Independent professional track — 14 days

Target

A solo professional or micro-activity with a website, client inquiries, quotes, suppliers, or content production.

Intended outcome

One low-risk workflow that is measured, reversible, and documented. The objective is not to create a general-purpose agent.

For a complete worked example, open the 14-day client follow-up case.

Days 1–2 — Choose the problem

• identify five repetitive tasks over one week;
• measure volume, minutes, and errors;
• exclude decisions involving health, employment, credit, rights, or other high-impact areas;
• complete the mandate and a use-case card.

Good first cases include drafting a reply, summarizing public documents, structured extraction, or comparing suppliers with human validation.

Days 3–4 — Data and tool

• list the data needed and the data that is prohibited;
• verify retention, training, region, and deletion with the supplier;
• choose among conventional automation, a simple model call, or controlled retrieval;
• keep passwords and secrets outside the context.

Days 5–7 — Evaluations

• select 20 to 50 representative cases;
• define a correct response, an escalation, and a severe failure;
• set the threshold before testing;
• compare time, quality, and correction rate with the baseline.

Days 8–10 — Shadow mode

Have the AI produce outputs without sending, publishing, or modifying anything. Review every result. Add newly discovered exceptions to the test set without retroactively changing the decision threshold.

Days 11–12 — Copilot

Allow an action only after explicit confirmation. Before approving, the person must see the input, output, and destination.

Days 13–14 — Decision

• continue if value and reliability are demonstrated;
• correct if a specific cause can be identified;
• stop if the case requires too much control or fails to outperform the simple process.

Keep a manual procedure, change log, versions, and costs. Reassess after any material model or data change.

Minimum foundation

• owner: the independent professional;
• simple AI register;
• backup and manual fallback;
• data rules;
• reusable tests;
• monthly review of cost, errors, and value.