Ambient clinical documentation in India: what changes in OPD

The demo is always the same. A clean room, one speaker, unaccented English, a tidy two-minute conversation about a sore knee, and out comes a beautifully structured note. Every clinician watching nods. Then you put it in an OPD where a consultant sees forty patients in a little over three hours, the patient’s son does most of the talking, the sentence starts in Telugu and ends in English, and a nebuliser is running two cubicles away.

Ambient documentation is the most promising clinical AI application I have worked on and the one most likely to be abandoned for reasons that have nothing to do with how good the transcription is. Doctors do not evaluate it on accuracy. They evaluate it on whether their day got shorter, and if the answer is no, they stop using it in the second week and nobody tells you.

What follows is what actually changes in the room, what breaks, and how to run a trial whose result you can trust.

What the doctor actually experiences

Strip away the positioning and the doctor’s experience is three things. Something has to be started and stopped. Something has to be read and corrected. Something has to be signed and land in the right place in the record.

Each of those is a friction point, and the doctor’s judgement forms on all three within about five consultations. If starting requires unlocking a phone and finding an app, they will forget on the busy days, which are the days that matter. If the note appears two minutes after the patient has left, they will have moved on and will review it at the end of the session, which means reviewing thirty notes in a batch and losing the context that made review fast. If the signed note lands in a free-text field rather than the structured sections the EMR expects, the coding team and the next clinician both lose.

I would spend more design effort on those three mechanics than on the model. They are where adoption is decided.

Consent, and who asks for it

The consent question gets handed to legal and comes back as a paragraph nobody will read aloud forty times a day. That is a design failure disguised as compliance.

What has to be true: the patient knows a machine is listening, they know what happens to what it hears, they can say no without it being awkward, and there is a record that they were asked. What that looks like in practice is a short spoken line from the doctor or the OPD nurse — one sentence, in the language of the consultation — backed by signage in the room and a consent flag captured at registration rather than in the consultation itself. Putting the capture at registration is the single change that makes this workable; it takes the burden off the clinician and gives you an auditable artefact.

The decisions you must make explicitly, because the vendor will otherwise make them for you:

Is audio retained at all, or only the transcript, or only the final note? The cheapest defensible position is that audio is processed and discarded within a short defined window, and you should have a strong reason to deviate.
Where does processing happen, and does anything leave the country? Answer this before the pilot, not during the security review.
Can recordings be used to improve the system? That is a separate purpose under DPDP and needs its own consent and its own decision. Deciding late means discarding data you have already collected.
What happens when the patient declines? There must be a path that takes five seconds and does not require the doctor to explain anything. If declining is inconvenient, consent is not real.
Sensitive consultations — mental health, sexual health, adolescent, anything where an attendant’s presence is itself an issue — need a default of off, not a default of on with an opt-out.

The Indian consultation is not the demo

Accuracy claims are made on clean single-speaker audio. Your room has none of those conditions, and the failure modes are specific.

Code-switching mid-sentence. Not “the patient speaks Hindi” — the patient speaks Hindi with English clinical nouns, English numbers, and a verb structure the language model was not trained to expect. Systems that switch language at the utterance level produce nonsense on sentences that switch at the word level.
Third-party narration. In a large share of Indian consultations the history comes from an attendant. The note must attribute correctly — “son reports” is clinically different from “patient reports” — and misattribution is the error type I would watch most closely, because it is invisible on a read-through and consequential in the record.
Drug names and dosages in accented speech. Two brand names that differ by a syllable, spoken quickly, in the middle of a Hindi sentence. Any system that writes a medication line without a hard vocabulary constraint will eventually write the wrong molecule.
The examination gap. Nothing is said behind the curtain, or what is said is a nurse being instructed. The examination findings are often not spoken at all, so the note will be confidently incomplete — which is more dangerous than being obviously incomplete.
Length. A six-minute follow-up in a crowded OPD contains very little speech. There is not enough signal to write a rich note, and a system that writes one anyway is filling gaps.

The last two are why I now insist on an error taxonomy that separates omission, misattribution and fabrication. They need different fixes, and a single accuracy number hides all three.

Template discipline

The instinct is to let the system write good clinical prose. Resist it. What you want is a note that fits the structure the speciality already uses and the fields the EMR already has — chief complaint, history, examination, provisional diagnosis, investigations, plan, advice — with the boundaries respected and nothing invented to fill a heading.

Templates should be per speciality and agreed with the clinical leads of that speciality, not designed centrally. An orthopaedic follow-up note and a paediatric first consultation have almost nothing in common. Where a structured field exists in the EMR — diagnosis code, drug, dose, frequency, duration, follow-up date — the system should populate the field, not write the value into narrative text. That is unfashionable work with no demo value and it is what makes the output useful to anyone other than the doctor who dictated it.

One discipline that has saved me trouble: an explicit “not documented” convention. If the examination was not spoken, the note says so rather than omitting the section silently. Clinicians trust a system that admits a gap far more quickly than one that produces a complete-looking note.

The quality review loop

Somebody has to read the notes, and it cannot be the person who bought the system. Set up a review with a clinician from the participating speciality, someone from medical records, and a coder. Weekly during the trial, monthly afterwards. They read a stratified sample — new patients, follow-ups, consultations flagged as multilingual, consultations under a certain duration, and any note the doctor rejected outright.

They score against a fixed taxonomy:

Omission — clinically relevant content said in the room and absent from the note.
Fabrication — content in the note that was not said or implied. Any instance is a stop-the-line event, not a metric to be averaged.
Misattribution — a statement assigned to the wrong speaker, or an attendant’s report recorded as the patient’s.
Structural error — content in the wrong section, or a structured field left empty when the value was stated.
Edit burden — how much the doctor changed. This is your real adoption predictor and the easiest thing to instrument.

Track edit burden per doctor, not in aggregate. The distribution will be bimodal — a group whose notes need a word or two and a group rewriting from scratch — and the second group is telling you something specific about their speciality, their accent, their consultation style or their template. Aggregate accuracy will hide that entirely.

Why adoption fails on workflow rather than accuracy

Every abandonment I have seen had a workflow cause, and none of them appeared in the evaluation criteria.

The note arrived after the patient left, so review became a batch chore. Signing required a different screen from the one the doctor was already in. The system worked on a phone but the EMR was on a desktop, so the doctor was holding two devices. The consultant’s secretary, who had been typing notes for eleven years, was not part of the rollout and quietly kept doing it the old way. There was no offline behaviour, and the network in the old block is what it is. The stop button was easy to miss, so recordings ran into the next patient’s consultation — which is a privacy incident, not an inconvenience.

And the biggest one: nobody redesigned what the doctor does with the time. If the note is written but the doctor still spends the same minutes in the room, the value shows up as a shorter evening, which is real but invisible to the person who approved the spend. Decide in advance whether the goal is a shorter day, a more complete record, better coding, or more patients per session. Those lead to different designs, and pretending you want all four is how a trial produces no answer.

What actually changes for the doctor

Honestly: less than the pitch says, and in a different place. The typing goes away and reviewing arrives. For a fast typist with a terse note style, that is close to a wash and they will tell you so. For a consultant who has been dictating to a secretary, the change is procedural rather than temporal. The clinicians who benefit most are the ones whose notes were thin because there was no time — their record quality improves sharply, which matters for continuity, for coding and for anything you later want to do with that data.

The second change is subtler and, I think, the real argument. The doctor looks at the patient instead of the screen. I have had consultants describe that as the reason they kept using it, after telling me the time saving was marginal. It does not appear on a business case and it is what drives voluntary adoption.

How to run a fair trial

Most hospital trials of this are rigged towards a yes, then fail at scale. To avoid that:

Pick two or three specialities with genuinely different note structures, and within each, a mix of enthusiasts and sceptics. A trial run only by volunteers tells you what volunteers do.
Measure a baseline first, for two weeks, before anything is installed: time from consultation end to note signed, note completeness against your own audit criteria, and the doctor’s session end time. Without a baseline you will be arguing about impressions.
Run for at least six to eight weeks. The novelty curve is real and the dip in week three is where you learn something.
Instrument usage per doctor per session, not just overall. Silent abandonment is the main failure mode and it looks fine in aggregate numbers.
Write the exit criteria before you start, including the criteria for stopping. Any fabrication in a medication or diagnosis line, or a consent failure, ends the trial pending a fix.
Include the secretaries, the OPD nurses and the coders in the trial population. They are part of the workflow whether or not they are in the plan.

If you’re starting this next quarter

Settle audio retention, processing location and the improvement-purpose question, in writing, before any vendor conversation gets serious.
Move consent capture to registration and write the one spoken sentence the clinician will actually say. Test it in each language the unit serves.
Agree note templates with two speciality leads and map every field to its EMR destination. If a field has no destination, drop it.
Fix the three mechanics — start, review, sign — inside the screen the doctor already lives in. Accept a worse model in exchange for fewer clicks.
Measure your baseline for two weeks with nothing installed.
Run the trial with sceptics included, the review loop live from week one, and a documented stop condition.
Decide what the time is for before you claim the time back.

Doctors will forgive a system that writes an imperfect note. They will not forgive one that adds a click.

Questions people ask

What is ambient clinical documentation and how does it work in a consultation?

Ambient documentation is software that listens to the consultation, transcribes it and drafts a structured clinical note the doctor reviews and signs into the record, so the doctor does not type. In practice the doctor’s experience reduces to three mechanics: something has to be started and stopped, something has to be read and corrected, and something has to be signed and land in the right EMR fields. Adoption is decided on those three within about five consultations, not on the model.

Why does ambient documentation fail in Indian hospitals even when the transcription is accurate?

Because doctors evaluate it on whether their day got shorter, and every abandonment I have seen had a workflow cause. The note arrived after the patient left, so review became a batch chore. Signing needed a different screen. The system ran on a phone while the EMR was on a desktop. The secretary who had typed notes for years was left out and kept doing it the old way. There was no offline behaviour in the old block. None of this appears in accuracy evaluations.

How should a hospital take patient consent for AI recording a consultation?

Not with a legal paragraph nobody will read aloud forty times a day. Capture the consent flag at registration, not in the consultation, backed by signage in the room and one short spoken sentence from the doctor or OPD nurse in the language of the consultation. That takes the burden off the clinician and gives you an auditable record. Declining must take five seconds and require no explanation; if declining is inconvenient, consent is not real. Sensitive consultations should default to off.

What does DPDP require for audio recordings from ambient documentation?

Decide three things in writing before any vendor conversation gets serious, or the vendor will decide them for you. Whether audio is retained at all, or only the transcript, or only the final note — the cheapest defensible position is that audio is processed and discarded within a short defined window. Where processing happens and whether anything leaves the country. And whether recordings can be used to improve the system, which is a separate purpose needing its own consent. Deciding late means discarding data already collected.

How does code-switching between Hindi, Telugu and English affect ambient documentation accuracy?

Badly, and in a specific way. Accuracy claims are made on clean single-speaker English. An Indian consultation switches language at the word level — Hindi sentence, English clinical nouns, English numbers — and systems that switch at the utterance level produce nonsense on those sentences. Add third-party narration from an attendant, drug brand names that differ by a syllable spoken quickly, and examination findings that are never spoken aloud. Test on your own OPD audio, not the vendor’s demo.

What errors should a hospital track when evaluating an AI scribe?

Not a single accuracy number, which hides everything. Use a taxonomy: omission, where something said is absent from the note; fabrication, where the note contains content never said — any instance is a stop-the-line event, not a metric to average; misattribution, where an attendant’s report is recorded as the patient’s; structural error, where content lands in the wrong section; and edit burden, how much the doctor changed. Edit burden per doctor is your real adoption predictor and the easiest thing to instrument.

How long should an ambient documentation pilot run before deciding?

Two weeks of baseline measurement before anything is installed — time from consultation end to note signed, note completeness against your own audit criteria, and the doctor’s session end time. Then at least six to eight weeks of trial, because the novelty curve is real and the dip in week three is where you learn something. Instrument usage per doctor per session; silent abandonment is the main failure mode and looks fine in aggregate. Write exit and stop criteria before starting.

Which doctors benefit most from ambient documentation, and which will not?

The clinicians who benefit most are those whose notes were thin because there was no time — their record quality improves sharply, which matters for continuity, coding and anything you later do with that data. A fast typist with a terse style will find it close to a wash and will say so. A consultant who dictates to a secretary sees a procedural change, not a time saving. The subtler benefit, and the real driver of voluntary adoption, is that the doctor looks at the patient instead of the screen.

What does an ambient documentation trial cost a hospital beyond the software licence?

The licence is the visible line. The hidden costs are the integration work to populate structured EMR fields rather than dump narrative text, per-speciality template design with clinical leads, a weekly quality review with a clinician, a medical-records person and a coder, and the time to include secretaries and OPD nurses in the rollout. Then decide what the saved time is for — shorter day, better record, better coding or more patients per session — because the business case only closes on one of those, and pretending you want all four produces no answer.

Should note templates for an AI scribe be designed centrally or by each speciality?

Per speciality, agreed with the clinical leads of that speciality. An orthopaedic follow-up and a paediatric first consultation have almost nothing in common. Where the EMR has a structured field — diagnosis code, drug, dose, frequency, follow-up date — the system should populate the field rather than write the value into prose. And insist on an explicit “not documented” convention: if the examination was not spoken, the note says so rather than silently omitting it. Clinicians trust a system that admits a gap far faster than one that produces a complete-looking note.

Who should own the quality review for ambient documentation in a hospital?

Not the person who bought the system. A clinician from the participating speciality, someone from medical records and a coder, reading a stratified sample weekly during the trial and monthly afterwards — new patients, follow-ups, multilingual consultations, very short consultations and anything the doctor rejected outright. The growth or digital function facilitates and instruments, but the judgement on clinical content has to sit with clinicians, or the review has no authority when it says stop.

When is ambient documentation the wrong investment for a hospital?

When you cannot fix the three mechanics — start, review, sign — inside the screen the doctor already lives in. When the EMR cannot accept structured fields, so the note lands as free text nobody else can use. When the pilot population is volunteers only, which tells you what volunteers do. When leadership has not decided what the reclaimed time is for. And when audio retention and processing location cannot be settled before launch. Doctors will forgive an imperfect note. They will not forgive an added click.