Synthetic independent-professional case — From meeting to reviewed follow-up

Fictional example. The person, clients, volumes, and results are synthetic. The scenario illustrates a cautious pilot; it does not prove the performance of a tool or profession.

Measured level: A1 copilot. The system drafts one step without connected tools or autonomous action. Its observed gain must not be used as a benchmark for a business agent or an orchestrated agency.

1. Starting point

Camille Rey is an independent operations consultant supporting six small companies. She normally holds 12 to 18 client meetings per month. After each meeting, she reviews her notes, prepares a summary, lists next actions, and drafts a follow-up email.

Some meetings are straightforward. Others contain hypotheses, unconfirmed decisions, or commercially sensitive information. Prices, contractual scope, and final recommendations directly commit the consultant.

Baseline over 22 authorized historical follow-ups

The selected problem is not “automate the client relationship.” It is: turn structured notes into a first summary and email draft without deciding, promising, or sending on the consultant’s behalf.

2. Mandate and boundaries

Written exclusions include meeting recording or automatic transcription; health, HR, dispute, and third-party-secret data; model access to email, calendar, storage, or billing; generated prices, contractual commitments, or final recommendations; automatic recipient selection or sending; and supplier training on client data without verified conditions.

Camille uses a local five-field form: authorized context, confirmed facts, decisions, actions, and points to clarify. Unnecessary information is removed before a draft is requested.

3. Simplest sufficient system

Three options are compared: a manual mandatory-field template; the same template with one model call producing structured output; and an assistant connected to meetings, email, and calendar.

The second option is selected. The form supplies the structure, and the model produces confirmed facts, actions with owners and dates, and an email draft. Price, commitment, and recipient fields do not exist. Camille always compares the proposal with her notes. The connected option is rejected because it adds data, permissions, and external effects that the test does not need.

4. Pre-registered evaluation

Twenty-four authorized closed files are minimized. Twelve tune the form and instructions; twelve remain frozen. The frozen set includes ambiguous dates, an unconfirmed decision, an action without an owner, and four cases that must be flagged instead of completed.

Frozen-set results

One secondary date is attached to the wrong milestone. The error stays in the log, and a visual date check is added to the review checklist.

5. Fourteen-day pilot

• Days 1–2 — Measure and bound. Confirm the baseline, eligible categories, manual fallback, and supplier data conditions.
• Days 3–7 — Tune and test. Use 12 tuning cases, freeze the instructions, then run the 12 decision cases. A critical failure stops the pilot.
• Days 8–10 — Shadow mode. Process five meetings, but reveal the generated drafts only after the manual follow-ups have been written.
• Days 11–14 — Copilot. Process nine eligible meetings. Camille reviews the notes, corrects the draft, adds commercial content herself, chooses the recipient, and sends through her usual email client.

6. Observed results

The results cover all 14 meetings. This small sample supports a continuation decision, not an annual return claim.

One date is attached incorrectly and caught before sending. Four drafts need major restructuring. The ten-minute median saving is useful, but it has not yet recovered the maximum eight setup hours and does not demonstrate higher revenue.

7. Gate decision

Decision: extend for 30 days as an unconnected copilot.

The pre-registered thresholds pass and follow-up becomes more consistent. The rework rate still rules out automatic sending and any extension to complete commercial proposals.

The extension keeps the form and date checklist, measures entry/review/correction time separately, reviews major-rework cases weekly, reruns the frozen set after every model or instruction change, stops on any invented price or commitment, and decides after 30 days whether time actually released justifies the tool.

8. Evidence anchor

Published studies calibrate the scenario but do not supply its results. The BCG–Harvard experiment found gains on some consulting tasks and lower accuracy outside the AI capability frontier. Noy and Zhang studied controlled professional writing tasks rather than an ongoing client relationship. OECD survey data report that one-person firms often perceive a workload reduction, while a 66-organization experiment found less email time but no detected shift in aggregate task quantity or composition.

Sources and transfer limits are recorded in the evidence note.

9. Evidence pack

The pilot retains the mandate, baseline, empty form, data rules, 12 frozen cases, versioned outputs, error log, time measures, and gate decision. Another person can see what was observed, what remains prohibited, and why the pilot continues.