Eighty-five percent accuracy sounds impressive - until someone doesn't get paid. 

That was one of the themes from multiple customer conversations at UiPath FUSION 2026 last week. In many enterprise AI conversations, if agentic AI gets it right 85% of the time, that's considered a success. In healthcare - where a miscoded claim means revenue being denied, a missed eligibility check delays patient care, an overbilled procedure leads to a government compliance audit - 85% looks more like a risk. 

During the event in Las Vegas, I spoke to two healthcare companies about what it takes to deploy AI in this environment: Medline, a manufacturer and distributor of essential medical supplies with automation embedded across its supply chain operations, and Omega Healthcare, a 40,000-person services company that manages revenue cycle operations for hospitals and health systems across the US. 

Sarah Stokes is the Director of Business Process Improvement for Supply Chain Solutions at Medline. Her role sits between business teams and IT, turning requirements into technical projects and technical constraints into language that the business can act on. In person, Stokes is incredibly modest about what she does, but from across the table, Medline's whole automation program depends on the strategic translation and coordination that the role involves. 

It started in accounts payable, processing thousands of invoices a day. Stokes walked into the CFO's office and said they needed to automate. Getting sponsorship helped to unlock funding, prioritization, and organizational buy-in. On the other hand, it meant deploying Robotic Process Automation (RPA) and Optical Character Recognition (OCR) - two major technologies, at the same time. This required significant levels of change management and training, which was a big ask for the business already working with high volumes of data. 

As it turns out, data was the biggest surprise. Medline maintained duplicate vendor records for a reason, and when automation hit those records, Stokes told me that the invoices wouldn't process correctly: 

Master data was probably the biggest challenge we were not expecting. We had duplicate vendor records intentionally because we'd segregate by different divisions. Boy, did that cause challenges from an invoice posting perspective.

Governance grew organically from there. A conversation about security morphed into architectural review boards, solution review boards, and a responsible AI board with multiple review checkpoints. Stokes explained that there was a healthy tension between the business pushing for speed and IT enforcing caution - this produced resilient, compliant solutions without hindering progress.

Moving to agentic automation opened up use cases that RPA wouldn't have been able to manage - where a bot follows steps, an agent reasons, compares, and makes decisions. Medline is using agents to solve what Stokes calls "human reasoning bottlenecks", the point in a process where a person has to think rather than just click. She explained that the need for precision is non-negotiable:

AI is hard, and it takes time. We were working on agent prompts, and that probably took us almost a year to get it just right. One of the keynote speakers [at UiPath FUSION] was talking about being 85% confident in the agent. That doesn't work in our world.

Stokes also emphasized that Medline's leadership position is that AI is not a headcount replacement tool, but a way to move employees toward higher-value, critical-thinking work. And automation proved how valuable and resilient it could be when COVID disrupted the healthcare supply chain - teams could spin up or adjust bots in minutes to contain downstream breakage and get ahead of problems that would have taken weeks to resolve with manual processes. When asked what she'd tell someone starting the same journey, Stokes came back to partnership:

You have to have a partner that's got equal skin in the game, because if you don't, you won't be successful together, and you're going to find traps in weird places that you can't anticipate. 

With more than 40,000 employees and 350-plus clients, Omega Healthcare manages the end-to-end process of making sure hospitals and health systems get paid for the care they deliver. Its CTO, Senthil Velayutham, has a strong background in software and AI, and his colleague Gautam Char, Chief Strategy Officer, has been driving the automation strategy at the company for longer. 

Omega Healthcare's ambition is to shift from roughly 25% software and 75% human today, to a 60/40 mix within three to four years. One thing that stands out is the focus on outcomes over inputs. As Velayutham put it:

Even though we try to measure how much is automation and how much is manual, what we ultimately care about is the outcome. The outcome our customers care about is that they provide services worth a million dollars - they would like to collect a million dollars. Their fair share, not more, not less. 

Omega Healthcare is increasingly going to market with outcome-based pricing, charging against the improvement it delivers rather than per transaction. As Char said, "We win more when the customers win more." It may sound simple, but the complexity underneath it is intense. With combinations of employer benefits, payer rules, provider contracts, and visit-specific variables, it is almost impossible for any single person to know every permutation. Char described three categories of automation that the organization deploys across its workforce: things that people shouldn't have to do at all, tools that simplify complex work so that a broader range of employees can handle it, and processes that are reimagined entirely - taking out the human steps altogether where the outcome can be reached more directly. 

Data governance is handled on a client-by-client spectrum - some customers are comfortable sharing de-identified data for model training, others allow customization but no sharing. Some permit a model, but no further training, and some say "Zero AI", and Omega Healthcare accommodates all of them.

If AI over-bills by coding a procedure at a higher level than the documentation supports, it can lead to a government compliance audit. If it under-bills, the hospital loses revenue. Char explained why there is no margin for error:

Human-in-the-loop is an interesting concept in our business because unless you're truly, truly sure this AI is accurate, a precision of 95% is not necessarily good in this business. If ChatGPT questions are 50% accurate, you live with it. Our business is a little more challenging. 

One thing that sets Omega Healthcare apart from many automation stories is that it has no culture resistance problem. With five years of automation under its belt, the practice has been normalized, and because the company is growing, job anxiety is alleviated. Char described it as "the reverse problem" - people push his team to automate faster, not slower. Not a bad problem to have. 

The two companies operate in different parts of the healthcare ecosystem - Medline manufactures and distributes the supplies, Omega Healthcare processes the claims and manages the revenue cycle. But there are commonalities. Both found that data quality was the first and hardest obstacle. Medline's duplicate vendor records and Omega Healthcare client-by-client data governance requirements both highlight the fact that enterprise data is messy, and the work of cleaning it up doesn't always get flagged in a vendor demo. Also, governance had to be built, not bought. Governance frameworks had to be developed iteratively over time - this was shown in the evolution of Medline's review boards and Omega Healthcare's tiered data permissions. 

Both found that the architecture had to be reshaped for absolute precision. A human review step can't be tacked on at the end of a process with 85% accuracy that could result in denied claims or compliance audits - the process has to be designed around the AI's limitations from the start. This means routing easy cases through automation and saving human judgment for the exceptions. This is what SMBC's Kei Yamamoto called "AI by design" in his conversation with CEO Daniel Dines, and it's what both Stokes and Char learned in their work with UiPath. 

Healthcare is one of the biggest stress tests against the claims that the enterprise AI industry makes about itself. Volume and compliance aside, the consequences of getting it wrong land on patients who don't get treatment, providers who don't get paid or hospitals that face investigations and audits. Each of the customers I spoke to described AI as augmentation, not replacement. Whether it's a tool to elevate employees, or a reason that staff don't fear for their jobs, people welcome automation when it both takes away the work they don't want to do, and the work that remains may be harder - don't tell me that you don't sometimes love your comfort zone of a task that you've done hundreds of times - but is also more interesting. I know that there are plenty of stories circulating that quote the phrase "higher value work" - but I've heard it from enough companies now to know that it means different things to different people. I've also sat across from spokespeople at other companies and other industries who will tell you, without hesitation, that AI has been helpful in reducing headcount.

One of the other heartening aspects for me was that progress was being made without using the most advanced models - the fundamentals of getting the data and governance foundations right came first, and it was done through iteration and experience rather than a template until the processes worked well enough to be trusted. Stokes' point about partnership also stood out as it was such a practical observation about what can break if the complexity gets too much for the vendor or the customer. 

Next up from FUSION: a conversation with UiPath CPTO Raghu Malpani on the platform strategy behind these deployments, and a closer look at the Dark Testing Factory - UiPath's vision for fully autonomous software testing.