Google put an agent across healthcare records. Keep the evidence gate human

THE SIGNAL IN ONE SENTENCE

Healthcare already has enough invisible machinery. Adding an agent should not create one more black box between a person and care. Google Cloud announced Gemini for healthcare on November 10. The preview is designed for health plans and healthcare providers. Google says the system can work across electronic health records, claims, care-management tools and administrative systems while preserving existing identity and role-based permissions. For providers, the agent can synthesize a patient's history across care settings, prepare a chart before a visit and assemble a prior-authorization packet against a payer's policy. For health plans, it can organize a request for clinical review and explain benefits and costs using the member's own plan documents. Google describes four pieces: healthcare-specific skills, connectors to existing systems, a partner ecosystem and a governed platform with audit logging, citations and model choice. The company says the product is built to support HIPAA compliance and is available in preview. That is the verified reporting. The analysis starts where the demo usually ends. In healthcare, a neat packet is not the decision. A citation is not the same thing as a correct interpretation. A policy match is not a clinical judgment. The system should prepare evidence, expose uncertainty and stop at a gate where an authorized person can inspect what it did. The first requirement is a case-specific evidence packet. Every generated chart summary, coverage explanation or authorization packet should identify the exact records used. Preserve document names, dates, authors, source systems and retrieval times. If a laboratory result was superseded, a medication list was stale or a note belonged to the wrong encounter, the reviewer needs to see that before the summary begins to sound authoritative. The packet should also identify what the agent did not retrieve. A missing specialist note can matter more than three pages of fluent synthesis. Show failed connectors, access restrictions, unavailable documents and conflicting identifiers instead of silently writing around them. Next, pin the policy and guideline versions. Coverage rules change. Clinical guidelines change. Formularies, benefit designs and authorization requirements change. A packet should name the exact policy document, effective date, jurisdiction, plan and section used for each conclusion. It should distinguish a payer rule from a clinical recommendation and both from a model inference. This sounds fussy because it is fussy. Healthcare decisions are full of exceptions, amendments and facts that look similar until the consequence lands on a person. The agent's reasoning trail should be reviewable without requiring a reviewer to trust hidden internal thoughts. Show the observable work: queries sent to each system, documents returned, fields extracted, rules applied, calculations performed, citations attached and confidence or ambiguity flags raised. Record the model version, skill version, connector version and configuration. If the system delegates to another model or sub-agent, preserve that handoff. Do not substitute a wall of machine prose for a trail. A reviewer needs a compact sequence that can answer five questions: what was asked, what evidence arrived, what rule was applied, what conclusion was proposed and what remains uncertain. Then put a real person at the consequence boundary. Google says the system provides human oversight. The useful question is where. An agent may gather records, reconcile formats and draft a packet. A clinician should retain authority over diagnosis, treatment and clinical interpretation. A qualified reviewer should make a coverage determination under the applicable policy. The patient or member should have a clear route to correct data and challenge the result. The human gate cannot be a ceremonial click after a machine has already framed the case. Reviewers need time, source access and authority to disagree. Interfaces should make contradictions and missing evidence more visible than the generated conclusion, not bury them behind a cheerful green check. Measure automation bias directly. Give reviewers cases where the agent is correct, cases where it is subtly wrong and cases where the record is incomplete. Track whether people detect the error, how long review takes, which interface cues help and whether workload pressure changes the result. Compare a clean summary with the evidence-first view. If the agent makes the wrong answer easier to approve than to investigate, the design has failed even if the underlying model looks good on a benchmark. Prior authorization deserves its own test floor. Use historical cases with known outcomes, but do not stop at agreement with past decisions. Past decisions can contain inconsistency and bias. Measure whether the assembled packet includes all relevant evidence, uses the correct policy version, identifies exceptions, avoids irrelevant sensitive data and routes ambiguous cases to the right specialist. Track false denials, false approvals, requests for avoidable additional information, appeal rates, time to resolution and differences across conditions, facilities, languages, disability status and other relevant groups. Measure the patient's waiting time, not merely the agent's processing time. Benefits explanations need a different standard. A member asking what care will cost is not asking for an elegant paraphrase. The answer should point to the plan document, state which facts are still unknown and separate an estimate from a guarantee. It should explain deductibles, networks, authorization and exclusions in plain language without inventing certainty. If a price depends on an unprocessed claim or a provider's final coding, say so. The system should not turn administrative complexity into a confident number that later becomes a surprise bill. Access control must follow the task. Google says the agent inherits user permissions and enforces need-to-know boundaries. Test those boundaries with adversarial cases. Can a scheduling worker retrieve a psychiatric note? Can a sub-agent keep data after its parent task ends? Can a copied citation reveal information the user could not open directly? What happens when a person's role changes while a long task is running? Audit logs should record attempted access as well as successful access. They should identify the human user, agent identity, delegated task, source system, fields touched, policy decision and output destination. Security teams need alerts for unusual retrieval patterns, repeated failures and bulk collection. Patients need a correction path. Healthcare records contain errors today. An agent can spread one error across a chart summary, authorization packet and benefit explanation at machine speed. Every output should offer a route to identify the disputed source, submit a correction and show where the old fact was reused. Correction should propagate through derived artifacts. It is not enough to fix the source record while an old summary remains in a downstream inbox. Preview status matters. Google's announcement names healthcare organizations and partners exploring the product, but it does not publish a clinical trial, a representative accuracy study or measured patient outcomes. It does not establish how often the agent misses a relevant record, applies the wrong policy, creates an unsupported synthesis or changes a human decision. The claim that a product supports HIPAA compliance is also not a blanket safety certificate. Compliance depends on contracts, configuration, access, retention, security operations, local law and actual use. A compliant system can still be wrong, confusing or unfair. The practical rollout is therefore boring on purpose. Start with one narrow administrative workflow. Run in shadow mode. Compare the agent's packet with expert review. Publish the error categories. Fix the interface. Add an appeal and correction route. Test access boundaries. Measure patient waiting time and staff burden. Only then let the tool move closer to a consequential decision. The agent may be excellent at gathering the pieces. That is useful work. But healthcare does not need a faster mystery. It needs a visible chain from source record to policy to proposed action, with a human being who can stop, correct and explain it. Keep the evidence gate human.