Yes for administrative fields, no for clinical ones. An AI agent can update administrative details when the source is verified and the old value is kept. Clinical entries, including allergies, medications, diagnoses and results, should be drafted by the agent and signed by a clinician, because a wrong entry is acted on before anyone can reverse it.
Why is a patient record harder to undo than it looks?
In the database, almost every change is reversible. Records systems keep versions, and an amendment can be struck through and corrected. That is why teams sometimes treat record edits as low risk.
The risk is not in the data. It is in what people do with it. A wrong allergy entry is read by a prescriber the same afternoon; a dropped medication is missed at the next ward round. By the time the field is corrected, the decision built on it has already been made. Our guide to why autonomy in regulated industries is a question about reversal makes the general case: judge an action by whether its consequences can be reversed, not whether its row can.
So a patient record holds two kinds of field. Some are read by schedulers and billing teams, where a mistake costs a phone call. Others are read by clinicians making treatment decisions, where a mistake can cause harm that no version history repairs.
What can the agent do, and what stays with a person?
| Change | The agent | Stays with a person |
|---|---|---|
| Contact details, preferred name, communication preferences | Updates them at depth 3 when the request comes from a verified patient channel | Disputed identity, or a request that conflicts with the record |
| Appointment and administrative notes | Writes them at depth 3 | Nothing routine |
| Medication list reconciliation | Drafts the reconciled list at depth 2, showing each source | The clinician who accepts or rejects each line |
| Allergies, diagnoses, problem list | Drafts the entry from a discharge letter or referral | The clinician signs it |
| Results and clinical observations | Reads them and flags gaps or mismatches | Filing, interpreting and acting on them |
| Deleting or merging records | Flags suspected duplicates | A records officer decides; the agent never deletes |
Two rows need comment. Merging duplicate records looks administrative, but a wrong merge mixes two people's histories, and unpicking it takes days of clinical review. And a patient's request to correct clinical content is a rights matter under most data protection regimes, so the response belongs to a person even when the agent prepares it.
A worked example
Take a community clinic that receives discharge letters from the local hospital. An agent reads each letter on arrival. One letter says the patient developed a rash on a penicillin-class antibiotic and lists a new allergy.
The agent drafts an allergy entry, quotes the line from the letter, and records the letter as its source. The entry is held for the patient's named doctor, with the draft and the source side by side. The doctor checks, signs, and the entry goes live. In the same pass the agent updates the patient's mobile number from a verified portal request at depth 3, keeping the old number in the history.
Both actions sit in the record. The first shows who signed it and when. The second shows what changed, from which channel, under which bound.
What controls make it safe?
- Field-level scope. The agent's role scope lists the fields it may write, not just the system it may reach. Clinical fields are draft-only.
- Provenance on every source. A message from a verified patient portal is a different class of evidence from a line in an unverified email. The agent records which one it relied on.
- Amend, never overwrite. Every change keeps the prior value. Deletion is not a capability the agent holds.
- Named holds. A clinical draft waits for a specific clinician or role, not a shared queue nobody owns.
- A reconstructable record. Anyone reviewing an incident should be able to reconstruct the agent's decision from the record: source, verdict, bound and person.
- Change control for scope. Adding a field to the agent's write list is a governed change. Use the autonomy change request template and the process in how to run change control for AI agents.
Health data is special-category data under GDPR and protected health information under HIPAA in the United States. Both point the same way: limit what the agent can touch to what the job needs.
Where Pryme fits
In Pryme, writing a clinical field is a separate capability from writing a contact field. Each is certified on the organisation's own cases and carries its own depth. The minimum chain checks role scope, certificate, sources, the customer's limits and platform guardrails at the moment of the write, and a clinical draft is held for a named person. Every verdict is filed in an append-only record.
Questions people ask
Can an agent write up a consultation into the record?
It can draft the note from a recording or dictation. The clinician reviews and signs it before it becomes part of the record, because the note is the basis for later clinical decisions and the clinician is accountable for it.
What if the agent finds an error in a clinical entry?
It flags the error with its evidence and drafts the correction. A clinician decides whether to amend, and the original entry stays visible in the history.
Does keeping a version history make unattended edits safe?
Not on its own. The test is whether anyone acts on the field before a mistake would be caught. For clinical fields, someone usually does.