Keep Everyone's AI Skills Current, and Show It When Asked

T13-L05 · Adoption & enablement · Level 5 Operator · 17-minute read, plus 4 hours of doing

An attendance list from January

Last winter the institute rolled out its hosted assistant to all eight research groups (T13-L04). Everyone was invited to a 60-minute session in January. There's an attendance list. Nine months later, a funder's questionnaire asks how the institute meets the EU AI Act's rule on AI literacy, meaning the skills and understanding people need to use AI sensibly at work. The coordinator opens the list and sees the problem at once.

Since January, 34 people have joined: PhD students, guests, a new group. None of them was at the session. Two groups now use paid external models, and one group runs a transcription workstation for patient interviews (T09-L04); nobody was trained on either. The assistant's menu has gained a document feature that stores files. And three people on the list of those allowed to use external models left in the summer.

In a company it's the sales department six months after its Copilot rollout. 22 people have seats, the rest of the company has the free chat, and the works council asks how people are trained, and whether the training records are being used to judge anyone.

You are at Level 5 Operator. You run AI for everyone, over years, while people and tools change underneath. This book shows you how to keep their skills current and show it, with a register that doesn't depend on you.

After this you can

• Map each role to what its people must be able to do with the AI tools they use, written as tasks you can check.
• Check those abilities with short scenario tests on real work, and record the result instead of attendance.
• Keep one register that stays in step with arrivals, role changes, departures and access, checked every month.
• Trigger a refresher when a tool, rule or risk changes, and record the decision even when none is needed.
• Show an auditor, the works council or a funder what you do and why, without your register turning into a file on people's performance.

Before you start

• Books: T13-L04 for the rollout, its access groups and the works or staff council agreement. T13-L03 for the one-page working rule and its eight test cards, which this book reuses as checks. T12-L05 for the list of AI systems your organisation runs. The four-quadrant method is in T01-L01.
• Your organisation's learning platform: Moodle at many universities and institutes, or Microsoft Forms and a SharePoint list in a Microsoft 365 company. Use what already exists.
• The access lists for each AI tool (the licence groups, the list of external-model users) and the monthly list of arrivals and departures from HR or administration.
• The works council or staff council, and the data protection officer: a training register is personnel data.
• About 4¼ hours: 17 minutes to read, about 4 hours of doing, over three weeks.

Last verified 23 September 2026. Since the 2026 amendment, the EU AI Act asks organisations to take measures to support their staff's AI literacy, without guaranteeing any level. No certificate is required; the Commission suggests keeping an internal record. Supervision started on 2 August 2026. Details for experts at the end of this book.

The idea in one page

Start from work, not from courses. For each role, write down which AI tools its people use, on which data, for which tasks, and what could go wrong. Then write what they must be able to do, as tasks someone can watch: "sorts eight cases into green, yellow and red and stops every red one", not "understands AI". Green, yellow and red are the working rule's three kinds of material from T13-L03: fine to use, only through the approved route, and never. A PI, the lead researcher of a group, needs to decide when to stop a tool; a PhD student needs to check every reference in a draft; the person running the transcription workstation needs to stop the job and delete audio on schedule. Each role gets only what its work requires.

Check the ability, not the presence. An attendance list shows who was in the room. A short check on the role's own tasks shows who can do them. The eight test cards from T13-L03 are a good start: a piece of material and a task, with the right answer known in advance. Add one or two role-specific cases, such as a draft with a wrong figure to catch, or a stop to carry out. Record the result, the date, the check's version and who ran it.

One register, two views. The role map says what each role must be able to do and why. The person record says, for each person, which role they have, which checks they've passed, when, and when they're due again. Keep both in one place the organisation owns, with a deputy who can run it.

Keep it in step with people and access. Once a month, compare three lists: who arrived, changed role or left (from HR or administration); who has access to each AI tool (from IT); and who has a current check (from the register). Three kinds of mismatch show up: access without a current check, a check without access (harmless), and access for someone who has left (fix today). This comparison is called a reconciliation, and it's what keeps the register true after you've built it.

Refresh on change, not only by calendar. When a tool, a rule or a risk changes, decide what it means for each role, and write the decision down. A new button: a two-line note. A new kind of data, a new action the tool can take, or a new stop rule: a short check again before people use it. No change in what people must do: record "reviewed, no refresher". A yearly review catches what the change log missed.

A record, not a surveillance file. The register holds what's needed to show that people can do the task: role, check, date, result, next due. No prompts, no usage figures, no comments on performance. The works council or staff council agrees what's in it, who can see it and how long it's kept. A missed check leads to an offer of a session within a week, never to a note in someone's file.

What you show. For an auditor, a funder or the works council: the role map, the register, the change log with refresh decisions, and the reconciliation log. Three questions should be answerable in minutes: what must this role be able to do, has this person shown it recently, and what did you do when the tool last changed?

Worked example: from an attendance list to a register, four quadrants

Both versions follow the four quadrants from T01-L01. The lab one is told in full. Names, numbers and dates are illustrations.

Lab: eight groups and a steady stream of new people

Structure it. Dr Paulsen, the institute's research coordinator, writes the problem down. "Every person who uses an AI tool at the institute can show, within the last year, that they can do what their role requires with it. When a tool changes, the people it affects get what they need within a month. The register keeps working when I'm away and when a quarter of the staff change each year."

Understand it. She starts with who uses what, not with training. From T13-L04's rollout plan, the list of AI systems the institute runs and one question to each group leader, she lists the roles:

• Everyone uses Chat AI, the institute's hosted assistant, under the one-page rule.
• External-model users in two groups use paid models for manuscript editing, published material only, with their PI's signature.
• Clinical groups' staff work near restricted patient data and must never let it reach any AI tool.
• The transcription operator and deputy run the workstation from T09-L04: stop the job, delete audio, redo a draft.
• PIs approve external-model use, check outputs in their group and decide when to stop a tool.

Then she compares three lists: the January attendance list, the administration's list of 34 arrivals and 21 departures since, and IT's list of external-model users. Four things stand out. Nobody who joined after January has any record. Three departed people are still on the external-model list. The transcription operator's deputy has never run the stop procedure. And the assistant's new document feature, which keeps uploaded files by design, arrived in June without anyone deciding what it meant.

She also talks to the staff council, the elected representatives of the institute's employees. They support the training and ask one thing: that the results are never used in appraisals or contract extensions.

The finding that changes her problem is the arrivals list. She had been planning a better training session. But a quarter of the institute changes every year, so any one-off session goes stale within months. Back to Structure it: "Every arrival gets their role's check in their first two weeks, as part of onboarding. The register is reconciled with the arrivals, departures and access lists every month. Changes to the tools go through a written refresh decision."

Choose tools.

• Moodle, the institute's learning platform, because every member of staff already has an account through the institute login and it records completions and quiz results. One short course per role: the one-page rule, two worked examples, and the check as a quiz, one multiple-choice question per card with green, yellow and red as the answers. The alternative is a Microsoft Forms quiz with a SharePoint list, which is what the company in this book uses.
• The eight test cards from T13-L03, adapted per role, as the check. Pass rule: every red card stopped, and at least seven of eight right.
• A practical check for the transcription operator and deputy, done in person: carry out the stop-and-delete procedure from the run sheet while someone watches.
• One spreadsheet on the coordination office's SharePoint site for the role map, the change log and the reconciliation log, next to the Moodle export.

Build it: the role map. One row per role, with the abilities written as checkable tasks:

Each row has a reason next to it, and the director signs the map.

Build it: the checks and the register. The Moodle course for Everyone takes about 30 minutes: the one-page rule, two short worked examples from the institute's own work, and the eight cards as a quiz. Moodle records the date and result. The quiz's Results page has a Grades report with the mark for each question, so the coordination office can see that every red card was stopped, and downloads it as a spreadsheet each month. The person record in the spreadsheet has one line per person and role:

staff ID | role | check | check version | date | result | next due | access reconciled on

No names of groups' projects, no quiz answers, no prompts. The staff council reads the column list and agrees it, along with who can see the register (the coordination office and the person themselves) and how long lines are kept after someone leaves (one year, then deleted). The DPO adds the register to the institute's record of processing, the list of every use of personal data that the GDPR requires.

Build it: the monthly reconciliation. On the first working day of each month, the coordination office's assistant, Jan, runs the comparison. He exports the administration's arrivals, role changes and departures, IT's external-model list, and the Moodle Grades report, and pastes each into its own sheet of the spreadsheet.

Every list carries the staff ID, so one lookup formula per comparison finds the mismatches. For example, next to each person on IT's access list, =XLOOKUP(A2, Register!A:A, Register!E:E, "no check") fetches the date of their check from the register sheet, or says no check. The three comparisons are access against checks, access against departures, and role changes against the role each check was for. The first run finds:

• 34 people with access and no check: all get the course link and two weeks. Thirty complete it; four are away, and get a date.
• 3 departed people on the external-model list: IT removes them the same day.
• 2 people who changed from the clinical group to another group: their clinical role line is closed, their new role's check is due.

Each run gets one line in the reconciliation log: date, what was found, what was done, and who did it.

Understand it: the refresh decision. Paulsen goes back to June's change, the document feature. The question isn't "do we need training?" but "does this change what any role must be able to do?" It does: people could now upload files that stay stored. The rule gets a line (published papers only), and the Everyone check gets one new card with a file to upload. The change log records it:

2026-06-15 | Chat AI document feature (keeps files) | Changes what Everyone must do: yes
Decision: new rule line; new card 9 in Everyone check; existing users: 2-page note
and card 9 as a 5-minute quiz within 4 weeks | Decided by: coordinator, DPO | Done: 2026-10-10

A month later the assistant's menu gets two new self-hosted models. The decision: reviewed, no refresher; nothing changes in what anyone must do. One line in the log.

Understand it: the deputy test. The transcription deputy does the practical check with the operator watching, from the run sheet alone, and passes. Then Paulsen asks Jan, her deputy for the register, three questions an auditor might ask, and he answers them from the register without her: what must a clinical group member be able to do (the role map), has this new PhD student shown it (her line, passed on 3 October), and what did you do when the tool last changed (the June log entry). It takes him six minutes. The funder's questionnaire gets the same three answers, and the role map as an attachment.

What it taught her: the courses took two days. What kept the register true was the monthly comparison of three lists, and it takes Jan twenty minutes.

Company: 22 seats, 180 people, one works agreement

Structure it. Sven, from the T13-L04 rollout, writes: "Everyone with a Copilot seat or the free chat can show what their role requires, and the works council can see that the records are used for nothing else."

Understand it. Two roles cover most people: Copilot seat holders (the 22 in sales) and Chat users (everyone else). The Copilot rollout's works agreement already says usage isn't used to judge people; the council now asks the same for training records. And IT's licence groups from T13-L04 give a clean access list, so the reconciliation is easy to automate.

Choose tools and Build it. A Microsoft Forms quiz per role, with the cards from the sales team's rule plus two Copilot-specific cases: a summary that includes a file the person shouldn't see (the right answer is to report it), and a customer draft to rewrite. Results go into a SharePoint list, the register, readable by Sven, his deputy and the person themselves. The works council agrees the columns, a one-year deletion after leaving, and one sentence: a missing or failed check leads to a session within a week and a pause of the seat until it's done, and to nothing else. The monthly reconciliation compares the two licence groups with the register and HR's leavers list. In month two it finds a seat holder whose check lapsed during parental leave; the seat is paused on her return until she's redone the quiz, which takes her twenty minutes. When Microsoft adds agents, small helpers colleagues can build inside Copilot, the refresh decision is "changes what seat holders can do: yes"; IT keeps agents switched off until the rule has a line on them and the quiz has a card.

In both places the register stayed true for one reason: someone other than its author compared it with the real lists every month, and every change to a tool got a written decision.

What goes wrong

Attendance taken for ability. Symptom: a list of who came to a session, and nobody can say what they're able to do. Fix: a short check on the role's own tasks, with a known right answer and a pass rule. Record the result, not the presence.

One course for everyone. Symptom: the PI sits through a prompt-writing session, and the workstation operator never practises a stop. Fix: a role map that gives each role only what its work requires, with a reason per line.

The register goes stale. Symptom: new arrivals aren't in it, and people who left still have access. Fix: the monthly reconciliation of arrivals and departures, access lists and check results, done by a deputy, with one log line per run.

Refreshers by calendar only. Symptom: a tool gains a feature that stores files, and nobody is told for a year. Fix: every change goes through a written refresh decision, including "reviewed, no refresher".

A training register that becomes a performance file. Symptom: quiz scores discussed in appraisals, and people stop taking the checks honestly. Fix: agree the columns, readers and deletion period with the works or staff council, and write down that a missed check leads to a session, not a consequence.

Too much in the register. Symptom: prompts, free-text comments or usage figures next to people's names. Fix: role, check, version, date, result, next due. Everything else stays out.

The register lives with you. Symptom: the funder's questionnaire waits until you're back from holiday. Fix: the register in a place the organisation owns, a deputy who runs the monthly comparison, and the three auditor questions tested on the deputy.

Do it yourself (4 hours, over three weeks)

1. Structure it (15 min). Write who uses which AI tools, how often people join and leave, and what you must be able to show and to whom.
2. Understand it (45 min). List the roles from your AI system list and rollout plans. Compare your current training records with the arrivals, departures and access lists. Talk to the works or staff council. Rewrite your problem sentence.
3. Choose tools (15 min). Pick the learning platform your organisation already has, and where the role map and logs live. Name your deputy.
4. Build it: role map (45 min). One row per role: tools and data, what they must be able to do as checkable tasks, the check, the pass rule, when it's due again, and the reason. Get it signed.
5. Build it: checks and register (60 min). Adapt the T13-L03 cards per role and put them into a quiz. Add practical checks where a role must carry out a procedure. Set up the register's columns, and agree them, the readers and the deletion period with the works or staff council and the DPO.
6. Build it: reconciliation (30 min). Run the first monthly comparison of the three lists with your deputy. Fix what it finds, and log it.
7. Understand it: refresh and deputy test (30 min). Put the last tool change through a written refresh decision. Ask your deputy the three auditor questions and time the answers.

Exit check

Written finding: a role map and a training register for one real department, with:

• the role map: each role's tools and data, what its people must be able to do as checkable tasks, the check, the pass rule and when it's due again, with a reason per line and a signature;
• the register: one line per person and role, with check, version, date, result and next due, and the columns, readers and deletion period agreed with the works or staff council;
• one monthly reconciliation of arrivals and departures, access lists and check results, with what it found and what was done;
• one change put through the refresh decision, including what was decided for each role.

It passes when every person with access to an AI tool has a current check result or a dated reason why not, and a deputy answers the three auditor questions from the register alone.

Rule to remember

Record what people can do, and check the register against who's really there every month.

Details for experts

Checked on 23 September 2026. You don't need this to follow the book; it's for setting things up or answering a colleague. It isn't legal advice.

Further reading and tools

Next books

• T13-L04 · Roll Out an AI Assistant to a Whole Department, With Access, Cost and Consent Settled First: the rollout, access groups and works agreement this book builds on.
• T13-L03 · Write a One-Page AI Rule Your Team Can Use Without Asking You: the rule and test cards behind every check.
• T12-L05 · Keep One List of Every AI System Your Team Runs, and Make It Survive You: the organisation's list of AI systems, which the role map starts from.
• T15-L05 · Keep Your Group's AI Use Compliant and Running After You Leave: the same register-and-successor pattern for one research group.
• T09-L04 · Transcribe Recordings That Must Stay in the Building: the workstation whose operators need a practical check.

Tools

• Microsoft Copilot: Catalogued. The assistant in the company example; its licence groups feed the reconciliation.
• Langdock: Catalogued. A managed assistant some organisations approve; the same register applies.