Your agents are on the ward. Each one answers to a clinician.

Agents are drafting discharge letters, reconciling medicines and answering patients. TraceMem ties each one to the clinician it works for, shows it only what its task needs, and holds anything sent to a patient or outside the hospital until that clinician signs.

Or call us on +1 650 550 1418

Ward 7B

Handover 14:00
Bed 3 · discharge-agenttok_4k9w

Works for

Dr Amara Okafor, consultant

Verified by Microsoft Entra ID

The model saw

Name, NHS number, diagnoses, procedures and medicines as tokens

Its task may read

  • Identity details
  • Diagnoses
  • Procedures
  • Medicines
  • Allergies

Not in its view

  • Blood results
  • Mental health notes
  • Safeguarding record
  • Genetic results

Held for signatureThe letter goes outside the hospital, so it waits for Dr Okafor to sign.

Illustrative ward. Names, beds and times are made up; patients appear only as tokens.


Each task sees only its part of the record.

Minimum necessary, enforced by the data product instead of trusted to the agent. Pick a task: the record lights only what that task may read, and a request for anything else is refused.

Governed Data Access
Patient recordtok_4k9w

As discharge-agent reads it: 5 of 9 sections

  • Identity detailsName, NHS number, date of birth, addressTo the model as tokens
  • DiagnosesCommunity-acquired pneumonia; type 2 diabetesTo the model as tokens
  • ProceduresChest X-ray; intravenous antibioticsTo the model as tokens
  • MedicinesAmoxicillin 500 mg three times a day; metformin 1 gTo the model as tokens
  • AllergiesCodeine (nausea)In view
  • Blood resultsNot in this task’s viewRefused if asked
  • Mental health notesNot in this task’s viewRefused if asked
  • Safeguarding recordNot in this task’s viewRefused if asked
  • Genetic resultsNot in this task’s viewRefused if asked

A task that does need mental health notes can be given them, and they reach the model obfuscated, which cannot be reversed. A genetic result can be set to stop the request altogether.

Illustrative record. The patient appears only as a token.


The rules for AI in care are arriving from every side.

Some of these already apply and some are still coming. Between them they ask who reviewed what an AI produced, what it was allowed to see and what patients were told, and only a record kept at the time can answer.

  1. 1 Jan 2025 · US

    California AB 3030

    Generative AI messages to patients about their care need a disclaimer and a way to reach a person, unless a licensed provider has read and reviewed them.

    Source (opens in a new tab)
  2. 1 Jan 2025 · US

    California SB 1120

    A licensed clinician, not an algorithm, must decide a health plan's medical-necessity denials.

    Source (opens in a new tab)
  3. 26 Mar 2025 · EU

    European Health Data Space in force

    The regulation that will move patient summaries, prescriptions and results across the EU, and open health data for secondary use.

    Source (opens in a new tab)
  4. 27 Apr 2025 · UK

    NHS England guidance on AI scribes

    Ambient scribes need clinical safety assurance, a data protection impact assessment and a technology assessment before use. Last updated July 2026.

    Source (opens in a new tab)
  5. 1 May 2025 · US

    Section 1557 and decision support tools

    Providers must identify and reduce the risk of discrimination in the patient care decision support tools they use.

    Source (opens in a new tab)
  6. 4 Aug 2025 · US

    Illinois limits AI in therapy

    AI may support therapists with admin work, but may not make therapeutic decisions or talk to clients therapeutically. Up to $10,000 per violation.

    Source (opens in a new tab)
  7. 1 Jan 2026 · US

    Texas TRAIGA

    Health care providers must tell patients in writing when they are interacting with AI.

    Source (opens in a new tab)
  8. 5 Feb 2026 · UK

    UK rules on automated decisions apply

    Under the Data (Use and Access) Act, significant automated decisions taken from this date need safeguards: telling the person, human review on request, a right to contest.

    Source (opens in a new tab)
  9. 30 Jun 2026 · UK

    Data Security and Protection Toolkit

    NHS organizations submitted their 2025-26 assessments, now aligned with the NCSC Cyber Assessment Framework.

    Source (opens in a new tab)
  10. 29 Jul 2026 · UK

    MHRA on ambient voice technology

    Scribes intended only to transcribe, summarise or draft for a clinician to review are not medical devices. Ones that support diagnosis, or act without a clinician's review, are.

    Source (opens in a new tab)
  11. 2 Dec 2027 · EU

    High-risk AI in access to care

    AI that decides eligibility for healthcare services, or triages emergency calls, becomes high-risk: human oversight and logs expected.

    Source (opens in a new tab)
  12. 2 Aug 2028 · EU

    AI in medical devices becomes high-risk

    AI inside devices that need third-party assessment under the MDR or IVDR takes on the AI Act's high-risk obligations.

    Source (opens in a new tab)
  13. 26 Mar 2029 · EU

    Health Data Space: first exchanges

    Patient summaries and ePrescriptions start moving across the EU, and secondary use of most health data opens.

    Source (opens in a new tab)

Dates checked against each regulator’s own publication, September 2026. Not legal advice.


Patient data has a file with regulators already.

None of these involved an AI agent. Each was a system a hospital relied on, and each came down to who could see patient data, for what purpose, and whether anyone was checking.

Royal Free NHS Foundation Trust

2015 · ICO, 2017

1.6 million records

ICO finding (opens in a new tab)

Patient records were passed to a technology partner to test a clinical app, and patients were not adequately told how their data would be used.

What it came down to

Health data handed to a new system without a stated, limited purpose, or an impact assessment first.

For an agent, with TraceMem

Every agent read carries a declared purpose, and the model sees only what that task needs, as tokens.

Governed Data Access

HagaZiekenhuis

2018 · Dutch Data Protection Authority, 2019

€460,000

AP decision (opens in a new tab)Court ruling (opens in a new tab)

Staff with no care relationship looked at a well-known patient's record. A court later reduced the fine to €350,000.

What it came down to

No two-factor authentication on patient records, and access logs that were not regularly reviewed.

For an agent, with TraceMem

Every read is on the record with the clinician behind it, and a control that is switched on but seeing nothing is flagged.

Decision Trace

Advanced Computer Software

2022 · ICO, 2025

£3.07 million

ICO fine (opens in a new tab)

Ransomware came in through a customer account, disrupted NHS 111 and exposed the data of 79,404 people.

What it came down to

An account without multi-factor authentication, trusted with health systems.

For an agent, with TraceMem

Every agent action carries a clinician verified by your own directory, on every call, or it does not run.

Verified Identity

Every month you wait, agents read records nobody can account for.

Put in your own numbers. What your agents read and sent before a governed path is in place can never be recorded afterwards.

100
6 months

21 working days a month. Your numbers stay in your browser.

In 6 months of waiting

100,800

Agent actions with no record of which clinician they worked for, what they read or who signed what they sent.

Next 18 monthsNo recordOn file

Put the first ward on the record this month.

Pick whichever is quickest for you. However you get in touch, a person replies. With our forward-deployed engineers, TraceMem can be running inside your network in days.

What would you like?
Where are your agents working?

Pick any that apply.

When do your agents act on production systems?

Optional. A sentence is enough.

45days, free

Free for 45 days on our Docker-based trial image, inside your own network. When you are ready, it upgrades to the full enterprise product.

Forward-deployed engineers, ours or our partners', can set it up in days and help you define policies and integrations. This may be charged separately.