Use this before the firm gives anyone an AI assistant for their work, and again before each new group gets one. It settles who goes first, which tasks are measured and how, and how the firm watches the quality of its most experienced people's work. The order follows Brynjolfsson, Li, and Raymond ("Generative AI at Work," The Quarterly Journal of Economics 140, no. 2, 2025, pages 889 to 942), a study of customer-support agents at one software firm; the measures are the firm's own.
Firm / scope (people, teams, the assistant's intended uses):
Prepared by / date:
Owner (who decides at each review whether to widen, hold, or withdraw):
Assistant and service used, its terms for retention and training on inputs, and what it may access:
Before anyone starts
- [ ] The firm has decided which documents, accounts, and systems the assistant may see, and whose access it uses.
- [ ] Each person in the rollout knows the assistant's output is a draft they own, and what may not go into it.
- [ ] Every task below has a baseline recorded for six to eight weeks before access.
Baseline period: Rollout start:
Who goes first
List people in the order they get access. Start with the newest people on repeatable written work, then experienced people on tasks they do rarely, then senior people on their core work.
| Order | Person and role | Months at the firm | Tasks the assistant is for | Group (new / rare task / senior core) | Access date |
|---|---|---|---|---|---|
| 1 | |||||
| 2 | |||||
| 3 | |||||
| 4 | |||||
| 5 |
- [ ] No one is first because they asked first or because they are senior.
- [ ] Senior people's core work starts only with the quality watch below in place.
What the firm measures
Two measures per task, both countable without new tools. Turnaround is days from request to a reviewed draft. Review findings are the corrections a reviewer makes.
| Task | Who does it | Turnaround before (median days) | Turnaround after | Review findings before (per piece) | Review findings after |
|---|---|---|---|---|---|
Who records the measures, and where:
- [ ] The same reviewer, or the same checklist, is used before and after, so findings are comparable.
- [ ] The firm reads the results as a structured look at a few people, not as a statistic.
Quality watch for the most experienced people
| Senior person | Peer reader | Pieces read this month (two) | Findings against the checklist | Unassisted piece this quarter (task, date) | Drafts sent largely as the assistant wrote them |
|---|---|---|---|---|---|
- [ ] Each month, a peer reads two pieces of each senior person's assisted work against the checklist the firm used before.
- [ ] Each quarter, each senior person writes one of each recurring piece without the assistant.
- [ ] The file notes when a draft went out largely as the assistant wrote it.
- [ ] The firm's library of example work keeps the senior person's own versions, not the assisted ones.
Checklist used by peer readers (name and version):
Newer staff
- [ ] Each quarter, each newer person writes one piece without the assistant, read by the same reviewer as their assisted work.
What the unassisted pieces show:
Review
90-day review date:
Decision for each task: widen access / hold / withdraw for this task
Reasons (from the measures and findings):
Approved by / date:
Next review (at least every six months, and after any change of assistant or its access):
Common failure modes
- Seats go to the most senior people first, and nobody measures anything.
- There is no baseline, so the 90-day review compares the assisted work with memory.
- Senior people's drafts are never read by anyone, so a slip in quality shows up first outside the firm.
- The example library fills with assisted drafts, and the firm's best work stops being anyone's own.
- Newer staff produce good drafts quickly and never write one alone.
This plan decides who gets an assistant and how its effect on the work is watched. It does not assess the service's security or contract terms; use the firm's vendor review for that.