Which AI Scribe Belongs in the Consultation Room?

• The seven-requirement decision framework, from data processing and the data processing agreement to evidence of accuracy and a clean exit
• The omission analysis: why omission, not hallucination, is the dominant failure mode
• The EU AI Act: the phased timeline, the four risk categories, and when a scribe becomes "high risk"
• MDR/IVDR: the medical device question and the route a manufacturer's claim sets in motion
• The GDPR: the DPIA obligation, Articles 9 and 35, and the division of roles between controller and data protection officer
• The WGBO (Dutch Medical Treatment Agreement Act): the record-keeping duty and the twenty-year retention period
• Professional disciplinary law under Article 47 of the Wet BIG (Dutch Individual Healthcare Professions Act), and where liability comes to rest
• Human oversight, and the consent and transparency flow
• Cases from practice: the lethargy case, and "identical technology, different legal status"
• The underestimated pitfalls: secondary use of data and automation bias
• Checklists and worksheets for the negotiating table
• Frequently asked questions, glossary and full accountability, including all 33 sources

Choosing an AI scribe is a governance decision, not a purchase. You do not judge an ambient scribe on the time savings everyone feels in the demo, and not on the ease of use a salesperson shows you. You judge it on governance, on what it demonstrably does not leave out of the patient record. This guide therefore reverses the procurement conversation. Not what this product can do, but what it can cost you if it stays silent at the wrong moment. You ask that question before the signature, not after.

The blind spot lies in the asymmetry of reviewing. You check what is there, you read the note back, you correct a wrong word. But nobody checks what is missing, because an omission has no place on the page and draws no attention. The dominant failure mode of an AI scribe is not the visible fabrication, it is the silent omission, and it only surfaces once you can no longer look back. This guide makes that blind spot visible and gives you the instrument to force that silence into the open before the signature.

What the guide gives you is a defensible decision, not a guarantee. It does not promise you a faultless scribe and it does not remove the risks of the technology. It gives your choice a defensible foundation at the board table, with your data protection officer and, if it comes to that, before the disciplinary tribunal, because you built it on governance rather than on a promise. What the disciplinary tribunal ultimately assesses is your own conduct, including the reviewing and correcting of the note. The guide strengthens the reasoning behind your decision, it does not replace that assessment.

• The seven-requirement decision framework: the core of the guide, your procurement instrument for the negotiating table. Seven concrete requirements against which you hold a vendor to account, from data processing and the data processing agreement to evidence of accuracy and a clean exit. For each requirement, the question you ask verbatim, the reason for it in law or in risk, and how you make it part of the purchase. The framework shifts the burden of proof for the product claims away from your signature and onto the vendor.
• The omission analysis: why omission, not hallucination, is the most dangerous failure mode. With the hard figures from the research: up to 76.3% of the errors were omissions, and only 35.8% of the correctly reported elements remained consistently correct across all platforms. Every figure appears in the guide with its source.
• The legal landscape, made workable: the EU AI Act and the question of when a scribe becomes "high risk", MDR/IVDR and the question of whether the scribe is a medical device, the GDPR with the DPIA obligation and Articles 9 and 35, the WGBO (Dutch Medical Treatment Agreement Act) with its record-keeping duty and the twenty-year retention period, and professional disciplinary law under Article 47 of the Wet BIG (Dutch Individual Healthcare Professions Act).
• Five figures that bring order to the material: the phased timeline of the EU AI Act, the four risk categories, the MDR/IVDR route, human oversight, and the consent and transparency flow.
• 33 verifiable sources: every statement in the guide rests on one of 33 real, traceable sources, [S1] through to [S33]. No invented authorities, no free-floating quotations.
• Checklists and worksheets: ready for the negotiating table, so that you can hold the seven-requirement framework directly against a quotation.
• Cases from practice: the lethargy case, in which one platform left out the finding of lethargy while other platforms did not, precisely the variation between scribes that makes evidence of accuracy non-negotiable, and the case of "identical technology, different legal status", in which two products built on the same technology acquire a different legal status depending on what the manufacturer claims.
• The underestimated pitfalls: the secondary use of data, meaning reuse for model training or onward sale, and automation bias, the tendency to trust output precisely because it comes from a system.
• Reference material for the whole decision: frequently asked questions, a glossary and a full accountability section, so that everyone around the table speaks the same language.

This guide is for everyone who takes part in the decision, or shares responsibility for it, when an AI scribe is introduced. For directors and practice owners who put their signature under the contract. For quality and privacy officers and the data protection officer, who prepare the DPIA and advise on it, since the duty to carry out the DPIA itself (Article 35 of the GDPR) rests on the controller, while the data protection officer advises and supervises. For procurement staff and information managers who select the vendor and assess the data processing agreement. And for the clinician who ultimately signs the note: the physician, the dentist, the pharmacist, the healthcare psychologist, the psychotherapist, the physiotherapist, the midwife, the nurse.

Because that is where the variable sits that nobody can delegate. The model writes the note, you put your signature beneath it, and under Article 47 of the Wet BIG (Dutch Individual Healthcare Professions Act) the liability remains with you, not with the model. A system can leave out a finding, but a disciplinary tribunal does not assess a system, it assesses the practitioner who signed. This guide makes sure you are not signing in the dark.

In this market, independence is scarce, and that is what this guide rests on: no vendor has paid for inclusion, and every judgement stands apart from whoever sells the product. The question is who gets the last word on a scribe, the salesperson with the demonstration, or an independent reading of the evidence. Here, it is the latter.

Every statement rests on one of 33 verifiable sources, not on hype, not on invented authorities or quotations. The picture that emerges from them is unambiguous: the regulators insist, without exception, on the human above the machine. And the landscape is shifting, the law is moving, and practice is running ahead of oversight.

Where the market mostly shows demonstrations and promises of time saved, this guide makes the consensus workable, calmly and precisely, so that you can decide on the basis of what is established.

All 33 sources are real and traceable, [S1] through to [S33]. The figures on failure modes come from the research and are reproduced with their source: up to 76.3% of the errors were omissions, and only 35.8% of the correctly reported elements remained consistently correct across all platforms. The adoption picture is the same: roughly one in three practices now works with a scribe, while the pace of adoption runs ahead of validation and oversight, and fewer than half of clinicians know their own professional body's position on liability. The signature is put down long before the governance is settled. This guide closes that gap. Inclusion was never for sale: no vendor had any influence on the assessment.