Not the clinicians, at least not first. In a hospital, agents usually start in patient access, medical records and coding, so the team leads there train first, alongside the clinician responsible for digital clinical safety and the information governance lead. The executive owner needs a short grounding too. Wider clinical staff follow once the agent touches their work.
Why start outside the wards?
Because that is where agents go first. Booking outpatient appointments, chasing referral information, drafting clinic letters from dictation, suggesting codes for completed episodes: this is high-volume work with clear rules, and it is where hospitals put their first agents. The people who approve those agents' held actions are patient access supervisors, medical records managers and coding leads.
Clinicians matter enormously, but most of them will first meet the agent as a draft letter in their signing queue. They need to know what to check. They do not need to be certified reviewers on day one.
Who goes in the first wave, and why?
| Person | Why first | Training |
|---|---|---|
| Patient access supervisor | Approves held bookings, cancellations and urgent referrals | Reviewer track with practicum on the hospital's own bookings |
| Medical records and coding leads | Approve agent coding suggestions and record updates | Reviewer track with practicum on recent episodes |
| Clinical safety lead | Decides which agent actions carry clinical risk and sets red lines | Common grounding plus how the record shows each decision |
| Information governance lead | Answers for how patient data reaches the agent and who sees it | Common grounding plus provenance and the record |
| Executive owner | Accepts the risk and reports to the board | Common grounding |
Second wave: the clinicians who sign agent-drafted letters, ward clerks, and the digital team who build agents. Third wave: everyone else whose work the agent will touch.
What may the agent do, and what stays with staff?
- Depth 1 to 2: read referral letters and clinic lists; draft bookings, letters and coding suggestions for a person to approve.
- Depth 3: book routine follow-up appointments into set clinic templates and send reminders.
- Stays with a person: triage of urgent referrals, any change to a clinical record, any message giving a patient clinical information, and any safeguarding concern.
Depth 4 has little place here; a hospital's limits should say so.
An example: a referral that should not have waited
Take a hospital's outpatient booking team, where an agent books routine referrals into the next available slot. A referral letter from a family doctor arrives marked routine, but the text mentions unexplained weight loss and a lump. The hospital's own limits make those phrases a red line, so the agent drafts the booking and holds it.
A trained patient access supervisor sees the hold reason and passes the letter to the consultant's secretary for clinical triage that day. An untrained one sees a routine referral with a slot already found and approves. The patient waits weeks. This is why the patient access supervisor is in the first wave, not the consultant: the supervisor sees the hold first.
What makes the first wave ready?
A written test, then a supervised period of real held decisions from their own service, with no missed red line, signed off by someone outside their team. Health data is special-category data under GDPR and falls under HIPAA in the US, so the first wave should also show they can spot agent output that exposes more patient information than the task needs. This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Agent certificates lapse after 90 days and a run that clears no depth drops an agent to depth 1 the same day. The first wave must be able to run bookings and coding by hand when that happens.
A training needs analysis template for AI agents helps confirm the first wave in your own hospital. Then run a readiness assessment in a hospital and roll the training out across the hospital. The answer differs in a bank, where accountability is personal by regulation. The full approach is in training teams for AI agents.
Where the Academy fits
Pryme Intelligence Academy's Operator track suits patient access supervisors and coding leads: about 14 hours of study, a 40-question exam passed at 32 of 40, and two weeks of real review decisions signed off by a Pryme assessor. Lessons can be listened to as audio, which helps staff on rotating shifts.
Questions people ask
When should consultants be trained?
When the agent starts drafting documents they sign. A short session on reading provenance and checking drafts is enough for most.
Should the digital team be in the first wave?
Those who build the agent need the builder track before it is configured. They are usually trained in parallel with the first reviewers.
How many supervisors should be certified per service?
Enough to cover every session when the agent runs, plus one. Held referrals should never wait for a single person to return from leave.