The Boring Update That Makes Healthcare AI Real

Most healthcare AI still sits outside the actual work. ACCESS gives agents a structured update a clinician can review, act on, and bill.

CMS published ACCESS for Primary Care Providers and Referring Clinicians. The model starts July 5, 2026. It runs for ten years on Original Medicare patients with chronic conditions. The point is not another policy layer. The point is a structured care update that arrives at initiation, escalation, and completion. A primary care doctor or referring clinician reads it, does one coordination task, documents it, and bills a co-management fee.

The broader ACCESS model uses technology-supported chronic care and Outcome-Aligned Payments tied to measurable health outcomes. That is the CMS language. The useful part is simpler. The care update now has a source, an owner, a reviewer, and a claim.

The Handoff

That small loop is where agents and models finally meet reality.

An agent that can chat about diabetes is a toy. An agent that ingests the ACCESS update, pulls the latest BP, HbA1c, PHQ-9, or PROMIS score, flags the gap against the EHR, drafts the exact coordination note, and waits for a named clinician to accept or reject it is infrastructure.

The model is only one piece. The harness around it does the real work: validating the schema, keeping PHI inside the right boundary, logging every route, and tying the output to a billable action. Without the harness the model is just another paragraph in the chart that nobody reads.

agent harness

care update -> reviewed action -> billing proof

01

Care update

track, measures, plan, escalation

02

Harness

schema, tests, PHI boundary, trace

03

Agent

drafts review packet and exceptions

04

Clinician

reviews, decides, documents action

05

Proof

EHR note, G-code, outcome evidence

harness gates

schemarequired fields present
sourceCMS update retained
privacyPHI route selected
modellocal or hosted run logged
reviewhuman owner required
billingclaim matched to action

review packet

The output is a packet, not an order.

clinician gate
BP / LDL / PHQ-9measure packet
med list / problem listEHR context
G0676 / G0677 / G0678claim path
reviewer / timestampaudit trail

The model drafts, compares, and flags. The clinician decides what belongs in the chart.

The agent is useful only inside a harness that validates the update, selects the model route, records the source, and requires a clinician review before chart or claim changes.

The Router

Model choice stops being theology and becomes routing. Run the PHI-heavy extraction on a local model inside the clinic's firewall. Use a private endpoint for summarizing long updates when the contract allows it. Send redacted policy comparison or first-draft reviewer notes to a hosted frontier model. Let a plain rule engine count the annual billing limit. The harness picks the cheapest, safest path that still produces inspectable work.

Route everything to the biggest model and you leak risk and money. Keep everything local and you leave reasoning on the table. Good operators route.

model router

local, private, hosted, or no model

routing matrix

PHI stays inside

Local model

extract fields, classify updates, flag missing measures

controlled remote

Private endpoint

summarize long updates with business associate controls

redacted payload

Hosted frontier model

reason over policy text or draft reviewer notes

no model needed

Deterministic rules

validate G-code, count annual billing limits, reconcile claims

trace

PHI extraction

proof: source update

local model

clinical summary

proof: review note

private endpoint

policy question

proof: citation log

hosted redacted

billing limit

proof: claim check

rules engine

source retained

Every model output keeps the CMS update, EHR context, and reviewer action attached.

action bounded

Agents prepare the work. Billing, chart, and referral changes wait for the named reviewer.

The harness chooses the smallest model path that fits the task and privacy boundary. Local inference, private endpoints, hosted frontier models, and deterministic rules all have jobs.

The Handoff Gets Paid

The payment is small, roughly $30 per co-management service after review and one coordination activity, plus a $10 one-time onboarding modifier. CMS says there is no beneficiary cost share. Primary care providers and referring clinicians do not even have to enroll in ACCESS to get paid. They just need to see the structured update and act on it. The technical FAQ puts the payment details in CMS language.

The G-codes are G0676 for cardio-kidney-metabolic, G0677 for musculoskeletal, and G0678 for behavioral health. That is the incentive: pay for the review that usually vanishes into an inbox. The update now has a source, an owner, a reviewer, and a claim. That quartet turns AI from demo to durable system.

Data Is the Constraint

By July 2027 the ACCESS organizations must feed structured data into a CMS Aligned Network, HIE, or similar exchange network so it becomes queryable inside normal EHR workflows. Narrative updates will not cut it forever. The EHR has to hold the source, timestamp, model route, reviewer, action taken, and claim path.

The measures are concrete: blood pressure, LDL-C, HbA1c, weight, uACR, eGFR, PHQ-9, GAD-7, PGIC, PROMIS physical function, PROMIS pain interference, and pain intensity. These numbers do not replace clinical judgment. They give the harness a measurable packet and the clinician a focused set of facts to act on.

The care update is the smallest unit that an agent can prepare, a clinician can inspect, and a billing system can prove.

The Test

Most current healthcare AI fails this test. It has no packet and no paid reviewer at the end. It generates summaries no one is accountable for. It suggests changes no one is paid to consider. It sounds clinical without touching the handoff.

Investors and operators should stop asking which flashy tool a company bought. Ask where the ACCESS update lands, who is named as reviewer, which fields stay structured, which model routes touch PHI, whether revenue cycle can see the right G-code attached to the documented coordination step, and whether the monthly close can reconcile the claim against the original update. The bridge between source and paid review is the artifact to inspect. Everything else is theater.

What Stays Human

Clinical judgment stays with the clinician. ACCESS organizations have to be Medicare enrolled, name a physician Medical Director, meet HIPAA, FDA, and state licensure rules, and submit to CMS monitoring. The primary doctor still knows the patient's full history, local realities, and what actually belongs in the chart.

The system asks the outside chronic care group to send a tight, structured update instead of vague narrative. Agents and models exist to cut the clerical drag around that decision. They must not launder the decision itself.

The Small Test

Draw the path today. Pick one track: cardio-kidney-metabolic, behavioral health, or musculoskeletal. Map the referral to the ACCESS organization, the initiation update to the EHR inbox, the clinician review to the coordination note, the note to the G-code, and the claim to reconciliation. Then map the model routes: local for sensitive data, hosted where safe, rules where deterministic. Name the owner at every step.

When CMS starts publishing risk-adjusted outcomes in Winter 2028, you will see whether the system moved the measures or just produced more reports.

If ACCESS works, it will not be because patients got another app or because a model wrote pretty clinical prose. It will be because the update arrived on time, the harness routed it correctly, a real clinician reviewed and acted, the data landed where the EHR could use it, the claim matched the documented work, and the patient's numbers improved. That is what healthcare AI has to become: less assertion, more inspectable handoff.