The model is the cheap layer.

When clients hire us to put AI in production against real data, they are not paying for the model. They are paying for the architecture around the model. That architecture catches the failures the model can't catch on its own.

The model is the cheap layer. The harness is the moat.

That sentence makes more sense when you look at what "production" actually means in this context. Production AI runs against real customer data. It runs while regulated decisions are being made. It runs in front of users who don't know the system just had a bad minute. It runs through model upgrades, prompt drift, and the long tail of inputs nobody planned for. The model produces the response. The harness is what makes the response trustworthy enough to ship.

What happens without one

A model without a harness is fine in a demo. The demo runs the happy path on a quiet dataset with the engineer in the room.

In production, the same model has to handle the hundredth weird input of the day. It has to fail without taking down the dashboard. It has to log enough that a regulator can answer "what did the system tell user X on day Y." It has to keep working when the API has a bad afternoon, when a prompt drifts, when a new model version subtly changes the output shape downstream.

A model alone can do none of that. Every AI feature that "works in the prototype and breaks in production" failed at the harness layer, not the model layer.

What the harness does, in principle

The harness is the gap between AI that demos and AI that runs. It treats the model as an unreliable component that produces useful outputs most of the time, and it builds the surrounding system to make those outputs safe to act on.

That means catching the responses the model gets wrong before they reach the user. Measuring every call so degradation is visible before customers complain. Making sure the system fails predictably when something does go wrong, instead of producing a worse failure no one can debug. Enforcing the rules a regulator cares about at a layer the model cannot work around. Storing what the system did in a form that survives the engineer who built it leaving.

None of that is the model. None of it is glamorous. All of it is load-bearing.

Why model-shopping misses the point

A vendor that pitches you on a model has shown you the cheap layer. The model will be different in a year. There will be a faster one, a cheaper one, a smarter one. Companies that bet on a specific model end up rebuilding the application every quarter.

A vendor that walks you through a harness has shown you the durable part. The harness, if it was built right, doesn't need to change when the model does. The model becomes a parameter, not a foundation.

What to ask a vendor

When you're evaluating an AI partner, the question worth asking is not which model they use. It is what happens when their system gets it wrong. How do they catch it. How does it degrade. How would they prove to a regulator next year what the system told a user last week.

The vendors who can answer those questions have a harness. The vendors who can't are betting on the model behaving.

That is the moat.

— James