NAO MEDICAL | CarePlan, the Physician Pre-Visit Copilot
A physician pre-visit copilot in the EHR: trends, open loops, and top actions ready before every visit
BACKGROUND
Nao Medical is an outpatient healthcare provider whose physicians deliver time-constrained visits using longitudinal records spanning encounters, labs, medications, and referrals.
THE OPPORTUNITY
Physicians spent valuable prep time reconstructing patient context, while subtle trends and open care gaps could remain buried in the record. Surfacing those signals before the visit could reduce prep time, improve quality performance, and help clinicians focus the encounter.
OUR APPROACH
Ajaia built CarePlan as a Chrome extension layered over the existing EHR, which meant no migration, no vendor change control, and no disruption to how physicians already work. The build ran in six phases.
We started with the physicians, not the technology. Through structured interviews and a ranking exercise with clinicians at Nao Medical, we mapped where pre-visit prep actually consumed time and where information was most often missed. Clinicians ranked longitudinal trend analysis highest and rated the referral loop as the most valuable workflow to close. Those two became the product's anchors rather than secondary features. We also defined what CarePlan would deliberately not do, which set the guardrails for everything downstream.
Insight quality depends entirely on data freshness. We built automated retrieval so CarePlan pulls the latest encounters, labs, medications, and referrals on its own, without a physician prompting a refresh or an administrator running an export. Every record ingested carries its source and date, which is what makes citation possible later. Where data was incomplete or stale, we designed the system to say so rather than infer around the gap.
We engineered for auditability from the first commit. Outputs are deterministic, so identical input produces identical output across runs. Every claim CarePlan surfaces links back to a dated entry in the patient's record, so a physician can verify a statement in one click instead of trusting it. PHI moves only to a BAA-covered, zero-retention endpoint, never to a consumer AI service. Security review, access controls, and audit logging were completed before any clinician touched the tool, not retrofitted after.
This was the largest phase of the build and the one that determined whether CarePlan shipped.
Ground truth benchmarking. We assembled a golden dataset of de-identified patient records and had clinicians independently review each one to establish what a correct output looked like. CarePlan's trend detection, open-loop detection, and recommended actions were scored against that clinician-adjudicated baseline, measuring both what the system caught and what it missed.
Citation and grounding validation. Every generated claim was tested for traceability to a dated source in the record. Any output that could not be grounded was treated as a defect, not a rounding error.
Determinism and regression testing. An automated regression suite runs against fixed reference charts on every prompt, model, or logic change. Output drift is caught before release rather than reported by a physician.
Adversarial and edge-case testing. We tested against messy real-world conditions: conflicting entries, duplicate records, missing labs, medications discontinued but never removed, and referrals with ambiguous status. Medication suggestions were tested against active regimens specifically to confirm the system would not propose something that conflicts with what a patient is already taking.
Alert fatigue calibration. We tuned sensitivity thresholds with clinician input so CarePlan flags what warrants attention and stays quiet otherwise. A tool that surfaces everything gets ignored, which is a product failure even when the model is technically correct.
Failure behavior. When evidence is insufficient, CarePlan states that explicitly instead of producing a confident guess.
CarePlan ran in shadow mode first, generating outputs that clinicians reviewed side by side with the chart without affecting the visit. That gave us a direct read on accuracy and usefulness at no clinical risk. We then rolled it to a small pilot cohort of physicians with a structured feedback loop, refining trend logic, action ranking, and interface density based on what clinicians actually used versus what they skipped. Features that did not earn their place in a 20-minute visit were cut.
Deployment happened per user through the browser, so onboarding took minutes and required no EHR downtime. We paired rollout with short workflow-based training rather than feature training, so physicians learned CarePlan as part of how they open a chart. Post-launch, we monitor usage, suggestion acceptance rates, flagged outputs, and prep time, and feed that telemetry back into the regression suite so the system improves against real clinical use rather than assumptions.
The governing principle: confirm, never auto-execute
Across all six phases, one constraint held. CarePlan is a copilot, not a decision-maker. Referrals and medication suggestions are presented as options and confirmed by the clinician in one click. Nothing is ordered or executed on a physician's behalf. Ajaia assesses the appropriate level of AI autonomy for each industry and client, and in an outpatient clinical setting the answer was clear: the system prepares the decision, the physician makes it.
THE SOLUTION
A Chrome extension over the EHR that turns a medical record into a pre-visit summary, flagging care gaps and visualizing data trends.
RESULTS
Both outcomes are ones Nao Medical group can see directly in its own operations.
Prep time fell because physicians no longer reconstruct context. They open the chart to an assembled picture: trends already identified, open loops already surfaced, and the highest-value actions for the visit already ranked. The work that used to happen in the minutes before an encounter now happens automatically before the chart is opened.
Quality and compliance measures improved because the misses that used to surface downstream now surface upstream. Lab values drifting inside the normal range, referrals placed months ago with nothing back on file, and follow-ups that quietly fell off get raised before the patient sits down rather than at the next audit. Closing those loops consistently is what moved the measures.
That consistency carries into quality-bonus capture. The measures a practice is paid on are the same ones CarePlan keeps in front of the physician at the moment they can still act on them, so quality performance and revenue performance move together instead of being tracked separately after the fact.
Adoption held because trust was engineered into the product rather than requested from clinicians. Every suggestion cites dated evidence from the patient's own record, nothing executes without physician confirmation, and identical inputs produce identical outputs. Physicians verify rather than assume, which is why the tool became part of the workflow instead of another window they close.
Ready to see similar results?
We'll help you turn uncertainty into an actionable plan built for measurable impact.