The Clinical Frontier · 12 October 2026 · Issue 025 Google’s diagnostic AI reached The Lancet this week. India’s primary health centres are the right place to think about what comes next.
In this issue
- Google published its first Lancet study of AMIE on 8 October 2026: 98 real patients at a US urgent primary care clinic, zero safety stops, 90% inclusion of the final diagnosis across the AI’s top seven suggestions.
- The study design, with a live physician supervising every AI conversation, sets the evidence baseline for what credible real-world clinical AI deployment looks like.
- India’s 25,000-plus primary health centres face a documented doctor shortage. AI-assisted pre-assessment could extend the reach of the PHC system, with the right language, connectivity, and governance safeguards.
- Anthropic’s Claude Haiku 5.5, released one day earlier on 7 October 2026 at approximately 90% lower cost than Haiku 4.5 for requests under 100k tokens, shows where the inference cost curve for patient-facing AI is heading.
The signal
On 8 October 2026, Google published a prospective real-world feasibility study of its AMIE conversational diagnostic AI in The Lancet. Google describes this as its first paper published in the journal’s main title.
The study ran from April through November 2025 at Beth Israel Deaconess Medical Center’s urgent primary care clinic in Boston. Patients who had scheduled an urgent appointment used AMIE for a structured pre-visit chat from home before seeing their physician. A supervising physician monitored every conversation in real time and could intervene at any point.
Key findings, from the primary source:
- 98 patients completed both the AI pre-visit chat and the in-person physician appointment.
- AMIE included the final diagnosis in its first seven suggestions in 90% of cases.
- AMIE’s top-ranked suggestion matched the final diagnosis in 56% of cases.
- Zero conversations required a safety stop under the study’s prespecified safety criteria.
- Safety supervisors recorded one hallucination and added clinical context in five interactions.
- Supervising physicians reported that AI summaries helped them prepare for the visit in 75% of cases (33 of 44 physicians who reviewed transcripts).
- Physicians felt the AI potentially changed how they handled the appointment in 57% of cases.
One finding that deserves equal attention: fewer than half of patients felt AMIE would keep their information confidential or seemed honest and trustworthy. Patient trust in AI remains an open design problem, not a solved one.
Google itself states that larger clinical trials are needed before assessing patient-facing AI at scale. The study is a feasibility signal, not a deployment recommendation.
How it actually works
AMIE conducts a structured pre-visit interview: it asks open history questions, follows up on responses, and generates a differential diagnosis list ranked by likelihood. At the end of the chat, it produces a summary for the supervising clinician covering the presenting complaint, relevant history, the differential, and any red-flag symptoms.
The physician then enters the consultation already oriented. In the Lancet study, the AI summary sat alongside the patient’s electronic health record in the physician’s view.
The workflow is not a replacement for the physician. It is a structured handoff: the AI gathers and organises; the physician diagnoses and decides. The value demonstrated in the study is showing that this handoff can work safely in a live clinic, not just in a controlled simulation.
Why India’s PHC system is the right frame
India operates more than 25,000 primary health centres under the National Health Mission. Many are understaffed, particularly for qualified MBBS doctors at the primary care level. The gap between patient load and doctor availability at the PHC is one of the most persistent constraints in India’s public health delivery system.
The practical problem the AMIE study addresses, how to use a physician’s limited consultation time more effectively, is directly relevant here. A PHC doctor seeing 60 to 80 outpatients in a session has limited time for structured history-taking with each patient. An AI-assisted pre-assessment that collects and organises the chief complaint, relevant history, and red flags before the consultation begins could let the doctor focus on clinical judgment rather than information gathering.
Key considerations for the Indian PHC context:
- Language. The AMIE study used English. India’s PHCs serve patients across 22 scheduled languages and hundreds of regional dialects. Any India adaptation requires multilingual conversational capability, not just translation of English prompts.
- Connectivity. Urban PHCs and sub-district hospitals may have reliable broadband. Many rural PHCs operate on intermittent connectivity. AI pre-assessment workflows need either an offline-capable design or a reliable low-bandwidth fallback mode.
- ABDM integration. India’s Ayushman Bharat Digital Mission provides a FHIR-based health records infrastructure. An AI pre-assessment tool that writes structured outputs (chief complaint, history summary, preliminary differential) to an ABDM-linked record could make that clinical context available to any subsequent provider, not just the PHC doctor who saw the patient initially.
- DPDP Act 2023. Any patient interaction through an AI system collects personal data. Under India’s Digital Personal Data Protection Act, the data fiduciary (the hospital or app operator) must establish lawful purpose, obtain verifiable consent, and meet breach notification obligations. The patient trust finding from the AMIE study (fewer than half of patients felt the AI was trustworthy about privacy) makes this not just a compliance checkbox but a genuine product design challenge.
The cost angle: Claude Haiku 5.5
One day before the AMIE paper was published, Anthropic released Claude Haiku 5.5 on 7 October 2026. The pricing is relevant to any discussion of AI at PHC scale.
Haiku 5.5 is priced at USD 0.10 per million input tokens and USD 0.50 per million output tokens for requests under 100k tokens. Anthropic states this is approximately 90% lower than Haiku 4.5. For a high-volume PHC pre-assessment scenario where one patient interaction might involve 2,000 to 4,000 tokens of input and output, the model inference cost per interaction is roughly USD 0.001, before infrastructure and integration costs.
Haiku 5.5 is not AMIE: it is a general-purpose fast model, not a trained diagnostic AI. But it signals the direction of the inference cost curve. As that cost falls, the compute component of a patient pre-assessment interaction becomes negligible relative to the clinical staff time it could free up.
Anthropic’s Haiku 5.5 system card also notes that the model is more likely to stress the need for prompt medical care in relevant situations than its predecessor. That is a small but meaningful alignment improvement for any health-adjacent deployment where the AI might be the first conversational touchpoint before a clinical consultation.
What Indian health-IT teams should do now
State health departments and NHM programme managers evaluating AI for PHC outpatient workflows should treat the AMIE Lancet paper and its commentary as the current evidence baseline. The study design, particularly the real-time physician supervision and preregistered safety criteria, is a useful template for any Indian pilot.
Health-tech product teams building outpatient queue management, appointment pre-assessment, or triage tools should examine whether a structured AI interview layer, with mandatory physician oversight, could reduce per-consultation time without increasing clinical risk. The finding that AI summaries potentially changed how physicians handled 57% of appointments suggests the pre-assessment layer affects consultation quality, not just throughput.
Clinical AI researchers at Indian institutions should note the patient trust finding. Patient-facing AI in India faces additional trust barriers: limited prior exposure to AI in healthcare settings, concerns about data use by government or private entities, and health literacy gaps that affect how patients understand what an AI can and cannot do. Any PHC pilot should treat community trust-building as a first-class design requirement, not an afterthought.
The takeaway
The AMIE Lancet study is the most credible real-world evidence to date that a conversational AI can safely assist in primary care pre-assessment, under physician supervision. It is a feasibility study at one US clinic, and it carries important caveats: a single centre, English-only, with a live physician on every call, and with patient trust in the AI remaining low.
India’s PHC system has the scale and the need. The question is not whether AI-assisted pre-assessment will eventually reach India’s primary health centres. It is whether the implementations that arrive will be built with the same rigour: prospective clinical design, real-time physician oversight, pre-registered safety criteria, and transparent reporting of where patient trust still needs to be earned.
The Clinical Frontier is a daily briefing from HCITExperts for India’s health-IT community. Primary sources: AMIE feasibility study (The Lancet, 8 October 2026); Google AMIE announcement (Google Blog, 8 October 2026); Introducing Claude Haiku 5.5 (Anthropic, 7 October 2026). More tomorrow.