Epic Connectors AI agents expose a contract gap for CIOs
Epic's forthcoming Connectors tool will let its AI agents complete work inside other vendors' systems, raising audit, liability and approval questions most health systems have not written into contracts.
Epic Connectors AI agents are about to test a assumption buried in nearly every health system technology contract: that the entity clicking inside a vendor's application is a person. Modern Healthcare reported in late October that Epic will launch a Connectors tool enabling its AI agents to complete tasks inside other vendors' applications, with availability framed around 2027. The engineering is the easy part. The audit trail, the error liability and the approval pathway are not written yet, and they sit with the buyer.
What Connectors actually changes in the stack
Epic's agent tooling has a visible lineage. The company previewed Agent Factory, a capability for building and orchestrating agents, at HIMSS26 in March 2026, then expanded its AI announcements at its User Group Meeting in August. Connectors is the logical next step: agents that do not stop at the edge of the Epic environment.
CIO press read the move as strategy rather than feature. HealthSystemCIO framed Connectors as reopening the best-of-breed versus in-suite question, because an agent that can act competently inside a niche vendor's application reduces the practical friction of keeping that vendor. It also, less comfortably, reduces the practical friction of replacing it. Either way, the integration logic that health systems spent a decade building around interfaces and APIs now has a layer above it that makes decisions.
Scale is not hypothetical. Epic reported roughly 9 billion records exchanged and about 60,000 organizations participating on TEFCA as of August 2026, and TEFCA overall surpassed 1.7 billion transactions by the fall. Cross-vendor movement of data is routine. Cross-vendor action by software is not.
Most health system AI governance assumes a human user and a single vendor of record. Connectors breaks both assumptions.
Three questions legal and IT should answer before 2027 renewals
The first is permission. Does the third-party vendor's contract actually allow non-human agent access, and does an agent action void support obligations or warranty terms? Many master service agreements were drafted with named users, seat counts and credential-sharing prohibitions in mind. An Epic agent authenticating into a revenue cycle or imaging application may technically violate language no one wrote with agents in view.
The second is evidence. When an Epic agent and a niche vendor's system disagree about what happened, whose log is authoritative? Reconstructing a clinical or financial error requires a defensible record, and two partial logs with different timestamps and different identity models is not that. Health systems should be asking, in writing, for log retention terms, identity attribution for agent actions and a defined process for reconciling conflicts.
The third is approval. Which committee signs off on an agent that was never separately purchased because it arrived inside a platform release? Becker's recently ran a roundup of 15 Epic updates in 30 days, a release velocity that outruns quarterly governance meetings. Prior Becker's reporting found hospitals entering 2026 with documented AI governance gaps, including incomplete inventories of the AI already running in production. Agents delivered through an existing contract will not trigger a procurement review unless someone designs that trigger.
Why HTI-5 shifts more of this burden to buyers
ASTP/ONC's HTI-5 final rule published in the Federal Register on October 28, 2026. Reporting indicates it keeps roughly 20 EHR certification criteria slated for repeal and defers action on information blocking. The EHR Association issued a statement the week prior.
The timing matters more than the content. Certification has functioned as a baseline federal proof point that certain capabilities exist and behave in defined ways. As that baseline thins, exactly as agent-to-agent action expands, less of the assurance comes from a federal attestation and more has to be extracted through contract terms, testing and vendor diligence. Health systems that treated certification as a shortcut for due diligence will need to rebuild that work internally.
A practical starting position for CIOs and CMIOs
Start with an inventory that distinguishes agents from models. Most AI inventories catalog algorithms and their clinical purpose. Few catalog which software has the ability to take an action in another system, under what credential, and with what rollback path. That distinction is the one that determines liability exposure.
Then move the questions into renewal cycles now, not in 2027. Third-party contracts coming up for renewal in the next four quarters are the cheapest place to add agent access clauses, log authority language and warranty carve-out protections. Vendors facing a competitive renewal are more flexible than vendors responding to an incident. Finally, give one named committee explicit authority over agent activation, including agents that arrive inside releases already paid for. The approval gate has to sit at activation, because that is the only moment the organization still controls.
