Every AI founder selling into large organisations eventually hits the same wall, and most misdiagnose it. The demo goes well. The benchmarks are strong, often stronger than the incumbent’s. Then procurement enters, the process slows, and eighteen months later the contract goes to a competitor with a visibly less impressive model.
The instinct is to blame enterprise inertia. Usually the real answer is that the buyer was evaluating something the founder was not selling, and one industry has just made that gap unusually visible.
The market that made it explicit
American health insurers covering more than thirty million older adults are paid according to the medical conditions documented in each member’s records. Software reads years of clinical notes and extracts codes that determine payment. Real scale, real money, and for a decade a market that bought on extraction performance.
Then federal auditors, scaled to roughly two thousand certified coders on a rolling quarterly cycle, began selecting individual outputs and demanding full justifications reconstructed from data retained at the time. Reviews of three insurance plans published this spring found 81 to 91 percent of certain sampled high-risk diagnosis codes unsupported by the records behind them. A major insurer settled federal claims for 117.7 million dollars over technology-assisted review programmes.
The buying criteria inverted within roughly eighteen months. Performance stopped being the differentiator, because the vendors all cluster within a few points of each other and none of that clustering answers the question an auditor asks.
What the buyer is actually purchasing
This is the part worth internalising if you sell AI into regulated environments. The enterprise buyer is not purchasing capability. They are purchasing the ability to defend a decision that your system made, in a room you will not be in, years after you made it.
That reframes every part of the product. Modern evaluations of ai risk adjustment systems test whether each output ships with its source evidence and the rule it satisfies, whether decisions reproduce under the model and rule versions that existed at the time, whether the system flags items against the customer’s short-term interest as readily as items in favour, and whether the human review step presents evidence rather than conclusions.
Notice that none of those are model properties. They are architecture and data-retention properties, and a competitor with a weaker model and better architecture wins every one of them.
The three founder mistakes
Selling accuracy into a defensibility market. When a buyer asks how accurate the system is, the honest translation of the question is usually: what happens when this is wrong and somebody officially asks me about it? Answering with a benchmark number answers a question nobody asked.
Treating evidence trails as a roadmap item. Provenance cannot be retrofitted meaningfully, because the data was never captured. A vendor promising to add explainability next quarter is promising to explain decisions it did not record, which the sophisticated buyer already knows is impossible.
Building only in a profitable direction. In the healthcare case, systems that only ever surfaced findings increasing customer revenue became the central evidence in enforcement actions. Any product whose outputs conveniently always favour the buyer’s short-term interest is creating a pattern that somebody will eventually query, and the buyer’s counsel knows it before the founder does.
The strategic upside
The encouraging half of this is that architecture is a genuine moat in a way that model performance is not. Performance advantages compress within a release cycle. An architecture built around per-inference evidence, versioned reasoning, and bidirectional correction takes a competitor years to reach, because it requires rebuilding data models rather than swapping weights.
The vendors gaining share in healthcare right now are largely those who built for accountability before anyone demanded it. That looked like wasted engineering for several years, and then the market moved and it looked like foresight.
For founders in any regulated vertical, the question worth asking early is simple. If a regulator picked one output from your system at random and demanded a complete justification, produced entirely from data you already store, could you deliver it? If not, you are competing on a metric your eventual buyer will stop caring about, probably faster than you expect.
