A try-first rule for the whole team: how to keep skills at work when everyone uses AI

Why a team needs a rule at all

Individually, using AI for everything looks like speed. As a team, you can end up with nobody who can do the task unaided, check the AI's work, or train the next hire. The evidence is early, but it points one way.

• Trust in AI goes with less checking. In a survey of 319 knowledge workers (Lee and others, Microsoft Research and Carnegie Mellon, CHI 2025, peer reviewed), more confidence in the AI was linked to less critical thinking about its output, and more confidence in one's own skill was linked to more. It is self-reported and shows a link, not a cause.
• Unaided performance can slip. At four endoscopy centres in Poland (Budzyn and others, Lancet Gastroenterology & Hepatology, 2025, peer reviewed, observational), experienced doctors' detection rate in colonoscopies done without AI fell from 28.4% to 22.4% after AI assistance was introduced, across 1,443 procedures. Other factors could explain part of it, and it says nothing about AI-assisted results.
• New skills form less well when the task is handed over. In a trial with 52 mostly junior developers (Shen and Tamkin, Anthropic, 2026, a preprint), the AI group scored about 50% on a skills quiz against about 67% without AI. Those who asked the AI conceptual questions kept learning.

None of this says "ban AI at work". It says: decide on purpose which skills you want people to keep, and protect those.

Set it up in one meeting

1. List the skills you cannot afford to lose

   Ask: if the AI were down for a week, or wrong, what must we still be able to do? Diagnosing a fault, pricing a deal, writing to an upset customer. Keep the list short: five items is plenty.

2. Sort tasks into two columns

   "Try first" for work that uses those skills. "Just ask" for chores: formatting, meeting summaries, boilerplate, lookups. Being generous with the second column is what keeps the first one honest.

3. Define the attempt

   Small and written. Two or three lines: my read of the problem, what I would do, what I am unsure of. It takes two minutes. It then goes to the AI with the question, so the AI is checking a view and not supplying one.

4. Change one review question

   In code review, document review and one-to-ones, ask: "What was your first take, and what did the AI change?" Nobody is in trouble for either answer.

5. One line per accepted output

   When AI output goes into shared work, the person adds one line on why it is right: in the commit message, the comment or the ticket. If they cannot, it is not ready.

6. One unassisted rep a month

   Each person does one small real task from the try-first column with no AI. It is a gauge, not a test. If it felt much harder than last quarter, that is useful to know early.

7. Review the rule after six weeks

   Ask the team what to move between columns.

What to avoid

• No monitoring. Do not count people's prompts or read their chats. The moment the rule feels like tracking, people hide their AI use, and then you can see nothing.
• No blanket ban. People switch tools or use personal accounts.
• No shame for "just ask". If someone is on a deadline, the answer is the answer. The rule is about the default, not about every case.
• Juniors first. People still building a skill have the most to lose from handing it over, and the most to gain from a senior who asks "what have you tried?"

Wording people can paste

Each person can make their own AI ask for their take first. This is the work line from our coach instruction:

I use you at work. For thinking tasks (plans, analysis, decisions, messages that matter) ask what I think first. For routine chores, just help.

The free coach setup builds the full instruction (choose Work) and shows where to paste it in each AI. Check your company's AI policy first: it decides which tools and accounts may be used for work data.

Questions