
What healthcare teams need to build before using OpenAI's agents
OpenAI's new agents can copy information from a patient's chart into an insurer's portal the way a staff member would. To do that well, they need a workflow built for them, with the right connections, rules and checks.
On November 16, OpenAI shared how its computer-use agents are being put to work in healthcare (OpenAI, 2026). These agents click, type and move between the apps a team already uses, like an EHR and an insurer's portal. CVS Health has them in production, Cigna Healthcare went from concept to deployment in under eight weeks, and Abridge and Ambience are developing and demonstrating their own.
OpenAI's example will sound familiar to anyone who has worked a front desk or a care team. A clinician recommends an MRI. The information the insurer needs is already in the patient's record, but someone still has to pull it out and type it into a separate portal before the patient can get the scan.
An agent that can see and use a screen the way a person does can take on that copying for the team. That matters because a lot of healthcare software was never built to share information with other systems. Usually, connecting two systems means using an API, which is a direct connection that lets one program send data straight to another. Insurer portals and parts of the EHR often don't offer one, so for years the only way to move information between them has been a person with two windows open. A computer-use agent doesn't need that connection. It signs in, reads the record and fills in the portal form the same way a staff member would.
That doesn't make it plug and play. An agent only does useful work inside a workflow someone has built for it, with the right connections, rules, and checks. Building that well requires high-quality, informed engineering work.
The work is already waiting
Prior authorization is a good way to see how much of this copying work lands on care teams. Prior authorization is the approval a doctor's office has to request from an insurer before the insurer will cover a test, drug or procedure, like the MRI above.
Physicians complete an average of 40 prior authorizations a week, and the work takes about 13 hours of physician and staff time.
Two in five physicians have staff dedicated to it. (AMA 2025 Prior Authorization Physician Survey).
That's repeatable, screen-heavy work, the kind these agents were built for. In a demonstration, Ambience's agent uses evidence from the patient's chart to prepare a prior authorization request on an insurer's website. When the record is ambiguous, it stops and asks the clinician, then waits for approval before it submits.
Cigna's agent is fast because the workflow behind it was built first
Cigna Healthcare's case managers used to navigate more than 10 separate clinical and administrative systems to get the full picture of a member before outreach. Now computer use helps bring that information together in seconds (OpenAI, 2026).
For that to work, a lot had to be in place first. Someone had to connect the agent to each of those systems, decide what to pull from each one, and shape the results into what a case manager needs in front of them before a call. The agent is fast because that workflow was built for it.
In a lot of organizations, the steps an agent needs to follow live in people's heads. They sit in the habits of whoever has done the job for six years, and in the workaround nobody wrote down because it only comes up on Fridays.
What it takes to get an agent ready
Getting an agent ready means building the workflow it runs. That includes connecting it to each system the task touches, writing the rules for when it stops and asks a person, and setting up checks that catch its mistakes before they reach a patient or an insurer. Each of those has to hold up on the days the chart is messy or a screen looks different than it did last week.
Then come the guardrails. Someone has to decide which systems and records the agent can reach, where it hands work to a clinician, and how every action gets logged so it can be audited later.
The platforms give you the controls for this. OpenAI says organizations can add review and approval steps, limit which applications and records an agent can reach, and keep audit logs. CVS Health checks its agents' work with validation checks, human oversight, screenshots and action records. Those controls still have to be designed and set up for your workflow, by people who understand both the work and how the agent behaves.
Why it takes a team
The people who know the work are busy doing it, and the people who can build healthcare agents aren't in the day-to-day work. A useful agent needs both. It takes time with the staff who run the workflow, watching where the real steps differ from the written ones, and it takes healthcare-focused engineers who can turn what they learn into something that runs safely every day. Gumbo brings both, then builds agents and automations in the tools you use.
How we start
At Gumbo, we talk to your people, then build the plan. With Northbound Treatment, that meant hearing from 125 people before the plan. From there we build agents and automations in the tools you use, with clear approval steps so people stay in charge.
If you're looking at agents now
Pick one workflow that eats hours every week. Sit with the people who run it and draw every step, including the ones that only happen sometimes (these can be tricky to remember, so take your time with it…it always pays off to spend the upfront time planning). Mark each place a person has to make a decision, no matter how small.
That map is useful whichever agent you choose, and it's the part a model can't write for you.
There's more on our page for AI Agents & Automation.
AI systemsAI made the draft fast. Sign-off is still the slow part.AI drafts medical content in minutes. Review takes as long as it always did, and most of that time is spent waiting between reviewers.November 11, 2026Steve Caldwell3 min read
Product developmentYour meeting app works three days a yearA rented conference app ends when the meeting does, and the vendor keeps the data. Members and sponsors now expect something that runs all year.November 11, 2026Steve Caldwell3 min read
AI systemsEvery practice you add brings another systemPhysician groups grow one practice at a time, and each one arrives with its own software. Connecting what's already there gets staff their hours back sooner than replacing it.November 11, 2026Steve Caldwell3 min read
Product strategyDoctors tell your MSLs what they need. Most of it never gets used.Only 30% of medical insights get used, though almost every KOL is willing to share them. The loss happens between the visit and the people who could act on it.November 11, 2026Steve Caldwell2 min read
AI systemsThe answer is already written downField reps call the same two specialists for answers that are already in a manual. A knowledge base that shows its source gives those specialists their time back.November 11, 2026Steve Caldwell2 min read
Product developmentIntroducing Roux: The AI Interviewer Built by GumboGumbo builds AI products, and today we're launching one anyone can use. Meet Roux, the AI interviewer that gathers your whole team's input from a single link.October 1, 2026Steve Caldwell5 min read
AI systemsWhat Multiplayer AI Looks Like in PracticeClosing the gap between positive task value and flat or negative system value.August 25, 2026Steve Caldwell7 min read
AI systemsMultiplayer AI: Humans and Agents, Side by SideThe first wave of AI made individuals faster. It is also making organizations messier. April 20, 2026Steve Caldwell4 min read
We'd love to chat.
Bring us a problem, a question, or a crazy idea. Let’s see what we can do together.
Book a call