Agentic Index

Best healthcare AI agents ranked on the chart loop (2026)

Data last verified October 21, 2026

The Agentic Index grades 69 healthcare agent vendors against the same 14 capabilities as every other lane. These are independent ratings of healthcare AI agents, rated from each vendor's own public evidence rather than from analyst opinion, vendor briefings or paid placement. 4 of them document the full chart loop. That means the agent reads the patient record, reaches the systems the record lives in, governs who and what can touch protected health information, leaves a trace somebody can read afterwards, and a clinician governs what actually lands. That is 5.8% of the lane. This is a bar, not a leaderboard: a vendor either documents all five in its own public materials or it does not clear, and Partial evidence on any one does not count.

Before the list, the number that should change a procurement conversation. Not one of the 4 vendors that clears this bar documents where patient data is processed and stored. Deployment and data residency is documented in full by 10 of 69 across the whole lane, 14%, and by 0 of the 4 that carry the strongest governance posture in it. These are vendors with SOC 2 Type II, HITRUST and business associate agreements on the public record, and the question a hospital compliance office asks first is the one they do not answer. Treat it as open for every vendor on this page and get it in writing early.

The shape of the whole lane sits underneath that. This field can reach the patient record, and most of it does not document who may. Integrations and tool calling reaches 61 of 69 and workflow orchestration 48, so getting into the EHR, the practice management system and the claims pipeline is close to table stakes. Observability and auditability sits at 33. 24 vendors document both reading the record and reaching it, and of those 24, 6 document no audit trail and 16 document no security and identity governance. Two in three of the vendors that can read your chart and write into your systems do not document the access controls over that reach.

Two more numbers for a vendor call. Oversight is common, and the record behind it often is not. 30 of 69 document clinician oversight and guardrails in full, and 13 of those 30 document no audit trail at all. A clinician signs off and nothing durable records what the agent read to produce the output, what it changed, or on what basis. For work that lands in a legal medical record and gets read back in an audit, a billing dispute or a malpractice matter, that is a gap rather than a control. Ask to see the record of a decision, not the approval screen. And memory is absent and self verification is thin. Memory and state persistence sits at 0 of 69 and testing at 16. The coding lane, by comparison, has memory at 19 of 78 after its second review pass. Persistent state is still the exception in agent markets, not a turn.

One note on the bar itself. Workflow orchestration is deliberately off it. It is documented in full by 48 of 69, more than every capability on the bar except tool calling, and an axis that 70% of a pool clears does no work separating vendors. Every count is printed below so you can build a different five, or a four, and recompute. Disagreeing with this bar is arithmetic rather than research.

The chart loop, and how the 69 vendors score against it

Step	Capability, and what has to be documented	Full	Only blocker
1.It reads the record	Knowledge grounding and RAGThe agent grounds itself in the patient record, the payer policy and the clinical source rather than in the transcript of the encounter alone. Documenting what was said is not the same as reading what is in the chart.	26 (38%)	4
2.It reaches the systems	Integrations and tool callingIt calls into the EHR, the practice management system, the claims pipeline and the scheduling stack, so the work lands where the record lives rather than in a summary somebody has to retype.	61 (88%)	0
3.It governs who can touch PHI	Security and identity governanceRole based access, identity controls and scoped permissions over protected health information. In this lane the agent is a workforce member with access to the chart, and access is the control that matters.	26 (38%)	5
4.It leaves a trace	Observability and auditabilityA durable record of what the agent read, what it wrote and on what basis, readable by a compliance officer, an auditor or an attorney a year later.	33 (48%)	0
5.A clinician governs what lands	Human oversight and guardrailsReview queues, confidence thresholds or sign off gates before anything reaches the legal medical record or a submitted claim. The licensed human stays accountable for the output.	30 (43%)	4

Full means the vendor publishes evidence meeting the capability in its own public materials, under the Agentic Index verification standard. Only blocker counts vendors that document the other four capabilities in full and fail on this one alone. Read the columns in order: reaching the systems is near universal at 61 of 69, the audit trail sits at 33, oversight at 30, and access governance at 26 and grounding in the record at 26 tie for thinnest. Security and identity governance is the most common single blocker, the sole failure for 5 of the 13 near misses, with grounding at 4.

The 4 vendors that document the full chart loop

Ordered by total coverage across all 14 capabilities, ties broken alphabetically, which is exactly what the rankings pages compute. Coverage measures breadth across the whole taxonomy, so a lower score here usually means a narrower product rather than a weaker one. Read what each one operates on before comparing scores, because none of them is built mainly for patient calls or clinical documentation.

1. 1.Cohere Health

   10.5 / 14 capabilities

   Operates in healthcare by: decides health plan prior authorizations and extends into payment integrity, appeals, care management, claims and quality

   The highest coverage score among the four that clear, and the payer side counterpart to the revenue cycle entries. Cohere Unify decides prior authorizations for health plans, approving 85 percent in real time and routing the rest to clinician review, and carries the same clinical intelligence into payment integrity, appeals, care management, claims and quality. It is Full on its API and on evaluation as well as the five on the bar, which most of this lane is not. Sold through a discovery session with no published rates; its return on investment figures are its own.

2. 2.CombineHealth

   9.0 / 14 capabilities

   Operates in healthcare by: runs provider revenue cycle work, coding, billing, A/R and denials, appeals and analytics, as named AI employees

   The revenue cycle as a staff roster. CombineHealth sells named AI employees for coding (Amy), billing (Mark), A/R and denials (Adam), appeals (Rachel) and analytics (Taylor). The design choice that earns the oversight grade is that the customer sets the review thresholds deciding what a human checks, and every decision carries a line by line rationale. Full on prebuilt agents as well as the five on the bar. Sold through a demo or scheduled call with no published rates.

3. 3.Ember Copilot

   9.0 / 14 capabilities

   Operates in healthcare by: prevents claim denials with cited coding checks and recovers the ones that land with drafted, tracked appeals

   The denial specialist, and the one that shows its rule rather than its answer. Ember works both ends of the denial problem: preventing claims from being denied and recovering the ones that are. On prevention it reviews every encounter against coding standards, payer policy and the practice's own contracts, checking CPT, ICD-10, HCPCS, modifiers, NCCI edits and documentation completeness, and returns corrections carrying the source rule from CMS, NCCI or payer policy. On recovery it identifies root cause, retrieves records, drafts the appeal packet with clinical evidence and tracks it through adjudication. It also handles ambient scribing across dozens of specialties and benchmarks payer rates to surface underpayments. United States based cloud, HIPAA and SOC 2, role based access. A $4.3 million seed led by Nexus Venture Partners with Y Combinator, founders out of Google healthcare AI and MIT CSAIL. Demo led with no public rate card, and providers can submit a few hundred cases for analysis inside about three days.

4. 4.Alaffia Health

   8.5 / 14 capabilities

   Operates in healthcare by: reviews health plan claims against the full patient record for payment integrity, utilization management and appeals

   The payer side entry built around the record behind each finding. Alaffia pairs autonomous agents with expert clinicians to review health plan claims forensically against the full patient record, across payment integrity, utilization management and appeals. Proprietary optical character recognition digitizes unstructured itemized bills and records, agents extract and structure the clinical facts and cross reference the claim against the record, clinical criteria and policy, and the most valuable claims are prioritized for human review. Every recommendation carries a clinical rationale and traceable citations, and licensed clinicians validate and sign off, which is what makes a finding defensible to a provider or a regulator. The company states it deliberately avoids the black box denial algorithms now drawing regulatory attention. SOC 2 Type II, HIPAA and HITRUST. A $55 million Series B in February 2026 led by Transformation Capital took total funding past $73 million.

The 13 vendors that miss by exactly one capability

These document four of the five and fail one. 4 of the 13 fail on the same capability, knowledge grounding, and that group includes some of the best known names in ambient clinical documentation. Documenting the encounter is not the same as grounding the output in the record, and this bar tests the second. A miss is a documentation finding, not a verdict on the product: several of these vendors may well do the thing and have not published evidence that meets the standard. If you are already holding a shortlist, this table tells you the single question to put to each one.

Common questions

What are the best healthcare AI agents in 2026?

Four of the 69 vendors in the Agentic Index healthcare agent lane document the complete chart loop in full: the agent reads the record, reaches the systems the record lives in, governs who and what can touch protected health information, leaves a trace of what it did, and a clinician governs what lands. Ordered by total coverage across all 14 capabilities they are Cohere Health, CombineHealth, Ember Copilot and Alaffia Health. That is 5.8% of the lane. Note what the list is not: two decide or review claims for health plans (Cohere Health and Alaffia Health) and two work the provider revenue cycle (CombineHealth and Ember Copilot). No product built mainly for patient calls or clinical documentation clears the bar at this review. Read what each operates on before comparing scores.

Which healthcare AI agents publish where patient data is processed and stored?

Ten of the 69, and not one of the four that clear the chart loop. That is the most uncomfortable number on this page. Deployment and data residency is the question a hospital compliance office asks first, and it is documented in full by 14% of the lane. Vendors that carry the strongest governance posture in the index, with SOC 2 Type II, HITRUST and business associate agreements on the record, are silent on where the processing happens or stop short of a choice. Cohere Health and Alaffia Health are the closest of the four, both at Partial, and Ember Copilot states United States based cloud infrastructure, which is a fact about its hosting rather than a choice a buyer can make. If your organization has a data location requirement, treat it as an unanswered question for every vendor on this page and put it in writing early.

Why do Abridge and the other ambient scribes not appear on the list?

Four of the 13 vendors that miss by exactly one capability fail on knowledge grounding, and two of those four are ambient scribes, Heidi and Abridge; a third scribe, Ambience Healthcare, misses on security instead. This is not a claim that they transcribe badly. It is a finding about a specific and narrow thing: documenting the encounter is not the same as grounding the output in the customer's own record, policy or protocols, and the axis tests the second. A vendor's own curated library of medical literature counts as Partial on the same reading, since the bar asks for a maintained structure over the customer's knowledge. Abridge carries the lowest coverage score of the near misses at 6.5 of 14 and fails on this alone. If you are evaluating a scribe, the question worth asking is what the agent reads from the chart before it writes to it, and what happens when the note and the record disagree.

Do healthcare AI agents keep a clinician in the loop?

Of the 69, 30 document human oversight and guardrails in full, the third most documented capability on this bar after tool calling and the audit trail, so the answer for 43% of the lane is yes. The problem is what sits behind the sign off. Thirteen of those 30 document no audit trail at all. A clinician approves the output and nothing durable records what the agent read to produce it, what it changed, or on what basis. In a lane whose output lands in a legal medical record and gets read back in an audit, a billing dispute or a malpractice matter, oversight without a trace is a governance gap rather than a governance control. Ask to see the record of a decision, not the approval screen.

How many healthcare agents can actually reach the chart?

Of the 69, 61 document integrations and tool calling in full, and 48 document workflow orchestration, so reaching the EHR, the practice management system and the claims pipeline is close to table stakes in this lane. 24 document both reading the record and reaching it. Of those 24, six do not document an audit trail and sixteen do not document security and identity governance. So two in three of the vendors that can read your chart and write into your systems do not document the access controls over that reach, even where they keep a record of what they did. That gap is the finding this page leads with.

Why is workflow orchestration not on the bar?

Because 48 of 69 document it in full, 70% of the pool and more than every capability on the bar except tool calling, and a capability that most of a pool clears does little discriminating work. A bar is only useful where vendors actually differ. The five capabilities on this bar were chosen because they separate the lane, and the counts are printed above so you can recompute the result against a different five, or against four, and disagree with the choice. Every number on this page comes from the same grid, so a bar you prefer is an arithmetic exercise rather than a research project.

How is this different from a KLAS report or a health system AI vendor list?

It grades what vendors publish, not what customers report or what analysts conclude. Every capability is graded against a fixed 14 point taxonomy from the vendor's own public materials, with no vendor briefings, no submissions, no paid placement and no vendor review before publication. The bar is mechanical and stated in full. It also names what each vendor fails rather than only what it does, which is why the near miss table exists and why this page can say that not one vendor clearing its governance bar documents data residency. A list assembled from vendor interviews cannot print that sentence.

Method. This page runs the horizontal grading method over one lane of the index: membership is the Healthcare agent category, 69 vendors out of 1,078 researched, 64 by primary category and 5 by secondary, every one carrying a complete 14 row capability grid, with none excluded for incomplete evidence. The bar is five of those 14 capabilities at Full, chosen because they are the loop a patient record actually travels, and it is stated in full above so you can disagree with it and recompute. Every grade comes from the vendor's own public materials under the Agentic Index verification standard. No vendor pays for placement and no vendor has reviewed this page. This is a capability record, not clinical, legal or compliance advice, and it is not a substitute for your own security review. Data last verified October 21, 2026. How this evidence is graded

Related: all 69 healthcare agents ranked on total coverage, the full healthcare agent capability matrix, how every vendor scores on the audit trail, how every vendor scores on deployment and data residency, how every vendor scores on security and identity governance, Hyro against Luma Health, platforms that require human approval before an agent acts, agent observability platforms, where the trace gets read, platforms meeting enterprise security and compliance requirements, production ops agents ranked for automated incident resolution, compare vendors side by side.