How ACOs Prioritize and Use AI

By Kevin Riley

This week I gave a talk at the TXAACO 2026 Conference -> Building the Right Tech Stack: How ACOs Prioritize and Use AI. I was privileged today to get to interview some inspiring healthcare leaders from Clutch Health, Beacon Health AI, & United Vein & Vascular Centers. 

From Large Teams to Mixed Teams

Ashley Medina's story is one of the most concrete illustrations of how AI is changing the structure of health care operations. Not long ago, she led a large national provider operations team within a full-risk MSO aligned with an ACO model. The work was manual end-to-end: Excel-based reporting, Google Forms feeding into spreadsheets, and an Ops Manager dedicated to pulling, cleaning, and analyzing data. "We were working hard," she noted, "but our ability to scale insight and act on it consistently was limited."

Today, her team at United Vein & Vascular Centers is much smaller and much more productive: a Health Plan Liaison, an offshore admin, herself, and AI.

"When I think about non-human worker agents, it's not theoretical," she said. "It's the difference between needing a team to produce insight and having that insight generated, structured, and ready to act on in real time."

In practice, AI has become her analyst, her first-pass strategist, and her quality check. She uses it to turn raw governmental data into market strategy, review contracts and flag gaps, understand line-of-business nuances across Medicare Advantage, Medicaid, and commercial markets, and identify patterns in referral leakage and network gaps. "Instead of spending time building the analysis, we're spending time making decisions and moving faster."

Her message to ACO leaders: start by identifying where your organization is spending human time on structured, repeatable thinking. In ACO environments, that's almost everywhere — reporting and performance tracking, attribution analysis, care gap identification, referral optimization, contract interpretation. 

"If those processes are still heavily manual, you're not just slower. You're introducing variability and missing opportunities."

The implication isn't that teams shrink for their own sake. It's that the composition and purpose of teams fundamentally shift. "You don't need as many people building spreadsheets or chasing down data. You need people making decisions, building relationships, and driving outcomes."

Agents Inside the EHR

Mark Pothen's company, Beacon Health AI, is taking a different angle: building agents that work directly inside the EHR, the same way a human would — logging in, navigating, taking action.

"We try to meet practices where they are," he explained. For those with strong existing workflows, Beacon records a human doing the workflow and spins up an agent to replicate it at scale. For practices without that foundation, Beacon brings the workflow itself. The use cases today are high-value and highly manual: transitional care management, quality measure closure, and risk adjustment.

The human-in-the-loop question is one Pothen thinks about carefully. For now, Beacon staff review every single agent workflow. But that review process is more than quality control — every review becomes a training label, and those labels improve the agents over time. "We've now built up a significant dataset from that," he said, "and we've started training an agent whose job is to benchmark the other agents. It watches their work and flags when something's off — with a confidence score. So the system is starting to quality-check itself."

Even with AI, Personalization Is the Answer

Craig Hauben has spent more than 30 years in health care, and he's direct about what hasn't worked: "Healthcare built a machine that treats sick people like data rows. That's why mailers get tossed. That's why robocalls get blocked. That's why your engagement numbers stay flat even when you spent more on outreach this year than last."

Clutch Health builds behavioral orchestration software for ACOs, payers, and risk-bearing providers — helping them close care gaps by actually moving members to take action. What distinguishes Clutch is that its platform originated not in health care but in retail, where it ran for more than a decade across thousands of brands and billions of consumer interactions. "Retail is unforgiving," Hauben said. "If a message doesn't land, the customer is gone, and the dashboard tells you in 90 seconds. That feedback loop forced us to get channel, timing, and content right years ago."

Applied to health care, the insight is straightforward: personalization is not a marketing trick. It's respect. It's reaching a 71-year-old diabetic patient in Tagalog at 7 in the morning because that's when she's awake, and that's the language she reads. It's texting a 28-year-old new mom instead of calling her at work. "The technology to do that at scale exists. We've been the bottleneck, not the tech."

On where AI agents fit, Hauben is blunt about the pace of change: "Two years ago, an AI agent was a glorified IVR. Today it listens, decides, drafts the right message, picks the right channel, sends it at the right moment, watches for the response, and adapts. No human in the loop unless one is needed. That's not coming. That's shipping right now."

His forecast for the next two years is stark: ACOs and payers will divide into two groups — those that treat AI as a feature they bought from a vendor, and those that treat it as the operating model of their member experience. "The first group keeps buying point tools and wondering why HEDIS and STARs barely move. The second group closes gaps faster, retains members better, and writes the case studies we read."

The Hard Questions

The panel also tackled the more difficult governance and adoption questions that every ACO leader is grappling with.

• On governance: As AI moves from generative to autonomous — from answering questions to executing tasks — the panelists agreed that the biggest risk isn't the technology itself but the absence of structured accountability. Who reviews agent decisions? Who owns the outcome when an agent acts on incomplete data? Building guardrails means defining those accountability structures before deployment, not after.

• On ROI: Where should ACOs look first — back-office automation or clinical decision support? The consensus: start with administrative automation, where outcomes are measurable, and risk is lower. TCM workflows, gap closure, prior auth support, and performance reporting all offer relatively fast returns with clear feedback loops. Clinical decision support carries higher upside but also higher stakes, requiring stronger governance before broad deployment.

• On automation bias: As agents become more reliable, there's a real risk that human reviewers stop checking their work. All three panelists flagged this as a genuine concern. The answer isn't more friction — it's intentional workflow design that keeps humans engaged in meaningful review, not rubber-stamping.

• On data quality: A common objection to AI adoption is that fragmented data across multiple EHRs makes it premature. The panel's view: waiting for a clean data stack is not a viable strategy. Agentic AI can help bridge data gaps — pulling from disparate sources, flagging inconsistencies, and generating useful insight even in messy environments. Imperfect data is not a reason to wait; it's a problem AI can help solve.

What This Means for ACO Leaders

The through-line across the entire conversation is this: AI is no longer a future consideration for ACOs. It is a present operational reality that is already reshaping team structures, workflow economics, and competitive positioning.

The leaders who are seeing results are not the ones who purchased an AI tool and waited for it to deliver value. They are the ones who asked a different question: What work in our organization requires structured, repeatable thinking — and what would it mean to have that thinking happen continuously, at scale, without human bottlenecks?

For ACOs, that question has a long list of answers. The organizations that start working through it now will be better positioned for the next contract cycle, the next quality measurement period, and the next round of value-based care expectations — whatever form those take.