The next AMIE evaluation needs to establish whether the proposed AI-assisted pathway improves care compared with an appropriate alternative. That means examining patient consequences, clinical decisions, total workload and understanding, not simply whether a supervised conversation can take place. The Lancet feasibility study published on 8 October 2026 provides a reason to pursue those questions, not their final answers.
There is already a relevant research development. Google announced a nationwide randomised-study collaboration with Included Health on 3 February 2026, and its 11 August 2026 update described that study as ongoing. The February announcement and August update concern research progress; they should not be presented as published comparative outcome results.
From feasibility to clinical value
A feasibility study asks whether a defined arrangement can operate and what needs to be learned before a more demanding evaluation. A comparative clinical study asks what changes because that arrangement is used rather than the alternative.
The distinction is practical. A patient may complete an AI conversation, rate it positively and provide a useful summary without experiencing a better health outcome. Another may benefit because the conversation identifies an unanswered question that changes assessment. A third may encounter additional work or confusion. Which pattern predominates must be measured.
The favourable AMIE feasibility findings make further evaluation worthwhile. They do not remove the need to define the intervention, comparator and intended benefit precisely.
What the announced collaboration does and does not establish
Google's 3 February 2026 announcement described a proposed nationwide randomised evaluation with Included Health, subject to the relevant review arrangements at that stage. The 11 August update subsequently called it ongoing. As checked on 10 October 2026, the cited programme updates do not provide completed comparative results. The public research accounts should therefore be read as evidence of the research programme's direction, not proof that it has achieved its aims.
The collaboration also should not be assumed to use precisely the same AMIE configuration, patient selection or workflow as the BIDMC feasibility study. Any published results will need their own description of the system and intervention.
The endpoints and decision rules proposed below are editorial recommendations. They are not an assertion that each appears in the announced protocol, nor a description of results already obtained.
Choose the comparison that matches the promise
For a service claiming to improve ordinary care, usual care is a relevant starting point. But researchers must define what ordinary care includes: existing reference resources, staff support and the actual appointment pathway. Comparing AI assistance with an artificially deprived clinical group would answer a less useful question.
If the proposed benefit comes from additional attention before an appointment, an equivalent human-supported intake process may also be informative. That comparison can help distinguish the value of collecting more information from the value of the specific automation used to collect it.
The unit of allocation matters as well. Assignment by patient, clinician or service creates different practical issues, including whether experience with the AI changes behaviour in the comparison pathway. These design choices should be explained rather than hidden beneath the word randomised.
A pre-publication outcome scorecard
| Proposed domain | Question that would matter | What should not substitute for it |
|---|---|---|
| Diagnostic consequences | Did the pathway reduce important missed or delayed diagnoses? | A longer differential alone |
| Appropriate action | Did decisions fit the available evidence and patient circumstances? | A plausible management paragraph |
| Safety and recovery | Were consequential errors detected, corrected and followed through? | The absence of one type of stopping event |
| Whole-episode workload | What changed across preparation, review, appointments and follow-up? | Minutes saved in one isolated task |
| Patient understanding | Did patients understand uncertainty and the next step? | Satisfaction alone |
| Access and equity | Who completed, needed help, declined or was excluded? | Performance only among successful users |
| Subsequent activity | Did the pathway generate useful care or unnecessary repeat work? | The number of messages or actions completed |
A strong publication need not make every domain its primary endpoint. It should make clear which claim it is designed to establish, which outcomes are exploratory and which remain unmeasured.
Keep incomplete pathways in view
A real service includes people who cannot finish intake, are redirected or stop using the system. Those outcomes need to remain visible even when a particular diagnostic analysis appropriately focuses on completed encounters.
For example, a fictional trial could find excellent summary quality among completers while many patients who need assistance never reach that stage. The summary result would still be valid for its analysed group, but it would not establish equitable access to the service. The illustration is a design warning, not an AMIE finding.
An evaluation should also describe what happens after escalation. A correct decision to hand back is only part of the pathway if nobody receives the concern or the patient cannot access the recommended next step.
Evaluate the people and the system together
The receiving clinician's performance may depend on how the AI material is presented, what sources are available and whether uncertainty survives summarisation. Evaluating only the model's final output misses that interaction.
The DECIDE-AI framework provides relevant principles for reporting early live clinical evaluation, including human factors. It is not a certification standard or a replacement for a suitable trial, but it reinforces the need to describe the working arrangement rather than only the algorithm.
For longer-term use, an additional research question concerns clinicians' retained ability to review, challenge and recover when assistance is incomplete or unavailable. That should be assessed in appropriate educational or simulated settings, not by withholding necessary support from real patients merely to create a test.
Mixed results should produce a mixed verdict
Suppose a future study finds better preparation but more total work. That could justify targeted use where the information benefit is valuable, while undermining a broad staffing-saving claim. Suppose it finds faster intake but no improvement in understanding and more repeat contacts. The speed result would not settle the service-value question.
Conversely, a neutral diagnostic-list comparison could coexist with clearer patient explanations or fewer unresolved actions. Those would be legitimate benefits if measured appropriately. The evaluation should not be designed so that only a headline about outperforming doctors counts as success.
These are hypothetical result patterns. They illustrate why adoption decisions should be linked to the promised benefit and the observed trade-offs, rather than a single favourable percentage.
What would justify wider adoption?
Wider use would be more defensible with representative comparative evidence, transparent reporting of failures and exclusions, appropriate independent assessment and a clear account of the work that remains with clinicians. The strength of evidence should match the consequences of the delegated task.
The immediate AMIE lesson is that real-world evaluation is possible and informative. The next lesson should concern what the service improves, for whom, at what operational cost and under which safeguards. That is a more demanding and useful standard than repeating that the system can hold a convincing conversation.
Frequently asked questions
Is the Included Health collaboration still only a February announcement?
No. Google's 11 August 2026 update described the study as ongoing, although the programme updates checked on 10 October did not provide completed comparative results.
Are the outcomes proposed here the confirmed trial endpoints?
No. They are editorial recommendations for interpreting future evidence, not a reproduction of the announced study protocol.
Would better preparation alone justify reducing appointment time?
Not by itself. A staffing or capacity claim needs evidence about the complete workload and consequences, not only clinicians' preparedness.
Turn an evidence question into focused professional learning →