OCS Labs wants AI to show its work

A physician spends part of a day seeing patients, part teaching, part doing research. Weeks later, someone asks how the hours were divided. The answer may travel into a spreadsheet, then a cost report, then an audit. Somewhere along that journey, a recollection acquires the appearance of a fact. This is the problem that makes OCS Labs worth a look.

The Boston-associated business, founded by Paul Swider, describes itself as an AI lab drawing on RealActivity’s work. Its proposition is to take methods developed around healthcare’s constraints and adapt them elsewhere. Finance, manufacturing and education appear among its intended markets. The interesting export is a way to keep an AI system accountable.

A memory becomes a record

RealActivity’s provider platform starts with an unglamorous task: allocating physician effort. It brings together information from enterprise resource planning and electronic health record systems, covering clinical work, research, teaching and administration. Providers review the assembled record and confirm their allocation. Hospital leaders can then use the dataset for reporting, compensation and workforce planning.

The distinction matters. A hospital needs to understand work that does not end in a billable patient encounter. Teaching and supervision still consume time. An incomplete account can make those hours hard to recognize. RealActivity’s response is continuous collection followed by a person’s verification, rather than a short sample reconstructed later.

“Memory is not data.”

RealActivity’s stated operating principle

RealActivity says its methodology has operated for more than fifteen years across a fifteen-hospital academic health system and is accepted by CMS as an alternative to manual time studies. Those are the company’s reported credentials. The long-running methodology should also be distinguished from the newer agents built around it.

Ava, its Activity Verification Assistant, sits inside Provider Activity Studio. It helps a provider question the record and complete the review. The physician retains the allocation decision. That division of labor is specific enough to borrow: decide which judgment belongs to a person before giving an agent a conversational interface.

The agent gets a patient

The patient side begins with a different sort of fragmentation. A lab report sits in one portal, medication information in another, a question for the doctor somewhere in the patient’s head. Tula, RealActivity’s open-source health-agent project, brings records and documents into a self-hosted workspace. Its skills include pulling records through SMART on FHIR, preparing visit materials and drafting portal messages.

The name has a personal origin. Swider’s mother, Tula, died from brain cancer; the project is named in her memory. The documentation puts caregivers alongside patients as intended users. That choice helps explain why organizing information across encounters matters as much as answering an isolated question.

Tula’s public code is licensed under Apache 2.0. Aria is the commercial organizational platform around those skills, with identity, consent, policy and audit controls. My Aria is the patient-facing experience. The commercial products remain in active development at pilot stage. A working personal agent and a hospital deployment carry different responsibilities.

The bill, and the error messages

Tula’s reference deployment guide estimates sixty to ninety minutes of setup and roughly $30–$50 a month for an Azure virtual machine plus API tokens. It assumes comfort with a command line. That is a personal deployment estimate, subject to usage and configuration. A hospital has integration, oversight and operating requirements beyond that small server.

The guide preserves the awkward parts: private-key permissions that block access, package-installation permission errors, a service name that differs from expectations, a bot waiting for pairing approval. These are documented setup problems, rather than evidence of a failed company strategy. They show how much ordinary systems work stands between available code and a useful agent.

For OCS, the stated commercial routes are framework licensing and bespoke development. RealActivity separately sells paid discovery, readiness assessments, governance planning and implementation. Its engineers work inside customer environments. This places the offering near specialist enterprise services, where the buyer needs someone to connect software, institutional rules and the people responsible for both.

What travels beyond healthcare?

Wheelhouse AI CoE is the framework OCS lists for licensing. Training, governance tools and incident reporting sit alongside custom AI research. An organization considering it is buying a method for deployment as well as technical help. The relevant alternative may be its own engineering team assembling those practices, or another consultancy doing the integration.

The transferable lesson is straightforward: capture evidence, preserve its origin, assign approval and keep a usable record. Healthcare-specific assumptions need fresh examination in another industry. A financial workflow has its own authorities and reporting duties. A manufacturing workflow has different consequences when an automated action goes wrong. Shared controls still require local design.

There is a useful recent example of keeping claims bounded. In October 2026, the Tula project reported passing SMART Health Check-in self-tests using a synthetic patient chart. Its documentation distinguished those results from live connectathon testing planned for November. A standards test is a defined achievement; it does not settle every question about deployment.

The approach also has conditions. Provider verification needs accessible records and clinicians who can review them. Self-hosting needs someone to maintain the workspace and understand its model connections. An organization unwilling to assign responsibility for consent or escalation would leave the proposed controls unfinished. Tula’s stated remit is health organization and literacy, rather than diagnosis or treatment. Choosing the right job for the software is part of making the design work.

A team can copy that discipline immediately. Pick one workflow. Specify the evidence an answer must carry. Give someone authority to approve consequential actions. Measure the review burden. OCS Labs’ proposition depends on organizations willing to do that work. The demo becomes more valuable when the next person can inspect what happened.