Healthcare fraud, waste, and abuse drain an enormous amount out of the system every year. The National Health Care Anti-Fraud Association puts the conservative figure at three percent of total health spending, and some government estimates run as high as ten percent, which would be more than $300 billion a year. When the leakage is that large, the pull toward more aggressive AI detection makes complete sense.
What the past two years have also made clear is how badly the other side of this can go. AI decisions nobody could explain, legitimate claims swept up in an automated net, denials that no human ever actually reviewed. Those failures have produced lawsuits, unflattering headlines, and a sharp increase in regulatory attention. So, the question facing payer leadership has quietly shifted. It is no longer about whether AI can catch more, because with roughly 84 percent of large insurers already using AI somewhere in their operations, that part is settled. The harder question is whether a payer can catch more without creating friction it cannot stand behind later.
The natural instinct is to chase detection accuracy, as though a sharper algorithm were the whole answer. In payment integrity it usually is not, because a more powerful model that cannot show its reasoning quietly becomes a liability. Every flag it produces might be challenged by a provider, examined by a regulator, or pulled into an audit, and a black box has nothing useful to say in any of those rooms. What actually limits a program is whether the decision still holds up once someone asks how it was reached.
What friction costs a payer
Friction is not a soft cost. It shows up directly on the ledger and in the risk register:
• Provider abrasion and network strain when false positives put legitimate practices under suspicion
• Appeals and rework volume that quietly consumes the savings the detection program was meant to create
• Member harm and the reputational exposure that follows a wrongful denial
• Compliance and audit risk when a determination has no reviewable trail behind it
A program that generates these costs faster is not winning. It is moving the loss from one column to another.
The responsible approach to payment integrity has settled into a fairly clear pattern, and the leading players are open about it. Cotiviti describes its fraud, waste, and abuse work as a combination of machine learning, expert rules, and human analysis working together, rather than detection handed off to an algorithm and left alone. The wider industry is tightening its own governance expectations too. AHIP’s commitments taking effect through 2026 say that determinations should come with clear, understandable explanations, and that any decision resting on clinical factors stays under review by a qualified professional. Those principles increasingly set the tone for how responsible AI is expected to behave across payer operations.
The enforcement climate makes the stakes very concrete. The Department of Justice recovered more than $2.9 billion through False Claims Act settlements in fiscal 2024, with healthcare accounting for the single largest share, and it has since stood up a dedicated fraud enforcement division as scrutiny continues to build. In that environment, a payment-integrity decision that cannot be explained stops being a purely operational concern and starts looking like legal exposure. Hybrid, governed operations have stopped being the cautious choice. They are quickly becoming the standard the regulators and the category leaders are writing into the rules.
Agentic AI earns its place in payment integrity by taking on the investigative work that surrounds a determination, while leaving the determination itself to people. Picture a single flagged claim. Instead of handing an investigator a bare alert, an AI Worker gathers the supporting evidence, pulls the provider’s billing patterns and claims history, drafts a structured case file that lays out its reasoning, routes it to the right reviewer, and logs each step along the way for audit. The agent absorbs the hours of reconstruction that normally slow an investigation down, and the human still makes the call.
That arrangement is what “without friction” actually looks like, and the friction it clears runs in two directions at once. Investigators get fuller, better-prepared cases faster. The provider on the other end of the claim is also far less likely to be wrongly frozen or dragged into an appeal, because the evidence has already been assembled and weighed before anyone acts, and a clear explanation travels with whatever decision the member or provider eventually receives.
None of this works if explainability is only decorative. Real governed explainability means every flag arrives with its evidence attached, the reasoning is something an investigator or compliance officer can follow without a data science degree, there is a complete trail running from the first signal to the final determination, and people hold clearly defined authority over the decisions at each stage.
That same inspectability turns out to be a payer’s best defense against a quieter danger, which is bias. A model trained on years of historical claims can gradually learn to over-flag particular kinds of providers, specialties, or patient populations, and what started as an artifact in the data hardens into a pattern of unfair scrutiny. In healthcare that pattern carries genuine regulatory and reputational consequences. A black box hides it until the damage is already done. When every flag comes with inspectable evidence and a traceable rationale, a payer can watch for disparate impact, ask why one category of claims keeps surfacing, and adjust the model before the problem becomes systemic. Human review at the point of decision adds a second safeguard, catching the cases where a statistically reasonable flag would still have been the wrong call. Preventing bias is not some separate program bolted on at the end. It is part of what governed, human-reviewed operations were built to do in the first place.
ThoughtFocus earned the 2025 PROGRESS in Lending Innovations Award for a platform run by a workforce of AI agents that handled complex tasks end to end, inside the compliance demands of a heavily regulated financial domain. That is essentially the same agentic model payment integrity calls for. Combined with explainability built to show why a given decision was reached, it is the kind of engineering that makes the firm a credible fit for governed payment-integrity work. The point was never to build a sharper detection model. The harder and more valuable work is the operating discipline wrapped around the AI, so that speed and accountability show up together rather than at each other’s expense.
That experience is grounded in healthcare, not borrowed from somewhere else. With more than a decade working in the sector, ThoughtFocus helps payers and providers modernize exactly the kind of process-heavy, compliance-bound operations where payment integrity lives, using AI and automation designed to hold up under scrutiny. The same rigor that keeps a regulated healthcare workflow accurate and auditable is what allows an AI Worker to operate inside an investigation without ever becoming an unexplainable black box.
The payers who pull ahead over the next few years will not be the ones running the most aggressive model. They will be the ones who can catch more and still stand behind every decision when a provider, a regulator, or a reporter comes asking. Governed, hybrid, explainable payment integrity is how a plan earns detection and trust at the same time, and that is the only version of fraud detection that actually scales without friction.
