Electronic Health Reporter

By Inger Sivanthi, CEO, Droidal Healthcare Solutions.

Every few months, another health system announces it has deployed artificial intelligence across its revenue cycle. The press release follows a familiar script: reduced denials, fastero authorizations, staff hours reclaimed, efficiency unlocked. What almost never appears in that announcement is a second document, the one that defines how the organization will know, 12 months from now, whether any of that is actually true.

That absence is not an accident. It reflects something deeper about how healthcare has historically treated its administrative infrastructure: as a problem to manage rather than a system to understand. And now, as AI tools move from pilot programs into operational deployment at scale, that gap is now creating real operational risk as AI moves into live production environments.

I have spent more than twelve years working alongside revenue cycle teams, coders, billers, authorization specialists, and CFOs, and I can say with some confidence that most of the people closest to this work are deeply skeptical of headlines. They have seen technology promises before. They remember the EHR implementations that were supposed to streamline documentation and instead added hours to the physician workday. They remember the clearinghouse upgrades that reduced one bottleneck and created three others downstream. They are not cynics. They are people who have learned, through experience, that what a system claims to do and what it actually does inside a live operational environment are often very different things.

That skepticism is not resistance to change. It is exactly the kind of operational discipline that should shape how AI gets evaluated and deployed.

The challenge right now is that the industry has skipped that step. Conference stages are crowded with transformation narratives. Health systems facing tight margins and persistent staffing shortages feel genuine urgency to find operational relief. All of that is understandable. But urgency without accountability is how you end up automating broken processes rather than fixing them. And in the revenue cycle, broken processes do not just affect the balance sheet. They affect whether a patient gets a procedure approved on time. They affect whether a physician burns another hour on paperwork that should have taken ten minutes. They affect the trust that providers, payers, and patients depend on to make the system function.

What I find missing in most AI deployment conversations is a straightforward commitment to answering a basic question before the contract is signed: what does success look like, and how will we measure it independently? Through clean, pre-specified performance benchmarks, first-pass resolution rates, authorization turnaround times, denial overturn rates, measured against a documented baseline and evaluated at regular intervals by people inside the organization who are empowered to say when something is not working.

Part of the reason is structural. Revenue cycle operations in most health systems sit in a complicated organizational space, accountable to finance, connected to clinical operations, dependent on technology infrastructure managed by IT, and constrained by payer relationships that nobody controls entirely. That diffusion of accountability makes it genuinely difficult to assign ownership over AI performance. When a denial rate creeps up six months after an AI tool goes live, the question of who is responsible for diagnosing why, whether the technology team, the RCM leadership, or the vendor, rarely has a clean answer. So the question often goes unasked, or gets absorbed into the background noise of operational management.

The other part is cultural. Healthcare administration has a long tradition of accepting complexity as inherent rather than examining it as designed. Prior authorization, to take the most visible example, has become so procedurally dense that many organizations have simply built workforces around navigating it rather than questioning whether the navigation itself can be fundamentally restructured.

The scale of that problem is not abstract: according to CMS, more than 53 million prior authorization requests were submitted to Medicare Advantage insurers in 2024 alone, and of the denials that were appealed, more than 80% were ultimately overturned. AI can reduce the friction of that navigation. But if the underlying logic of the process remains unchanged, if the criteria are still opaque, the payer responses still inconsistent, the documentation requirements still disconnected from clinical reality, then automation speeds up a broken system without healing it. That is a meaningful difference, and it is one that outcome measurement frameworks need to be designed to capture.

What better practice looks like, in my view, is fairly concrete. It starts with a pre-deployment audit with a clear-eyed inventory of where the revenue cycle is actually failing, not where it looks like it might benefit from technology. It requires that AI tools be evaluated against those specific failure points, with defined thresholds for what improvement looks like at thirty, ninety, and one hundred eighty days.

It demands that operational staff, the people who work inside these processes daily, have a formal mechanism to surface when a tool is creating new problems, not just solving old ones. And it insists that model performance be reviewed on a scheduled basis, because the payer landscape does not hold still, and a model trained on last year’s coverage criteria may be quietly degrading against this year’s.

None of this is technologically complicated. It is organizationally disciplined. And that distinction matters, because the conversations health systems need to have about AI accountability are not primarily conversations with vendors. They are internal conversations about how seriously the organization intends to govern its own operations.

Policymakers have a parallel responsibility. As federal and state attention increasingly focuses on prior authorization reform and payer transparency, there is an opportunity to embed outcome reporting requirements into any regulatory framework that governs automated administrative decision-making. An AI system that accelerates a payer’s denial process without improving clinical appropriateness is not a healthcare innovation. It is an efficiency tool for the payer, not an improvement in care decision-making. Regulators should require that distinction to be measurable and reported, not left to vendor interpretation.

The potential here is real. The revenue cycle absorbs an extraordinary share of healthcare resources, resources that could otherwise support direct patient care, workforce retention, or capital investment in underserved communities. Thoughtful AI deployment, governed by rigorous measurement, can free up meaningful capacity across the system. I have seen it work in contained, well-designed implementations. The problem is not that the technology cannot deliver. The problem is that without accountability frameworks, we will not actually know when it does, and we will not catch it when it does not.

Healthcare has spent years debating what AI can do. It is past time to build the infrastructure to find out what it is doing.