Both Sides Have AI Now. The Denials Are Still Winning.

Last week I wrote that AI took the stage at HFMA, and buyers came for evidence instead of theatre. This week the evidence showed up. It is not flattering.

The story everyone is telling right now is an arms race. Payers deploy AI to process claims faster. Providers deploy AI to fight back. Two sides, two models, may the better algorithm win. Here is what the script leaves out: it is not a fair fight, and the side that is losing already knows it.

The numbers buyers actually got this quarter

Insurers denied more claims on clinical grounds in 2025 than in 2024. Kodiak Solutions put the cost of that shift at roughly a 25% increase in net revenue leakage. The mechanism is not a mystery. Payers stopped sampling claims and started reviewing the full dataset, which means the old strategy of “they probably won’t catch it” is dead.

Meanwhile, the provider side of the arms race is mostly aspiration. Guidehouse’s 2026 Revenue Cycle Management Trends report found that only a sliver of organisations have fully integrated AI across their revenue cycle. Most have not started. A separate survey of revenue leaders across eighteen specialties found that payer behaviour has now overtaken staffing as the single biggest threat to revenue growth — the first time in years the top risk is external rather than internal.

Read those two facts together. One side has AI in production, pointed at everyone’s claims. The other side has AI in a pilot deck, pointed at a moving target. That is not an arms race. That is a head start.

Why the better model does not win

The arms-race framing assumes the contest is model versus model. It is not. The payer’s AI is trained on one rulebook and aimed at thousands of providers. Your AI is aimed at a rulebook each payer rewrites every quarter. The asymmetry is structural, and no amount of GPU spend closes it.

More importantly, the model was never the moat. Clean data and a tight workflow are. An appeal bot bolted onto messy documentation does not win appeals — it automates rework faster. If the denial was born at registration, eligibility, or prior authorisation, the smartest appeal engine in the world is just an expensive way to clean up a mess that should never have left the front desk.

The quiet part: who actually gets crushed

A three-hospital system has a revenue integrity team, a denials analytics group, and the leverage to get a payer on the phone. A four-provider independent clinic has one biller and a hope that the claim goes through.

When payers accelerate with automation, the large system feels friction. The small practice feels free fall. Smaller providers simply cannot catch the changes fast enough, and the gap is widening, not closing. The AI era was sold as a great equaliser for healthcare. In the revenue cycle, it is doing the opposite.

What to do instead of buying another bot

• Fix the front end before you automate the back end. Most denials are conceived at registration, eligibility and authorisation — not in the appeal. Money spent there returns more than money spent on appeal automation.
• Build payer-specific playbooks and keep them current. The rules change quarterly. Track the changes deliberately, or you are appealing against a rulebook that no longer exists.
• Measure first-pass yield and overturn rate — not “AI adoption”. Adoption is a vanity metric. Cash collected on the first submission is the only one that pays rent.
• Do not hand judgment to a model you cannot audit. If you cannot explain why a claim was coded or appealed the way it was, you have not reduced risk. You have hidden it.

The line that matters

AI did not level the playing field in the revenue cycle. It tilted it, and it tilted it toward the side that already had scale, data and leverage.

The providers who come out ahead in 2026 will not be the ones with the most AI. They will be the ones whose data was clean enough, and whose workflow was disciplined enough, for AI to matter at all. Everything else is theatre — and the payers stopped buying tickets a long time ago.