Doug Marcey, CTO at Coronis Health, does not want humans doing robotic work. Across 700 clients, 154 electronic medical record (EMR) systems, and around 100,000 cases a week, he wants them doing the work that requires human judgment. Everything else is a problem for what he calls accountable intelligence. 

Coronis Health announced its UiPath platform agreement in March 2025, initially pursuing a parallel strategy of traditional Robotic Process Automation (RPA) alongside intelligent document processing (IDP) for medical record extraction. Eighteen months later, the implementation has grown significantly - with Coronis building proprietary layers on top of the UiPath platform's RPA, IDP agentic automation, and orchestration capabilities. 

I spoke with Marcey at UiPath Fusion 2026 in Las Vegas, the same day that UiPath announced general availability of Cartographer (its process mapping product), Process Atlas, and Coding Agents, alongside the Map of Work concept and the decision ledger - a continuous mechanism that captures human overrides and feeds them back into process definitions. Several of these announcements apply to problems that Coronis Health was already addressing independently. 

Healthcare revenue cycle management is a particularly demanding environment for automation, with thin margins, system fragmentation, and regulatory complexity. Coronis Health takes in medical records from providers, turns them into claims, submits them to payers, and manages denials, appeals, and collections. The company is contingency-based so the economics of every tool decision can have an impact, right down to transaction level. 

Coronis Health works across 154 EMR systems and more than 25 practice management systems. The anesthesia division - the company's most valuable and most idiosyncratic service line - alone processes around 60,000 documents per week, including medical records, explanation of benefits, and lockbox feeds. Marcey described the environment's scale and fragmentation: 

Healthcare technology being what it is, it hasn't really come into the 2020s or even the 2010s yet when it comes to interoperability. Even when we have the ability to pull data programmatically from the system, we often don't have the ability to write back account information or claim data the way we need to. So we've been relying heavily on agents and RPA bots for what I think of as the last-mile piece - us writing back into the various different systems of record that our clients use, so that we can consolidate our workforce, have a more consistent single pane of glass approach for them, and then have the bots take the data and put it back into the right system at the end of the day.

This last-mile capability has delivered what Marcey says is the biggest operational win from the UiPath platform so far. Before the integration layer existed, Coronis Health had staff that were trained on specific EMR systems, and could not be moved between teams when demand shifted. The generic adapter layer provided by UiPath's RPA and agentic functionality removed this issue of workforce mobility. UiPath's Database Hub - which reads and writes to systems like SQL Server, Oracle and Databricks without custom code - and MCP Connector, which lets workflows call tools from any compatible MCP server as a native activity, takes away this type of integration challenge. 

One of the most technically significant parts of Coronis Health's architecture is document processing. With 100,000 cases a week, cost per transaction is as much of an optimization point as accuracy. Marcey's team has built a tiered system to tackle this, he explained: 

We have a pretty sophisticated platform now. We have a homegrown model that does a discriminator, looks at the record, figures out what's the best approach to deal with the record, and then routes it to the right point depending on how complex and how messy the record is. 

The tiers range from Apple's built-in Optical Character Recognition (OCR) running on Mac Studio hardware, processing around 30 pages per second (compared to roughly three pages per second from Poppler); standard IDP and Tesseract-based extraction; an open-source OCR model; and a commercial visual language model running on Coronis Health's own hardware for the most complex records. The pipeline also incorporates UiPath's IDP with Action Center for human-in-the-loop correction. 

This architecture mirrors the principle that UiPath presented this week with its deterministic-versus-cognitive framing - route to deterministic, tokenless execution wherever possible, and escalate to cognitive token-based processing only where it meets a certain level of complexity. Coronis Health was driven to its own development and implementation due to the economics of thin-margin, high-volume healthcare billing. Marcey elaborated: 

Medical billing is a weird profession where 50-75% of what we do should be fairly straightforward. Payers make it interesting. And then you have the complicated stuff - the Medicare cases where somebody has two additional insurances, or the trauma cases where prior authorizations weren't possible. We want to try to route to the cheapest possible way to handle it if we can. 

Coronis Health started its IDP pilot with the anesthesia division which can be the most difficult for document processing, with 14-20% of records still arriving handwritten. IDP could not handle those. For the 80% of records that are digitally sourced or scanned print, the combination of IDP and Action Center has worked, but with a very interesting metric. Marcey: 

If you look at the average, it's only moved down about 30 seconds on average. But what's interesting is the standard deviations have tightened a lot. The amount of time it takes for somebody to do a record is much more consistent. Whereas before, we'd average this, but then we'd have several cases that would take 10x the same amount of time to process because they're flipping through pages. 

Even though the average handle time stayed relatively static, the operational reality altered considerably. When the time it takes to process a record becomes more consistent, everything downstream gets easier - such as staffing forecasts, cost projections, and service-level commitments to clients. Quality also improved - the percentage of records with at least one undetected human error dropped from seven percent to three percent. 

Healthcare billing has compliance constraints that are complex, to put it mildly. Coronis Health submits claims on behalf of clients - and if the coding is wrong, the provider faces audit liability. Marcey made it clear where the boundary for automation sits: 

We're a services company, and our clients hire us for our experience and the quality of work that we do. We are never going to be in a place where we can go to a client and say, 'oh, sorry, the bot messed up.' That's not our thing. From a compliance perspective, we are going to stand behind the codes that we put on the record, the submissions we've done, and you're not going to end up in court for over-billing because your AI bot decided that they could get an extra RVU [Relative Value Unit] out of this other code. 

If this wasn't complex enough, there is another factor in healthcare billing, which is that the payers - such as United Healthcare and others - are deploying their own AI agents to automatically review incoming records and flag claims, and providers know it. Marcey describes the dynamic as a Cold War, with both sides of the transaction now deploying AI: 

I joked with a doctor recently - it's becoming a little bit like hiring, where you have tools now that help you to make your resume look good for the bot that's going to look at your resume. We're giving feedback to our doctors, saying here's how you can change your record to make your record more easily acceptable to the payers' AI systems. 

Meanwhile, a new wave of competitors is approaching doctors directly and offering AI tools that promise to handle billing without a services company - because of course they are. Marcey's assessment is: 

There's a whole gold rush of people being like, 'oh, it's really easy to build a tool that does X, but they have no idea the level of complexity they're stepping into. We're dealing with 154 different EMRs, over 25 different practice management systems. We don't have the luxury of pulling all the clinical data and things like that. 

Marcey told me that Coronis Health has built its own knowledge graph platform for denial prevention. The graph links diagnosis codes and procedure codes to specific sections in payer contracts, client-specific policies, and internal codebooks. When a claim comes in, the system queries the graph for relevant contract clauses, then passes that targeted context to an agent for evaluation, instead of trying to load an entire provider contract into a context window, which degrades performance at the volumes that Coronis Health handles. He continues:

Just because a record has a set of words doesn't mean those words are going to be reflected in the contract, in the right clauses and the places we need them to. So we have our codebooks, our contracts, our payer policies, all linked together through the codes in the graph. When the claim comes up, we can do that query for similarity, then do the next order in the graph and pull back just those sections of the contract that matter. And then we actually have a separate agent that figures out: 'Is this the right set? Am I missing something? Does this reference something else? Should we have done one more hop?'

A second model generates a root-cause summary for anything flagged, so Quality Control (QC) staff receive a targeted explanation rather than reviewing records blind. The QC team's time shifted from mostly confirming correct work - in steady state, coders were in the 95-98% accuracy range - to reviewing specifically flagged cases with context.

This feedback mechanism is very similar to UiPath's Decision Ledger, announced this week as a component of the Map of Work that captures human overrides with their rationale and feeds curated improvements back into the process definition. Marcey's team had already started building out their own version. He elaborated: 

Right now, we're harvesting ourselves. We have our own tool that keeps track of what the auditors change, and then we are ingesting that back in. We are using a large language model to help evaluate the changes and hopefully improve the prompt store or figure out what we need to add to the knowledge graph. We were already going down the path of having to build a version of that ourselves anyway. So it's always nice when we don't have to do that. 

The same applies to Cartographer - Marcey was very definite that his team does not always have enough business analysts. The collaboration capabilities shown on stage this week address issues he is dealing with right now, where subject matter experts contribute operating knowledge through Teams, including screen recordings and policy documents that the Cartographer agent analyses and integrates.

When it comes to change management, there is no genie in a bottle that can grant staff three wishes, and the company has a workforce with deep tenure - the longest-serving employee in the India region recently marked 27 years. These experts deliver exceptional quality in their work but can feel exposed by the pace  of change. Rather than talking about automation as replacing workers, Marcey gave me one of the best phrases I've heard all week: 

I don't want humans processing spreadsheets. I don't want humans pulling stuff from websites. I don't want people downloading files and manually uploading them. I don't want any of that stuff. But trying to make that transition as comfortable as possible for our long-tenured people has been, frankly, our biggest challenge in the last year. Because everybody's just seeing that the pace of us building stuff has accelerated to the point where it feels as though every week something new is happening. 

A citizen developer program had poor uptake. An enterprise AI chatbot deployment was considerably more successful - staff are comfortable with a conversational interface than a low-code development environment. Coronis Health has already deployed an internal chatbot called Aria in Teams that captures automation ideas from across the company, and Marcey's goal is to connect that directly to UiPath's now generally available to Coding Agents to generate some automations from these ideas. 

The economics of token-based AI fame up after Doug and I had finished our conversation - we talked about token costs and the strategic implications of a model pricing for a contingency-based business. He observed: 

If something goes from 50 cents a case to 75 cents a case, that can literally wipe out our entire profit on that case. I think what I'm expecting to see is that a lot more of our tokens are going to get spent on creating tools that don't use tokens. It's amazing what you can do with simple machine learning models - cost analysis, predictive analytics, the old school stuff. But building these models, getting the data sets together, training them - that requires specialized skills that Claude and others are getting pretty good at. 

This is why Coronis Health has invested in on-prem hardware - despite having been entirely cloud-based before. The fixed cost of hardware is easier to plan around than variable token pricing. Or as Marcey said, AI dragged them back into having a colocation facility. 

This was one of the richest and technically most powerful customer interviews I've done this season. I think one of the most useful metrics he touched on during our conversation was around the importance of the standard deviation, which in this case, tightened dramatically. Consistency is harder to sell than speed, but it lets you plan staffing, forecast costs and make offers to clients. More of this in automation use cases, please!

I've had conversations with several UiPath customers from the healthcare industry this week and some interesting themes are appearing. In the meantime, I think accountable intelligence is an important distinction. When you are the company standing behind the output, would you stake your reputation on it?