Ambient scribes raise the stakes for clinical safety

A device, not a gadget

Leading ambient scribe platforms, including Accurx Scribe (powered by Tandem), are registered with the MHRA as Class I medical devices, and their manufacturers hold DCB0129 compliance and DTAC documentation.1 NHS England has published explicit requirements for AI scribe technology, and an NHS ambient voice framework now lists reviewed suppliers.2

None of that removes the deployer's obligation: under DCB0160, the organisation switching the scribe on must run its own clinical risk assessment, maintain its own hazard log, and name its own Clinical Safety Officer. The manufacturer's file covers the product. It does not cover your patients, your pathways, your deployment.

The failure modes are clinical, not technical

Vendors are admirably candid about what can go wrong. Scribes can omit clinically important information, confabulate (hallucinate content that was never said) and substitute one term for another.13 In a discharge summary that is an inconvenience. In an ADHD assessment, where the diagnostic signal lives in nuance, collateral history, and what the patient didn't say, an omission is not a typo. It is a corrupted record that downstream clinicians will trust.

That is why the emerging consensus, including the vendors' own published recommendations, is a human-in-the-loop design: the scribe drafts, the clinician verifies, and the clinician remains accountable for accuracy and clinical impressions, supported by training that goes beyond software mechanics into critically appraising AI-generated text.3

Consent is a clinical act too

In mental health and neurodevelopmental services, scribes intersect with vulnerability. Sector analysis has warned that for patients with paranoia or histories of trauma, the knowledge that an unseen tool is listening can be triggering, and that consent must be authentic, voluntary and revocable, with clinicians prepared to forgo the scribe the moment it strains the therapeutic alliance.3 Features like mid-session pause and a commitment never to train models on transcripts exist precisely because these risks are real.3

Deciding when ambient listening is appropriate, patient by patient, is clinical judgement. Governing that judgement, writing the policy around it, and auditing it afterwards is the CSO's job.

What a safe deployment actually requires

• A deployer-side DCB0160 file: hazard log, risk assessment and Clinical Safety Case Report specific to the local service, signed by the CSO before go-live.
• A consent and pausing protocol: how consent is sought, recorded and withdrawn; when the scribe is simply switched off.
• A verification workflow: human-in-the-loop review of every drafted note, with accuracy auditing and drift monitoring over time.
• Clinician training: not just which button to press, but how to detect omission, hallucination and substitution in AI-drafted text.3
• An incident pathway: what happens when an AI-generated error reaches the record, and who reviews it.
• Change control: model updates change behaviour; each material change re-enters the safety process.

Why this lands on the CSO's desk

Every item on that list sits at the junction of clinic and software, exactly the territory the DCB0160 CSO is defined to own, and exactly where a generic compliance officer or IT lead is unqualified to sign. The organisations deploying ambient scribes safely in 2026 are the ones where a trained clinician holds the hazard log, designed the human-in-the-loop workflow, and carries the authority to pause the rollout. That person is not overhead. They are the reason the deployment is defensible: to the CQC, to commissioners, and to patients.