentation Medium Vendor claim, not independently confirmed
Media monitoring and reach of brand mentions with Meltwater Western markets Vendor reports 25% less time spent Low Meltwater vendor claim
Selling inside a chat assistant: OpenAI Instant Checkout US, global Withdrawn. Ran Sep 2025 to Mar 2026; about a dozen Shopify merchants integrated; Walmart measured roughly 3× worse conversion than a click through to its own site High OpenAI, Mar 2026; conversion figure via Forbes
Discovery in an assistant, purchase on the merchant site US, EU The pattern the market settled on after the above. Agent handles search and intent; the merchant keeps checkout Medium Industry consensus after Mar 2026; treat as direction, not a measured effect
Boundary General Good copy is not the same as a sold product. A person must handle a major customer negotiation and own the promise of a result. Not applicable Not applicable

Pattern: do not ask an agent to "write a post." Ask it to prepare a customer file, deal history, and three possible conversation plans before tomorrow's meeting. The agent prepares the person for work that should stay human.

Customer service and support

What the agent does. It answers common requests, drafts replies for an operator, sorts and routes tickets, and maintains a source-linked knowledge base.

Use case Region Effect Effort Source
First-line service and internal employee assistants Russia Scale has not been disclosed publicly Medium Banks and telecom companies such as Sber, MTS, and Alfa; scale not disclosed
First-line support at Klarna Western markets The AI assistant handled work equal to about 700 operators and two thirds of chats in its first month in 2024. The company later said it had gone too far and brought people back to improve quality in 2025. Medium Klarna, 2024; reversal reported by Bloomberg and CX Dive, 2025
Support and IT, the most common functions for agent adoption Global data 88% of organizations use AI in at least one function and 62% at least experiment with agents, but only 23% scale an agentic system anywhere, and no more than 10% within any single function Not applicable McKinsey, The State of AI in 2025, published Nov 5, 2025
Boundary General A ticket error may be cheap, but repeated reputation damage is expensive. Use the Human Review Matrix. Not applicable Not applicable

Analysis and reporting

What the agent does. It combines data from different formats, prepares recurring reports and market briefs, and classifies unstructured records.

Use case Region Effect Effort Source
Internal analytical assistants using GigaChat or YandexGPT in a corporate environment Russia Company-level effect has not been disclosed Medium Corporate deployments; no public figures
Lower cost in software development, IT, and analytics Global data 10% to 20% lower cost only where the company redesigned the process, rather than adding AI to the old one High McKinsey, 2025
Task-specific agents shipped inside enterprise software Global data Forecast: 40% of enterprise applications carry task-specific agents by end of 2026, up from under 5% in 2025 Not applicable Gartner forecast; a projection, not a measurement
Boundary General Models confuse number formats and extrapolate poorly. Manual sampling is required, and a person must define a good report before the run. Not applicable Not applicable

What the agent does. It compares contract terms, drafts a position with references to rules, prepares claims and responses, and performs an initial review of a document package.

Use case Region Effect Effort Source
Agent review of incoming contracts Russia The author reports about 90% fewer requests to external lawyers in a routine flow Medium Author's case, Volume 2, Chapter 1; self-reported
Legal copilots such as Harvey for drafts and research reviewed by a lawyer US, UK Faster initial research High Vendor claim, not independently confirmed
Documenting AI use in regulated work EU From Aug 2, 2026 the EU AI Act's transparency duties are enforced, with fines to €15M or 3% of worldwide turnover. Compliance work is now a line item, not a future risk High European Commission, enforcement began Aug 2, 2026
Boundary General Check legal citations and wording against the primary source or with a lawyer. A person makes the decision and signs. Not applicable Not applicable

HR, hiring, and development

What the agent does. It drafts job descriptions and offers, performs initial screening, prepares personal development plans, supports 360 reviews, and creates training materials.

Use case Region Effect Effort Source
360 reviews, personal development plans, and resume review in Russian services Russia Company-level effect has not been disclosed Medium Russian HR services; see the HR role, including Potok 360 and MTS AI
AI ability as a hiring requirement Global trend Shopify asked teams to show that AI could not do a task before requesting more people, described as "hire an AI before a human" in Apr 2025 Not applicable Shopify / Tobi Lutke, CNBC, 2025
Entry-level hiring after AI adoption US, UK, EU The clearest labour signal of 2026, and it lands on the entrance rather than the whole market: junior postings squeezed hardest, and employers loading junior roles with judgment and stakeholder work, a shift researchers call seniorization Not applicable Estimates vary widely by dataset and definition. See AI and the labour market before quoting a number
Boundary General Screening people has a high cost of error and a risk of algorithmic bias. A person must make final decisions about people. Not applicable Not applicable

Operations, back office, and procurement

What the agent does. It processes incoming requests, checks invoices and shipping documents, prepares paperwork, handles routine correspondence, and performs an initial review of bills.

Use case Region Effect Effort Source
Agents built on accounting systems, sometimes called RPAi Russia Company-level effect has not been disclosed Medium Large companies and banks, including Cloud.ru; no public figures
Scaling agents inside a function Global data 23% of organizations scale agents in at least one function, and no more than 10% do so in any individual function. Value comes from redesigning the process. Medium McKinsey, The State of AI in 2025, published Nov 5, 2025
Boundary General Start with processes that receive five yes answers in the score: repeated work, enough volume, a verifiable result, an affordable error, and past examples. Not applicable Not applicable

Software and product development

What the agent does. It writes and reviews code, prepares prototypes, tests, and documentation. At the far end, one person coordinates a team of agents.

Use case Region Effect Effort Source
AI-native agency, the SLSBMB case described by A. Gladkov Russia Staff reportedly fell from about 84 to 2 people, with about 36 research agents and self-healing systems. The account also states the cost: about 16 hours of work a day. High Founder's self-description, not independently confirmed
Agent projects overall Global data More than 40% may be canceled by 2027 because of cost, unclear value, and weak controls High Gartner, Jun 2025
Coding agents in normal engineering work Global data By Feb 2026 Claude Code alone was authoring roughly 4% of all public GitHub commits, about 135,000 a day. The common company pattern is a broad autocomplete baseline plus an agent used by senior engineers Medium Public commit analysis; see the tool map
Boundary General The stronger the hope of removing people entirely, the more failures and review work tend to appear. Not applicable Not applicable

Governance: AI in the boardroom

What the agent does. It prepares board materials and analyzes scenarios. In a few experiments, a company has formally added an AI system as an adviser, or has even claimed that it can vote.

Use case Region Effect Effort Source
An "AI director" on a board Asia Experiments include Mitani Sangyo's "Hokuto Taizan" in Japan and Samruk-Kazyna's "SKAI" in Kazakhstan, which the company described as a voting member High, experimental Public company statements, not independently confirmed
AI board adviser, Aiden Insight at IHC in Abu Dhabi Western markets Overview of the topic; the question of whether AI boards will outperform people remains open High, experimental Harvard Business Review, Nov 2025
Boundary General A vivid image of AI in governance does not make a machine accountable. Human directors make the decision and own it. Not applicable Not applicable

Where the approach breaks

These boundaries are not arguments against AI. They are what turns an application into a result.

Use caution when What practice shows
An error is expensive Do not hand an annual contract, reputation risk, or regulatory penalty to an agent without strict human review.
The work depends on a live relationship Major negotiations and difficult conflicts are where human judgment earns its premium.
The result cannot be checked If you will not know whether the work is good for six months, the agent can produce confident garbage. Define the test first.
A vendor is agent washing Gartner estimates that only about 130 of thousands of "agent" vendors offer real agent capabilities. Check whether the product is an agent or a renamed chatbot.
A company reverses course Even famous cases change. Klarna brought people back into support in 2025 to improve quality. Autonomy is not a system you set and forget.
You copy someone else's case A catalog does not create tool fluency. These are patterns, not magic prompts.

Sources (verified 2026-08-18)

Data is linked to a primary source. Vendor figures carry a "vendor claim" label and have not been independently confirmed. Links and numbers become outdated, so check the date.

Next steps

For personal and work cases from Volume 1, with before and after examples, see the case bank. To choose a starting point for your company, use the small-business staff map with the process score. For current role-based tools available from Russia, see the tool map.

Version: 2026-07. What changed: the catalog now uses columns for the use case, region, effect, effort, and source. The sources section contains links to McKinsey 2025, Gartner 2025, Shopify and Lutke, Klarna, and Yakov and Partners. Vendor numbers carry a "vendor claim" label. The Klarna 2025 reversal is included as an anti-case. Review once a quarter and after a material event.

Business cases: where AI agents already produce results