1. The signs
- Usage is reported as logins or prompts, and nobody reports whether the work is better.
- People use their own AI accounts because the approved one is slower, or because nobody said which one is approved.
- Errors in AI-assisted work are found by the client, not by your own review.
- "Can I use AI for this?" gets a different answer from each manager.
2. Why it happens
Most AI rollouts are run as a purchase: pick a tool, buy seats, run a lunch-and-learn. But AI changes how work gets checked, and nothing in a purchase changes that. Among firms reporting real returns from AI, most had redesigned the workflow around it; most of the others hadn't. GOL-ADP-04
Bans and silence make it worse. When there's no sanctioned, data-protected tool, use goes out of sight, and people who use AI for important work are often reluctant to say so. Work you can't see is work nobody checks. GOL-SAI-01 · GOL-SAI-03
And usage hides harm. On a task outside the tool's ability, consultants using AI were 19 percentage points less likely to get it right. A rising usage chart can sit on top of falling quality. GOL-ADP-05
3. What good looks like
- One page of rules people can find: approved tools by kind of data, what never to paste, who stays responsible, and how much checking each kind of work needs. PRM-GOV-01 · PRM-GOV-02
- Verify, then trust. Output is checked by a different method or person before it's used; a model checking itself isn't a check. PRM-VER-01 · PRM-VER-04
- Tested on your own work first: 30 to 50 real tasks with known answers, run before rollout and again after every model change. PRM-VER-05 · KIN-TWN-04
- Answers that cite their source, or say "not found", drawn only from documents you govern. KIN-TWN-01
- Results broken out by experience. AI tends to help less experienced staff most, so give it to them first, with a person responsible for every answer. GOL-AIA-01 · GOL-ADP-06
Seats bought, a launch email, a dashboard of prompts per week. Quality is assumed. The first real test is a client spotting a wrong number.
One workflow redesigned, a gold set of your own cases, a named reviewer for each tier of work, and a usage figure always paired with an outcome figure.
4. Your first three moves
- Write the one-page policy this month, and give people a sanctioned tool alongside any restriction. PRM-GOV-01 · GOL-SAI-01
- Build a gold set of 30 to 50 real tasks from one team, with the right answers, before you widen access. PRM-VER-05
- Pick one workflow and redesign it with the tool in it, then report usage and outcome side by side. GOL-ADP-04 · GOL-ADP-05
5. Stress-test it
The risky moment in an AI rollout is ordinary: a confident, well-formatted answer with one wrong figure in it, on a deadline. A TPC Stress Test puts your team in exactly that moment, with the tools and time pressure they'd really have, and shows who checks, who doesn't, and where the review step quietly disappears. The readout names three fixes you can act on.
The short version. Treat AI as a change in how work is checked, not a purchase. One page of rules, a sanctioned tool, a gold set of your own cases, a reviewer for each tier, and usage always reported next to outcomes.