The better the scribe gets, the weaker your safety net becomes
Every vendor answers the accuracy question the same way, and the answer is correct: the clinician reviews the note and signs it. Attestation is the control this whole category rests on.
It is also a control that weakens in proportion to how good the tool is. People are poor at sustained detection of rare errors in output that is nearly always right. The pattern shows up wherever automation is mostly reliable, in aviation, radiology assistance and driver support systems: as the failure rate falls, the human checker stops finding the failures that remain. Not through carelessness, but because vigilance for rare events is something humans do badly and cannot be trained out of.
So a scribe that is right ninety five percent of the time gets read carefully, and one that is right ninety nine percent of the time gets skimmed. The number of errors reaching the permanent record may not improve at all.
This is not an argument against the technology. It is an argument that the engineering problem is not where most people are looking.
Which is why vendor accuracy benchmarks mislead
Accuracy measured on a generated note, before a human touches it, is the wrong denominator. What matters is the error rate in signed notes, which is a product of model accuracy and review effectiveness, where the second term falls as the first rises.
Ask for that number instead. Most vendors will not have it, because obtaining it means auditing signed notes against what actually happened. That absence is worth knowing before a contract is signed.
Two follow up questions reveal more than a headline figure. Which specialties, accents and interpreter mediated consultations was the evaluation run on, and does that include your population? And how does performance trend over months in a deployed site rather than an evaluation set?
Review is worst at catching exactly the errors that matter
Not all errors are equally detectable, and the split is not random.
Errors that look wrong get caught. A garbled sentence, a drug name that makes no sense, a date that cannot be right, each of these interrupts a reader and gets fixed.
Errors that look right do not, and in clinical documentation those are the dangerous category:
Invented negatives. A symptom recorded as absent when it was never asked about. Nothing in the text signals a problem, because thorough documentation looks exactly like this.
Fabricated normal findings. Examination elements the model expects to accompany a complaint, generated because they usually appear rather than because they were performed.
Attribution drift. A condition described about a parent during family history, recorded as the patient's own.
Plausible dose errors. A strength or frequency that is wrong but inside the range a reader accepts without checking.
Each reads as competent documentation. The reviewer's attention is a filter tuned for implausibility, and these pass straight through it.
Design for review, not for generation
If review is the control, review is the thing to engineer. These are interface decisions rather than model decisions, and they are where a custom layer earns its cost.
Keep notes short. Length is the enemy of scrutiny, so verbosity is a failure mode rather than a feature.
Separate what was said from what was inferred, so a clinician sees at a glance which statements carry interpretation.
Mark low-confidence passages rather than presenting uniform prose. Uniform confidence is a lie the output tells about itself.
Link each assertion back to its moment in the encounter, so a doubtful line takes seconds to check rather than requiring recall.
Do not autofill sections that were never discussed. An empty examination field is safer than a populated one nobody performed.
Instrument the review, and alarm on it
Model accuracy is the wrong thing to monitor in production. Review behaviour is the leading indicator, and your own systems already measure it.
Track the time between a note being presented and signed, the proportion edited before signature, how much changes when edits happen, and the distribution across individual clinicians. Establish the pattern in the first weeks, while people are still reading carefully, and treat a later collapse in sign off latency as the signal it is.
That is the alert worth building. Not a drop in model performance, which you will not see, but the moment your organisation stopped checking.
Two others matter. Billing level distribution, because models write thorough sounding notes, thorough notes support higher levels of service, and drift stays invisible until an audit finds it. And a periodic human audit of signed notes against the clinician's recollection, the only way to get the number that counts.
Audio is protected data, and consent is not a checkbox
Ambient capture creates a recording before it creates a note, and that recording is a separate risk.
Establish in the contract whether audio is retained at all, for how long, who at the vendor can access it, and whether your recordings or the resulting notes train their models. Consultation audio is protected health information, so a business associate agreement is required in the United States and an equivalent arrangement elsewhere, and it does not transfer your obligations. The safest default is processed and discarded.
Consent has a wrinkle specific to this setting. Some jurisdictions require the consent of everyone recorded rather than one party, and a consultation room holds people who were never asked: family members, carers, interpreters, sometimes another patient audible beyond a curtain.
Independent security review should cover the capture path, the storage and the record system integration, not the model alone.
Where a distributed ledger does not help
We build blockchain systems, so this is worth stating plainly. The pitch that surfaces here is an immutable record of consent or note provenance, and it misreads the requirement twice.
Clinical records must be amendable. Corrections, addenda and late results are routine, and regulations generally require a record can be changed with the change and its author visible. Immutability works against that, and record systems already keep audit trails regulators accept.
The hard part of consent is obtaining it meaningfully from everyone in the room and honouring a withdrawal across every copy of the data. That is a process problem, not a storage problem.
Buy the engine, build the layer that protects you
For most organisations the engine should be bought. Vendors have speech models tuned to clinical vocabulary, certified integrations with the major record systems and validation work that is expensive to replicate, and building to parity is a poor use of budget.
What is worth building is everything in this article: the review interface, the instrumentation, the audit sampling, the monitoring that tells you when checking stopped. Those are specific to how your clinicians work, no vendor will prioritise them over their own roadmap, and they separate a deployment that is safe from one that merely looks safe.
If you commission that work, hold the pipelines, configuration and documentation at the end, with another team able to operate it without the original supplier. A control you cannot inspect is not a control.
RWaltz is a blockchain and enterprise software development company building custom smart contracts, dApps, and tokenization platforms that integrate with existing business systems. We work to a build-to-own model: clients hold their keys, repositories, and intellectual property; engagements are scoped honestly, including the cases where buying is the better answer; and security review is treated as continuous rather than a single sign-off.
📖 Read the full blog: https://www.rwaltz.com/blogs/ai-clinical-documentation-what-ambient-scribes-change-and-what-they-do-not
Connect with RWaltz:
LinkedIn: https://www.linkedin.com/company/rwaltzsoftware
X (Twitter): https://twitter.com/rwaltzsoftware
Facebook: https://www.facebook.com/RWaltz-Software-PvtLtd-255590135349493
Telegram: https://t.me/RWaltzCrypto
GitHub: https://github.com/rwaltzsoftware
Clutch: https://clutch.co/profile/rwaltz-software
Website: https://www.rwaltz.com
Top comments (0)