Multilingual Ambient Documentation in Indian Outpatient Clinics

The failure modes have a rough order of frequency. Language switching mid-utterance produces transcription that silently drops or garbles the switched segment. Speaker attribution fails when a relative answers for the patient, so the record attributes symptoms to the wrong person. Clinical terms and drug names are rendered as similar-sounding ordinary words, which is the most dangerous category because the output remains fluent and plausible.

Numbers and dosages are a specific risk worth separating out. Doses, frequencies, durations and measurements spoken quickly are frequently misrecognised, and unlike a garbled sentence, a wrong number reads as perfectly ordinary. Any evaluation should look specifically at whether numerical content survived, because a summary that is ninety-five per cent accurate overall and unreliable on doses is not usable.

Then there is the summarisation layer, which is separate from transcription and adds its own failure mode: a fluent, confident summary of a conversation that was partially misheard. The summary reads well precisely because the model has smoothed over the uncertainty, which removes the signal a clinician would otherwise use to notice something is wrong.

Any evaluation has to run on real consultations from your own clinics, with your own clinicians and patient population, with consent obtained properly for the recording and its use. A vendor demonstration or a test on scripted dialogue tells you nothing about the language mix you actually have, which differs substantially between hospitals and even between departments within one hospital.

Sample deliberately across the variation that matters: different clinicians, different departments, different times of day including the busiest clinic, consultations with and without an accompanying relative, and the range of languages your patients actually speak. A sample drawn from one enthusiastic consultant's calm afternoon clinic will produce an encouraging and meaningless result.

Score against the note a clinician would have written rather than against the transcript. What matters is whether the clinically significant content is present and correct — the history, the findings, the assessment, the plan and the medication — not whether every word was transcribed. A tool that misses conversational content and captures the clinical content accurately is doing its job.

The entire value proposition is time saved, and that calculation must include the time spent reviewing and correcting the output. A tool producing a draft that requires careful reading and several corrections may save no time at all, and clinicians will work this out within a fortnight regardless of what the business case said.

Measure it directly during evaluation: time to produce a finished, signed note with the tool against without it, for the same clinicians. Include the correction time and the occasional case where the draft is discarded and the note written from scratch. Report the distribution rather than the average, because a tool that saves time on most consultations and costs a great deal on some is experienced by clinicians as the latter.

Be alert to the more subtle risk, which is that review becomes perfunctory. When a draft is usually good, the natural response is to skim and sign, and the errors that get through are then the ones the clinician did not read carefully enough to catch. This is not a training problem; it is a predictable consequence of automation that is usually right, and it argues for keeping the highest-risk content — medication in particular — outside the ambient path or subject to a separate check.

Recording a consultation is a significant step and requires the patient's informed agreement, obtained in a way that is genuinely voluntary. A patient in a consulting room asked to agree to recording by the doctor about to treat them is not in a strong position to decline, so the request should be made before the consultation, in plain language, with declining made easy and consequence-free.

Establish what happens to the audio: where it is processed, where it is stored, for how long, whether it is retained after the note is produced, and whether it is used to improve the vendor's models. That last point deserves particular attention, since consultation audio is highly sensitive and any secondary use requires a lawful basis that a general treatment consent almost certainly does not provide.

Decide deliberately whether audio is retained at all once the note is signed. Retention creates a record of the consultation far richer than the note, with its own discovery and breach implications, and most hospitals conclude on reflection that they do not want it. Where it is retained, it should be governed as clinical data with the same access controls, not as a technical artefact sitting in a vendor's storage.

Performance varies substantially by setting, and a sensible deployment reflects that rather than treating the tool as a hospital-wide capability. Consultations that are longer, more structured and more likely to be conducted predominantly in one language tend to work better; rapid high-volume clinics with heavy code-switching tend to work worse.

Start where the evaluation showed it works and extend only where evidence supports it. A partial deployment that genuinely helps three departments is a better outcome than a hospital-wide rollout that two departments quietly abandon, and it avoids the reputational cost that makes the next AI proposal harder to get accepted.

Keep the clinician's authorship unambiguous throughout. The note is the clinician's, they are responsible for its content, and the record should show that they reviewed and signed it. Whether the record notes that a draft was machine-generated is a decision worth taking deliberately rather than by default, and there is a reasonable argument that it should, since it tells a future reader something about how the note came to exist.

Performance is not fixed. Vendors update models, and an update can improve or degrade performance on your particular language mix without any announcement that would let you predict which. Your clinics change too, as patient demographics shift and clinicians come and go with different speaking patterns.

Build in periodic reassessment on a fresh sample rather than treating the initial evaluation as settled. It need not be elaborate — a modest sample scored the same way as the original — and it is what catches a silent degradation before clinicians notice it as a general sense that the tool has got worse.

Ask vendors contractually to notify you of material model changes and to allow a period to reassess before a change takes effect. This is the same algorithm-change concern that applies to any deployed model, and ambient documentation is a case where a change can be substantial and entirely invisible in the interface. Keeping the evaluation results and the change history together in your governance record is what makes that reassessment comparable rather than a fresh opinion each time.