Specialist cases: where AI raises the ceiling

This is a live Practicum page for Volume 1. Verified: 2026-08-18. The field changes, so check the date. It catalogs techniques reported by working specialists: the input, how they framed the task, what came out, and where a person was required.

Discipline: these are community field reports, not verified return-on-investment cases. Every figure is a participant's own estimate and has not been independently confirmed. Take the pattern, including how the task was framed and checked, rather than the number. Another person's case is not your skill. Do not trust. Verify.

The page does not try to list everything AI can do. It shows where AI can raise a specialist's ceiling by making work possible that used to require a team or several days. It also marks the points that still need a person. This complements the personal before-and-after examples in the case bank, but organizes them by work category.

Career and job search

Pattern. Combine a resume with a job description to draft a direct cover letter. Analyze how your role is changing. Practice interview answers.

Example, self-reported and not independently confirmed: a participant used "resume plus job description" to produce a focused cover letter and estimated that the process became about twice as easy.
Boundary. Tests, review, and a merge request separate "it works" from "it broke with confidence." Autonomous agents need stop controls and approvals. See the anti-cases below.

The scale behind these individual stories, checked Aug 2026
Every case on this page is self-reported. One measured figure gives them context: by Feb 2026, Claude Code alone was authoring roughly 4% of all public GitHub commits, about 135,000 a day. So the pattern here is not a few enthusiasts. What individual stories cannot tell you is the failure rate, because nobody posts about the afternoon the agent broke the build. Read the cases for the method, and take the discipline in the boundary line above as the non-optional part.

Software development: roles, contracts, and checks

Pattern. Do not ask one model to "write the code." Use supervised roles: an on-call agent finds the issue, a mechanic fixes it, and a reviewer checks it. Require tests and a merge request instead of sending a raw script to production.

Example, self-reported and not independently confirmed: a repeated TypeError had produced "1,274 events over six weeks." The fix was reportedly built in "two calls, four tests, and a finished merge request in about 15 minutes from report to pull request."
Example, semi-technical work moved into production: a WordPress checkout plugin and a Google Shopping export were reportedly completed "in half an hour." AI raised the ceiling for someone who would previously have stopped at the edge of their technical ability.
Boundary. Tests, review, and a merge request separate "it works" from "it broke with confidence." Autonomous agents need stop controls and approvals. See the anti-cases below.

Analysis and a knowledge base

Pattern. Keep a personal base of notes and use an LLM as a lightweight retrieval system instead of starting each search again. Review dozens of sources while retaining links to the originals.

Example, self-reported and not independently confirmed: one participant compared a review of about 50 specialist articles that took one or two workdays by hand with about 30 minutes using a strong model.
Boundary. A model summarizes and extrapolates poorly. Check details, formulas, and physical claims in the primary source. The value sits in your base and quality criteria, not the chat itself. See the harness Knowledge layer.

Negotiations and contracts

Pattern. Compare contract terms, build a position with references to rules, and rehearse a difficult meeting before you enter the room.

Example, self-reported and not independently confirmed: in one disputed contract, the contractor reportedly admitted the mistake and gave a discount of about GBP 15,000. In another case, annual cost reportedly fell from 80,000 to 30,000. These figures come from the authors of the reports.
Example, meeting preparation: a simulation of the other side reportedly predicted about 90% of the questions asked in the real meeting.
Boundary. Check legal references and wording in the primary source or with a lawyer. A person signs and makes the decision. High-cost errors require review.

Compliance and review at scale

Pattern. Let AI perform the first review of hundreds of similar items against a checklist. A person handles disputed items and makes the final decision.

Example, self-reported and not independently confirmed: a compliance process reviewed "hundreds of creative items each month, with a success rate above 95%." One employee had previously done the work.
Example, qualifying leads through rejection: ask for reasons to reject a candidate. A participant said accuracy rose from about 70% to more than 95%.
Boundary. Public case samples are biased because strong enterprise cases often sit under nondisclosure agreements. Do not treat another team's success rate as a promise. Review a regular manual sample in any high-volume process.

Digital hygiene and a personal system

Pattern. Organize files, projects, and context before using powerful agents. Otherwise the agent scales the mess.

Example, self-reported and not independently confirmed: after an agent organized a home folder, "Downloads fell from 16 GB to 72 KB, 49 projects were moved, and the task took about two hours."
Boundary. An agent with access to files, email, or drives inherits your permissions. Start in a sandbox and require approvals. Do not tell it to finish everything without stopping.

Anti-cases: where the technique breaks

Use caution when	What practice shows
An agent has broad access	One autonomous agent reportedly connected to a laptop over SSH on its own. Agents need stop controls, approvals, a sandbox, and minimum permissions.
Voice bots call customers without care	Participants report that bots often annoy customers. A bad autopilot can cost more than having no autopilot.
The system says "done" too smoothly	The less you understand the path to the answer, the larger your cognitive debt and the risk of missing a confident error.
You copy someone else's case	A case library does not create tool fluency, just as a phrasebook does not teach a language.

Next steps

Build your AI working system with the personal harness builder. Check whether debt is building with the cognitive debt tracker. For personal before-and-after techniques, use the case bank. To examine your profession, use the profession catalog.

Version: 2026-07. What changed: community field reports now carry an explicit "not independently confirmed" caveat. This is the honest final status for anonymous community reports, which cannot be independently verified, not an open task. Review once a quarter. Every figure is self-reported and not independently confirmed. Do not treat it as a guarantee of your result.

Sources (verified 2026-07)

This page is based on community field reports that have not been independently verified. Each example says so. There is no separate primary-source list, because anonymous self-reports do not have a verifiable primary source. For data on AI results by business function, use business cases, which links to McKinsey, Gartner, and labeled vendor claims.