The Integration Tax: Why AI Pilots Cost More Than the Model

Ask a room of executives what their AI pilots cost and you will get a number for the model. Ask what the programme costs and the room goes quiet. That gap is not an accounting oversight; it is the shape of the problem.

The Bill Nobody Scoped

Inference is the cheapest line item in most AI programmes, and it is falling every quarter. The spend that actually accumulates sits downstream of the demo, in work that rarely appears in the business case:

• Data access: negotiating, provisioning, and auditing entitlements the pilot assumed it already had.
• Integration: wiring the model into systems of record that were never designed for a probabilistic caller.
• Evaluation: building the harness that tells you whether last week’s prompt change broke anything.
• Operations: on-call, observability, and the quiet cost of a system that fails in prose rather than stack traces.

Why It Compounds

A single pilot absorbing this tax is tolerable. The trouble is that most enterprises are running eleven of them, each paying it separately. Every team rebuilds the same evaluation harness, negotiates the same data access, and discovers the same integration constraints six weeks apart. The organisation pays the integration tax once per pilot instead of once per company.

What To Do About It

The fix is not fewer pilots. It is making the second pilot cheaper than the first, which means treating the scaffolding as the product: shared evaluation, shared tool permissioning, shared observability, and one control plane that every agent is obliged to go through.

Pick the workflow where failure is cheapest, instrument it properly, and give it a kill date. Then insist the next team inherits what you built rather than starting over. The compounding runs in both directions.