The AI Model Isn’t Your Advantage, The Knowledge Behind It Is

The following is a guest article by Brian Kenah, CTO at EnableComp

Model capability is climbing fast, and the tooling to build AI agents is now largely commoditized. The strongest models are a license away for any health system or vendor that wants them. That’s led a lot of technical leaders to the conclusion that winning means adopting the best models.

It works until the claims get complicated. Point a capable model at the wrong knowledge and it returns an answer that’s logically sound but confidently wrong. The model is missing something that takes years to build.

The Knowledge is the Hard Part

Getting an agent into production can take days or weeks with today’s tooling. The state-by-state, payer-by-payer knowledge it reasons against, however, takes an entire career to accumulate. A workers’ compensation claim in California follows different rules than the same claim in Texas, and both look nothing like a federal VA claim. Workers’ comp is just one category. Out-of-state Medicaid and motor vehicle claims each run on their own rules.

That knowledge can’t be purchased. It has to be built. Anyone can license the model. No one can license the hard-won judgment that tells a system when a claim is about to fail, and why. 

Why a Better Model Still Gets it Wrong

Most of what an expert knows was never written down. A veteran specialist has learned that a particular state’s workers’ comp carrier won’t pay without a specific attachment, even when the claim record shows as complete. An agent working without that knowledge produces output that looks right and is wrong. It fills every field and passes the internal checks, then the claim is denied for a reason the system never flagged. The model performed exactly as intended. The knowledge it needed had never been encoded.

This is a sequencing problem. Build the data and domain foundation first, then put the agents on top. Teams that reverse the order automate the mistakes they haven’t caught yet. When a payer changes a rule, claims keep being denied under the old one until someone encodes the new one, and every week of that lag is revenue a health system loses. In our work, that lag has dropped from quarters to weeks.

Models Improve, the Advantage Grows

Once the foundation is in place, every gain in model capability improves outcomes against that same encoded expertise. The team that did the hard work inherits every future model release for free, applied to a knowledge base no competitor has.

Domain experts encode what they know, and the system reasons against that expertise. Every claim it works and every correction a specialist makes feeds back in, so the system gets sharper the longer it runs. When it isn’t sure, a human steps in.

The market has already felt the cost of skipping that step. In its latest survey, McKinsey found only a third of leaders expected autonomous coding to handle even 30 percent of outpatient coding volume. A year earlier, about half had expected more than 35 percent. Expectations dropped once the tools ran into work they weren’t built to handle.

From Chasing Revenue to Preventing Losses

A system built on real domain knowledge does more than expedite claims. It flags the risk before a claim goes out and points to the accounts where the return is highest, so fewer high-value claims slip through. Specialists still review every decision that carries real money, which keeps human judgment where the stakes are highest. Because that expertise is written down and governed, the system can show why it made a call, not just what it decided. In an audit or a payer dispute, that is the difference between a defensible decision and a guess.

The result is fewer denials to work, because most never reach that point. The ones that do are expensive to fight, and that fight keeps getting harder as payers get faster and more automated at issuing them. Hospitals spent nearly $18 billion in 2025 overturning claims denials. Prevent them and there’s nothing to overturn. For a hospital, that means higher yield on its hardest claims and revenue the finance team can plan around.

New models keep coming every year. A more capable model without domain knowledge behind it just reaches the wrong answer faster. Pair that same model with expertise you’ve already encoded, and each new release compounds your results instead of repeating the mistake. Your competitors can buy the same models tomorrow. They can’t buy the knowledge underneath, and in complex claims, that decides who gets paid.

About Brian Kenah

Brian Kenah is CTO at EnableComp and the driving force behind EnableComp’s enterprise AI operating model. He has spent 25+ years building and scaling healthcare technology organizations, including CTO at Azalea Health, senior leadership at Sharecare, and 13 years leading product and engineering at NextGen Healthcare.

The following is a guest article by Brian Kenah, CTO at EnableComp

Model capability is climbing fast, and the tooling to build AI agents is now largely commoditized. The strongest models are a license away for any health system or vendor that wants them. That’s led a lot of technical leaders to the conclusion that winning means adopting the best models.

It works until the claims get complicated. Point a capable model at the wrong knowledge and it returns an answer that’s logically sound but confidently wrong. The model is missing something that takes years to build.

The Knowledge is the Hard Part

Getting an agent into production can take days or weeks with today’s tooling. The state-by-state, payer-by-payer knowledge it reasons against, however, takes an entire career to accumulate. A workers’ compensation claim in California follows different rules than the same claim in Texas, and both look nothing like a federal VA claim. Workers’ comp is just one category. Out-of-state Medicaid and motor vehicle claims each run on their own rules.

That knowledge can’t be purchased. It has to be built. Anyone can license the model. No one can license the hard-won judgment that tells a system when a claim is about to fail, and why. 

Why a Better Model Still Gets it Wrong

Most of what an expert knows was never written down. A veteran specialist has learned that a particular state’s workers’ comp carrier won’t pay without a specific attachment, even when the claim record shows as complete. An agent working without that knowledge produces output that looks right and is wrong. It fills every field and passes the internal checks, then the claim is denied for a reason the system never flagged. The model performed exactly as intended. The knowledge it needed had never been encoded.

This is a sequencing problem. Build the data and domain foundation first, then put the agents on top. Teams that reverse the order automate the mistakes they haven’t caught yet. When a payer changes a rule, claims keep being denied under the old one until someone encodes the new one, and every week of that lag is revenue a health system loses. In our work, that lag has dropped from quarters to weeks.

Models Improve, the Advantage Grows

Once the foundation is in place, every gain in model capability improves outcomes against that same encoded expertise. The team that did the hard work inherits every future model release for free, applied to a knowledge base no competitor has.

Domain experts encode what they know, and the system reasons against that expertise. Every claim it works and every correction a specialist makes feeds back in, so the system gets sharper the longer it runs. When it isn’t sure, a human steps in.

The market has already felt the cost of skipping that step. In its latest survey, McKinsey found only a third of leaders expected autonomous coding to handle even 30 percent of outpatient coding volume. A year earlier, about half had expected more than 35 percent. Expectations dropped once the tools ran into work they weren’t built to handle.

From Chasing Revenue to Preventing Losses

A system built on real domain knowledge does more than expedite claims. It flags the risk before a claim goes out and points to the accounts where the return is highest, so fewer high-value claims slip through. Specialists still review every decision that carries real money, which keeps human judgment where the stakes are highest. Because that expertise is written down and governed, the system can show why it made a call, not just what it decided. In an audit or a payer dispute, that is the difference between a defensible decision and a guess.

The result is fewer denials to work, because most never reach that point. The ones that do are expensive to fight, and that fight keeps getting harder as payers get faster and more automated at issuing them. Hospitals spent nearly $18 billion in 2025 overturning claims denials. Prevent them and there’s nothing to overturn. For a hospital, that means higher yield on its hardest claims and revenue the finance team can plan around.

New models keep coming every year. A more capable model without domain knowledge behind it just reaches the wrong answer faster. Pair that same model with expertise you’ve already encoded, and each new release compounds your results instead of repeating the mistake. Your competitors can buy the same models tomorrow. They can’t buy the knowledge underneath, and in complex claims, that decides who gets paid.

About Brian Kenah

Brian Kenah is CTO at EnableComp and the driving force behind EnableComp’s enterprise AI operating model. He has spent 25+ years building and scaling healthcare technology organizations, including CTO at Azalea Health, senior leadership at Sharecare, and 13 years leading product and engineering at NextGen Healthcare.