XCaliber Health

The coordination layer is the product and it is agent driven: Merlin role based agents act across EHR, billing, and scheduling systems in real time. Notably the company blends generative AI with traditional machine learning and microservices so deterministic tasks stay deterministic, which is a more disciplined architecture claim than treating everything as a model problem.

Among the clearest autonomy disclosures in the index because it is calibrated rather than uniform: agents run routine administrative workflows end to end, a human in the loop structure retains control of critical decisions, and the company states explicitly that the degree of autonomy varies by whether a task is operational or clinical. Distinguishing autonomy level by task consequence, and saying so, is the right design and a rarer disclosure than it should be.

No foundation model is named, no evaluation methodology is published and no accuracy figure was located for any agent.

One architectural statement deserves credit and is unusual enough to record. The company describes blending generative AI with traditional machine learning and microservices specifically so that deterministic tasks stay deterministic. That is a more disciplined position than treating every problem as a model problem, and it implies the team has thought about where probabilistic output is inappropriate. It is an architecture claim rather than a performance disclosure, and nothing published says which tasks fall on which side of that line.

The published figures are outcomes rather than model performance: chart review time reduced by 80 percent, prior authorisation turnaround from 14 days to under 24 hours, quality reporting accuracy improved by 30 percent, no shows reduced by 30 percent. Those describe workflow results and are useful, but none states an error rate, a confidence threshold or how often an agent is wrong.

Nothing identifies any party in the chain: no model or model family, no foundation model provider, no hosting arrangement and no sub processor list was located in two passes, and no retention position or statement on whether customer data trains the agents was found. The write behaviour makes one question specific to this record and it is not a privacy question in the usual sense.

A platform authoring more than one hundred and sixty thousand record updates daily across more than seven hundred thousand patients is adding content to the permanent clinical record, and nothing published states whether agent authored entries are identifiable as such in the chart.

That matters for a reason this index has not had to raise before: a clinician reading a note months later should be able to tell which content a person wrote and which a system generated, because the two carry different weight when deciding whether to rely on it, and a record that loses that distinction cannot be audited afterwards either. Provenance of authorship is the record level equivalent of the model provenance this axis asks about elsewhere. Ask whether agent authored entries are marked in the chart, what the vendor retains after an agent completes a task, and for a sub processor list.

Operational scale is specific and substantial for a seed stage company: more than 8 million chart updates processed, over 160,000 EHR updates generated daily across more than 700,000 unique patients, and a reported average of six hours of manual work saved per provider per day on prescription refills alone. Held back from A because the figures are vendor reported without named customer references or published methodology, and the company is early, with a $6.5 million seed announced in May 2026.

No retention period, no statement on whether customer data is used to train or improve the agents, and no de identification posture was located.

The write volume makes this more consequential than for a read only tool. Generating over 160,000 record updates daily across more than 700,000 patients means the platform is not only holding protected health information but authoring it into the permanent clinical record, where it persists, is relied on by clinicians, and follows the patient.

Two questions belong in diligence. What is retained by the vendor after an agent completes a task, and whether agent authored entries are identifiable as such in the chart. The second matters because a clinician reading a note months later should be able to tell which content a human wrote and which an agent generated, and nothing published says whether that provenance is preserved.

Noted as context rather than credit: the chief executive previously co founded a data privacy infrastructure company, which is relevant background but is not a disclosure about this product.

No business associate agreement terms and no HIPAA posture document were located. The word HIPAA appears alongside SOC 2 in a statement that the deployment models support compliance, which describes a capability rather than a commitment the company has made.

Business associate status is structurally certain: the platform is reported to process more than 8 million chart updates and generate over 160,000 electronic health record updates daily across more than 700,000 unique patients.

That volume is the reason to press. An agent writing to the record at that rate is acting on protected health information continuously and at scale, and the agreement governing it should be read before deployment rather than after. Ask specifically how the agreement treats agent generated content written into the chart, since the vendor is originating record entries rather than only reading them.

The company names HIPAA and SOC 2, but the phrasing is that each deployment model supports enterprise grade security, performance and compliance with them. For a platform offered on premise as well as in the cloud, supports compliance reads at least as plausibly as the platform being deployable inside a customer's compliant environment as it does as the vendor holding an attestation of its own.

No SOC 2 report type is stated, no report or trust centre was located across two differently phrased searches, and no other framework is claimed.

Context rather than excuse: this is a seed stage company that announced a 6.5 million dollar round in May 2026, and a mature attestation programme would be unusual at that stage. The note states that plainly so the C reads as a stage observation rather than a judgement on the team. It remains a real gap for a product taking autonomous action inside health system systems of record, and it is the first thing to ask for.

No FDA pathway applies and none is claimed. The platform coordinates administrative and operational workflow rather than diagnosing or selecting treatment, and the company is explicit that clinical decisions stay with humans.

Graded C because several regimes do reach the functions being sold and no position is published on any of them. Prior authorisation work sits under the CMS interoperability and prior authorisation requirements and the state laws now conditioning AI involvement in coverage decisions. Quality reporting automation feeds CMS quality programmes where the accuracy of submitted measures carries its own exposure. Patient outreach and scheduling contact reaches the consumer telephone consent framework where it is outbound.

The quality reporting function is the one worth pressing. A platform that improves reported measure accuracy is shaping what a health system attests to a payer or regulator, and a buyer should establish what review sits between an agent's output and a submitted measure.

No AI governance framework, model monitoring disclosure or bias evaluation was located.

The autonomy design is genuinely good and is graded highly on its own axis, with the degree of autonomy stated to vary by whether a task is operational or clinical. Governance is the layer above that and it is missing: who decides which tasks sit on which side of the line, how that classification is reviewed, and what monitoring detects an agent drifting from its intended scope.

Two functions raise concrete equity questions. Selecting a cohort of high risk patients for outreach determines who receives proactive contact, and flagging high risk discharges for post acute follow up determines who gets attention after leaving. Both are allocation decisions made by a model about people who never see it, and nothing published addresses whether the selection performs evenly across patient groups.

Two passes located no error rate, confidence threshold or evaluation methodology for any agent, and no warranty, indemnity or remediation commitment. The published figures describe workflow outcomes rather than model performance, covering chart review time reduced, prior authorisation turnaround shortened, quality reporting accuracy improved and no shows reduced, and none of them states how often an agent is wrong.

One architectural statement deserves credit and is unusual enough to record: the company describes blending generative models with traditional machine learning and microservices specifically so that deterministic tasks stay deterministic. That is a more disciplined position than treating every problem as a model problem, and it implies someone has thought about where probabilistic output is inappropriate, though nothing published says which tasks fall on which side of that line.

What makes the missing error rate consequential here is the write volume. The platform is described as generating more than one hundred and sixty thousand record updates daily across more than seven hundred thousand patients, so it is not reading the record, it is authoring into it at scale, and what it writes persists, is relied on by later clinicians and follows the patient. A wrong entry at that volume is not corrected by a reviewer noticing it. Ask for per agent error rates, which tasks are deterministic, and what human review precedes a write.

Coordination across systems is the entire premise, so integration breadth is the product rather than a feature: EHR connections spanning Epic, Cerner, athenahealth, and eClinicalWorks plus billing and scheduling platforms, exposed through a unified data layer that harmonizes data from multiple sources, with listings on the athenahealth and AVIA marketplaces. Marketplace listing is independent validation of integration quality.

The company states flexible deployment across cloud, on premise and hybrid environments. An on premise option is genuinely uncommon among agentic platforms and it matters, because it removes the data transfer question rather than merely locating it: a health system running the coordination layer inside its own infrastructure is not sending chart content anywhere.

Held at B rather than A because the statement is a menu rather than a specification. No hosting provider, region, tenancy model or subprocessor list is published for the cloud option, and nothing describes what differs between the three in terms of what the vendor can access or what leaves the customer environment.

Ask which option a proposed contract actually covers, and specifically where language model inference runs under each. A platform blending generative and traditional models can be on premise for the deterministic parts while still calling an external provider for the generative ones, and that distinction decides whether record content leaves the building.

No public pricing. Contact the vendor. Sold as an enterprise platform to health systems and provider groups through a direct sales motion the company states it is still building out with seed funding, which is worth noting: an early direct motion means implementation and support capacity are themselves worth diligencing.

Coverage is broad but enumerated by function rather than vaguely claimed, which is what this axis rewards. Named workflows span quality reporting and compliance monitoring, prior authorisation, hospital throughput and bed utilisation, patient navigation across care benefits and billing, refill and adherence management, discharge risk flagging, referral and care coordination tracking, and readmission monitoring. Buyer types are named as integrated delivery networks and management services organisations, and user roles as nurses, case managers and administrators.

Held at B rather than A on maturity rather than clarity. This is an early stage company with one agent line shipped and a wide stated span, so a buyer should establish which workflows are production proven in a comparable setting rather than assuming uniform readiness across the list.