The value of AI pre-diagnosis is not a substitute for doctors, but a shift from information gathering to shared decision-making

01 Facts & context

What has been done with the AMIE study: pre-hospital collection, summary handover and full custody

2026-10-20, Google Blog published an interview with Anil Palepu and Mike Schaekermann, Google Research Scientist, on his clinical reasoning and dialogue research system, AMIE. As described in the text, participants in the study may choose to have a dialogue with AMIE before their next visit, with the systematic collection of clinical history and the provision of a concise summary to their doctors; Each interaction between AMIE and the patient was supervised by a trained doctor, but the interview stated that the guardian did not need to interrupt or intervene.

The two researchers also described the way forward of the project: a larger follow-up study was being prepared to cover a wider range of patients in the United States, to include more patients and to set comparisons; The interactive form begins with the text and extends to the distribution of images in the text, and audio and video are being explored. The main point of the doctor ' s feedback during the interview was that when it was not necessary to spend much time checking underlying facts, they could focus on discussing treatment options and joint decision-making with patients. All of the above were disclosed in the interview and are not considered an independent validation of the system ' s clinical effectiveness.

02 GSP analysis

Strategic choice: “Pre-clinic concourse” first, not direct diagnostic subject

According to the public description, AMIE is located as an information medium between the patient and the doctor: it does not declare a diagnosis, it does not replace a face-to-face medical examination, but rather provides a structured and individualized medical history to the doctor before the consultation. This trade-off is noteworthy. The easiest story to tell by the Medical AI is “AI to see the patient”, but the path is to cross the three-step threshold of supervisory clearance, clinical responsibility and medical acceptance; In contrast, Pre-patient Collection and Abstract put AI in ancillary position, the subject of responsibility remains a doctor and the purchaser is more likely to find a drop point in the existing process.

Example of hypotheses: If a chain clinic purchases such a system, it is more realistic to embed it in the chain after an appointment to confirm a text message, to allow patients to complete the dialogue during the waiting period, to include a summary in the draft medical record and to visit it two minutes before a doctor is seen. This is a very different product commitment from “to allow AI to give a preliminary diagnosis”, where the former sells time and information integrity, the latter sells judgement and the latter requires much higher levels of evidence and accountability.

For the GSP, the case is inspired by the fact that when an AI is introduced in a professional service industry, low-responsibilities, high-frequency and measurable links such as “information handling and pre-packaging” are more likely to result in sustainable business closure than direct “core judgement”. The conditions for determining whether the link has a clear interface and a verifiable output; Automation only moves the chaos forward if the output is unreceived and unresearched.

03 GSP analysis

Value mechanism: Not saved by typing time, but by the cost of a cognitive shift in consultation

One of the details mentioned in the interview was that the standard admission form only allowed the patient to describe the problem and to tick a few, while AMIE could ask what kind of pain, when it started and where it was in the back. The value of such questioning is not merely more information, but rather the “clarification—requirement-and-clarification” cycle that was to be done during the doctor's visit is pre-empted. The most rare thing doctors have to do in the clinic is not to type speed, but to turn their attention: each time they ask the basic facts, they have to return to the “data-collection” situation from the “thinking with the patient” situation.

The economic value of such systems should therefore be measured by “saving high-quality dialogue time per consultation” and “reduced omissions and back-to-work”, rather than by “saving minutes”. Example of hypothesis: An average of twenty minutes per consultation in a specialist clinic, if the pre-patient summary allows a doctor to take four minutes less to check the underlying history, and four minutes more to discuss the dependence of treatment and patient concerns, the value of which is reflected in the rate of re-examination, programme achievement and patient experience, which are closer to the operational outcome than the word in the document.

Conversely, if the summary merely condensed the patient ' s original words into a single text, and the doctor had to go through it, the time saved would be close to zero and would even add to the burden of reading the digest. The condition of value is that the summary be “checkable, retrospective and organized in the order of clinical decision-making”, and the doctor can see at first glance what information is missing and what needs to be identified in person. This is a product design problem and the most appropriate pressure element in the procurement assessment.

04 GSP analysis

User journey: Low patient threshold, high level of trust at doctor level, non-consistent experience at both ends

There are two categories of users on this journey and claims are not consistent. The key to the patient is to be willing to start, to tell the truth and to be able to stop at any time. The interview mentioned that the participants in the study were “optional” to engage in dialogue with AMIE, a voluntary nature that was important: mandatory writing would bring resistance to the quality of the dialogue. Example scenario: an appointment is followed by a skipper entry, starting with two or three questions, and the patient answers a paragraph before asking the next paragraph, which is more likely to be described in detail than a one-time long questionnaire. In addition, the low threshold means that patients can use fragment time to complete their expressions without familiar medical terminology, and that the system has to take on the task of transforming mouth language into a structured thread, rather than requiring patients to learn first to describe symptoms.

The key to doctors is trust and control. The doctor needs to know that the summary comes from the patient ' s own account, which is the result of an AI inquiry, which is not answered, and that he needs to be able to return to the original dialogue. The arrangement of “doctor-supervised every interaction” in the interview showed that safety nets were placed on people during the research phase; In real operations, this safety net should be transformed into a hierarchical mechanism: regular dialogue will not be interrupted, and medical attention will be made to triggering a particular risk signal. Trust also relies on presentation: listing key findings, uncertainties, non-response sub-columns makes it easier for a doctor to judge quickly the level of confidence than a fluid narrative.

A third user between the two ends is often ignored: the appointment and the front desk. They decide when the entrance will be issued, whether the patient will be alerted when the patient is not finished, and who will be approached when the abstract is abnormal. If this role does not have a clear mandate, the most common result when the system is online is that the entrance is not pushed and the summary is not read. The journey must be designed in such a way that it is written into the process rather than assuming that it will function spontaneously. The other trade-off is that the information particles at both ends are not consistent: the patient is guided by problems, the doctor is given a list of clues with source markers, and the same data is expressed in two languages so that the experience does not drag each other.

The method of testing this section is a serviceability test that is separate from the end: the patient sees the completion rate and the reason for the mid-way exit, the doctor looks at the opening of the summary, the retrace ratio and the time taken to check it; Let the front desk go through the trigger and alarm process and record the card points. Three sets of indicators are established at the same time before the journey is closed.

05 GSP analysis

Organizational division of labour: who is responsible for the summary and determines whether the system will operate on a permanent basis

The division of labour revealed by the interviews was that researchers performed technical assessments and extended call patterns, while another researcher was responsible for clinical research design, collaboration with medical institutions and contribution papers; The study scene is also supervised by a trained doctor. The division of labour was established in the study, but needs to be repositioned when translated into the operating organization. The core issue is: who is responsible when the summary is wrong? If the answer is vague, the doctor will hedge the risk with “I will ask again” and the time saved disappears.

A more feasible arrangement would be to sever the responsibility into three paragraphs: product-side accountability for the quality of the dialogue and the rules for the production of summaries; The clinical director is responsible for the summary information structure and risk alert rules and has the authority to modify them; The first line of doctors made a final judgement as to whether or not to accept a summary, and that judgement was not considered “must be used”. Example of the hypothesis: An agency has a summary evaluation panel composed of two senior doctors, which examines dialogue and summaries on a monthly basis and, if missing, adjusts the logic of questioning rather than individual doctors. Responsibility thus falls on the hierarchy of rules, not on the user. This arrangement also requires the evaluation results to enter backwards into the version of the record of the query template and the summary field so that the changes are retroactive and comparable.

Procurement and compliance roles also need to be entered in advance. The answers to these questions have a direct impact on the online range of data storage, patient authorization, cross-agency access, scar audit. Leaving them for the later stages of the pilot usually results in a forced reversal of the already functioning process. Another vulnerable role is the nursing or pre-visit team: if the summary serves the nurse at the same time, it is necessary to identify who looks first and which version to see and to avoid the duplication of the same material by two sets of processes.

It is simple to test whether the organizational division of labour is on the ground: to draw an omitted summary, and to take three paragraphs of responsibility, to see if there are clear principals, time limits and records for each step. If a “all look together” or “see then”, it means that the division of labour has not been established and that the process is returned to the doctor under real pressure.

06 GSP analysis

Inputs and alternative explanations: do not use “doctor satisfaction” as evidence of impact

The main inputs for such projects go beyond model call costs and include access operations, summary quality maintenance, custody or conversion mechanisms, system docking and compliance. The interview mentioned that follow-up research would be extended to the United States, with more patients and comparisons, which is a good indication that comparable evidence of impact is the most lacking at this stage. Doctor feedback “helps to move from collection to collaboration” is a valuable signal, but it is a subjective experience that is susceptible to freshness and selectivity: doctors and patients who are willing to participate in research may themselves be more compatible and time-consuming.

There are at least three alternative interpretations. First, the time saved may come from the patient being prepared in advance, rather than from the AI questioning itself; Second, the summary improves communication, perhaps simply because the doctor has read the written material, and the same effect can be achieved by visiting nurses for the change of adult workers; Thirdly, in the short term, doctors are more involved in curiosity and in the long run. To distinguish between these interpretations, it is necessary to compare “Has there been a manual advance” “Has there been a summary” for different groups and to observe continuity for more than a few weeks. Communication is not improved by the word “summary”, and “AI” without opening the summary points to “read-as-you-go, open” as the key intermediary.

The input judgement should therefore be phased: it would take time to check with the doctor the proportion of the summary used for small-scale pilot validation before deciding whether to expand it. If summary usage is low and doctors repeat questions, it is stated that the value assumption is not established and that additional input will only magnify the ineffective process. The phasing should also set conditions for withdrawal, such as suspension of the expansion, a return to logic and presentation rather than a hard push through additional training or administrative requirements, when the opening rate for the summaries of two consecutive cycles does not reach the threshold.

When proof of effect is tested, process indicators can also be added as a first-line signal: the completion rate of the dialogue, the proportion of patients in the summary who fail to respond, and the number of visits to the basic facts by doctors in the room. These indicators are closer to the mechanism than “doctor satisfaction” and are available more quickly in small samples to determine whether a larger comparison study is worth entering.

07 Suggested actions

Order of implementation: starting with single-disease, single-hospital, voluntary participation

The first step should be to select a scenario of relative concentration of complaints, such as musculoskeletal pain or routine re-examination, rather than covering the whole curriculum at the outset. The reason is not technical constraints, but questions about logic and abstract structures that need to be field-driven; The narrower the scene, the easier it is for doctors to judge whether the summary is reliable and the easier it is to give enforceable amendments. At the same time, single-horizontal and single-section rooms have been restricted to avoid the inundation of the coordination costs of multiple access points. For example, if a back pain scenario is chosen and the questioning revolves around the nature of the pain, the start time, and the specific parts, the summary should first present these decision points rather than a full flow account, allowing the doctor to judge whether the information is adequate at first glance.

The second step is to embed the entrance into the established appointment process, to identify when an invitation is issued, how long the patient has not been completed, and whether the failure affects access. The initial phase should be voluntary and “jumping does not affect visits” should be written at the entry point in order to reduce the patient's strategic response. The summary is entered into the doctor's view in the form of a “draft” and the doctor decides whether to adopt it or not, leaving it to the conclusion field of the official medical record; The trade-off of this step lies in the exchange of controlled “half-finished” products for the trust of doctors rather than in the pursuit of a once-in-a-lifetime automation.

The third step is the establishment of a weekly round-up: extracting conversations and summaries, marking omissions, misreading and over-requirement of questions, which are to be decided by the clinical director and implemented by the product. The board must have a line of doctors present, otherwise the rules would become more and more like questionnaires. The expansion of the Section or the extended campus will be discussed only if the application rate and the reconciliation take several weeks of steady improvement; If indicators stagnate, priority should be given to overhauling the structure rather than scaling up.

08 Suggested actions

Indicators and experiments: use of “checking time-consuming” and “programme discussion ratio” to replace satisfaction score

Indicators should be designed around time. Process indicators include: the rate of penetration of invitations, the rate of completion of dialogue, the average length of dialogue, the success rate of the digest, the proportion of doctors opening the digest, and the proportion of doctors adding changes to the digest. Outcome indicators include the estimated time-consuming nature of doctors checking the history of basic diseases in the interview, the proportion of time used to discuss treatment programmes and patient concerns, and the number of secondary contacts due to lack of information. Empirical indicators can be retained, but only as an aid and not as a basis for decision-making, as satisfaction can easily be influenced by speech and freshness to hide whether the time structure has really changed.

In the experimental design, it is proposed to alternate rather than simply contrast the single room: the same group of doctors randomly enter the “with summary” and “without summary” processes on different dates, observe the difference between the time-consuming check and the discussion ratio, and track the decline over the four-week period. On the patient's side, the cost of coercion can be judged by comparing the impact of the voluntary entrance and the short questionnaire that must be filled out after an appointment. The trade-off here is the balance between statistical clarity and clinical interference, and the alternate design, while increasing the complexity of the shift, is able to separate the seasonal from the individual doctor.

Pre-conditions for cessation are required. Example of scenarios: If the application rate of the summary for four consecutive weeks is below the agreed threshold, or the doctor ' s replacement rate is high, the description of the summary is not aligned with the clinical decision-making order, the extension point should be suspended and the summary template re-structured. All thresholds are signed by the clinical head before the pilot is launched, avoiding ex post facto interpretation; At the same time, it was agreed who would have the right to adjust the threshold to prevent the pilot from being relaxed in order to meet the target midway.

09 Suggested actions

Applicable conditions and adjustment judgement: under what circumstances and under what circumstances should it be stopped

The conditions for the application of this model are more specific: there are obvious waiting windows before visits, patients are able to complete a word or voice conversation, doctor ' s hours are significantly squeezed by information gathering, and institutions are willing to assign clinical evaluation staff to the quality of the summaries. When these conditions are met, pre-patient AIS collection is more likely to result in measurable time transfers; Conversely, the benefits are reduced if the visits themselves are highly fragmented, the patients focus on acute conditions, or doctors rely on fixed templates. For example, in emergency cases or high-mobile clinics, patients who have not completed the dialogue have been called, the chain of questioning has broken and the summary has increased the burden of checking.

Adjustments can be made in three tranches. If the introduction and reconciliation of the summary are time-consuming to improve significantly, the strains and interactions can be expanded, but the recovery of the recovery and summary template for each extended disease cannot be directly reused. If improvements are limited but the doctor is willing to use them, priority should be given to the abstract structure and the portal to suspend the extension point. If the doctor generally bypasses the summary, or if the cost of guardianship and conversion is higher than the time saved, the extension should be discontinued and returned to manual pre-interviews or simplified questionnaires. The three tranches of judgement should be included in the project charter, with a clear view to the periodicity of observations and decision makers at each stage.

Finally, it is accepted that the benefits of such systems are indirect, reflected in the quality of doctor time and decision-making, and do not automatically translate into income. It is more appropriately managed as a process re-engineering project rather than a one-time software acquisition; The test for success or failure is whether the clinical time structure has really changed and not whether the system is online. This means that the budget should include ongoing evaluation and template maintenance staff, rather than just first deployment costs.