The demo begins with a polished note. The useful evaluation begins when the microphone drops a medicine name, two people speak at once, and the doctor changes the plan after the visit.
AI is not one clinical task. Its risk changes with the output and what people do next.
Takeaways
- Evaluate one defined task and consequence at a time.
- Put human review where an error can affect care or the record.
- Test ambiguity, interruption, correction, and exit.
- Verify legal and programme claims against current official sources.
This buyer scorecard is not medical, legal, regulatory, or procurement advice. It does not validate a product for a patient population or clinical use.
Sort tasks by consequence before features
Use three lanes during a trial.
| Lane | Example output | Minimum acceptance boundary |
|---|---|---|
| administrative | appointment summary or internal task draft | staff review before action or external message |
| documentation | transcript or draft visit note | clinician verifies patient, facts, omissions, and final record |
| decision-adjacent | prescription text, code, alert, or clinical summary | authorised clinician independently checks and accepts; no autonomous order |
The lane names are local. The boundary matters. A fluent paragraph can still contain the wrong patient, dose, chronology, or negation.
WHO’s AI-for-health ethics guidance centres human autonomy, safety, transparency, responsibility, and equity. Its regulatory considerations discuss documentation, intended use, validation, data quality, privacy, and lifecycle oversight. These are evaluation references, not proof that a particular tool is safe or approved.
Give every trial one task contract
AI TASK CONTRACT
Task and intended user: ___________________________________
Allowed input and prohibited input: _______________________
Output status: draft / suggestion / administrative aid
Who must review before the next action: ___________________
Errors with the highest consequence: ______________________
How uncertainty and missing context appear: ______________
Evidence retained for correction and review: _____________
Fallback when the service is unavailable: _________________
Exit and deletion evidence: _______________________________
Do not let a general label such as “clinical copilot” replace this contract. Repeat the exercise for each enabled task.
Test five hostile moments
Use invented data and an authorised test environment.
- Two speakers use the same surname.
- The doctor corrects a medicine after first mentioning it.
- Hindi and English appear in the same exchange.
- A clinically important negative statement is followed by background noise.
- The service becomes unavailable before the draft is accepted.
The reviewer should be able to compare source context, edit without hiding the change, reject the draft, and complete the work through a safe fallback. Record the result per case. A single average score can conceal the error that matters most.
The India medical-scribe workflow has a deeper documentation test. The AI-scribe comparison guide separates vendor claims from witnessed evidence.
Trace the data before the pilot
Draw each transfer: device, browser or app, model service, storage, support access, exports, analytics, subprocessors, and deletion. For every copy, record purpose, owner, access, retention, and exit treatment.
India’s Digital Personal Data Protection Act, 2023 and the final Digital Personal Data Protection Rules, 2025 require date-sensitive implementation analysis. Their provisions and commencement dates are phased. A buyer should have qualified counsel determine what is in force and how it applies to the exact roles, purposes, notices, safeguards, rights handling, and incident process.
Score evidence, not adjectives
| Test | 0 | 1 | 2 |
|---|---|---|---|
| intended use | vague | stated | task, user, limits, and prohibited use recorded |
| human review | implied | named | witnessed before consequential action |
| failure handling | absent | described | interruption and fallback tested |
| correction trail | overwrite | edit visible | source, edit, acceptance, and rejection reconstructable |
| privacy and exit | sales answer | documents shared | data path and deletion/export exercise completed |
Set the pass rule before the trial. A product with a strong total should still fail if a mandatory safety boundary scores zero.
Patient Square is an AI clinical platform. Its visit-note output is a draft for clinician review; prescription text is a draft, and ICD-10 output is a suggestion. Do not infer autonomous diagnosis, prescribing, coding, or completed ABDM certification. Review the current India security page, product scope, and agreement, then run the same hostile cases against the selected configuration. The right result is not an impressive demo. It is a task whose limits remain visible on a difficult day.