75% of your AI transformation budget is operating-model work.
Vendors price the technology layer because that is what they sell. The 75% of the spend that determines whether the deployment compounds — operating-model redesign, governance gates, talent overlay — is yours to fund and yours to design.
Walk into any enterprise AI procurement conversation today and the spend conversation is overwhelmingly about the technology layer: which foundation model, which orchestration platform, which agent SDK, which cloud’s Copilot tooling. The vendors who organize that conversation organize it that way because that is what they sell. The technology layer is real, important, and well-priced — and it accounts for roughly a quarter of what an enterprise will actually spend to land an agentic AI deployment that compounds.
The other three-quarters — the operating-model layer — sits below the technology, is unpriced because no incumbent vendor sells it, and is the layer that determines whether the deployment lands in production or stalls in pilot. Most operators discover this twelve to eighteen months in, when they realize the platform works fine but nothing is changing. By then the technology budget has been spent and the operating-model budget has not been requested. This piece names the rule: roughly 75% of an enterprise’s AI transformation budget will be spent below the technology layer, in the operating-model layer. Plan for it explicitly, or pay for it later by accident.
Why the vendors don’t price it.
The technology layer is well-priced for a structural reason: vendors compete to sell it. The operating-model layer is unpriced for the symmetric reason: no incumbent owns it. The closest adjacent category is management consulting, but the major firms organize around advisory hours, not delivery; the closest adjacent product category is identity-and-governance tooling, but those vendors solve narrow technical questions rather than the full operating-model redesign. The space between consulting and tooling — the actual operating-model work — is currently structural orphan. The operators who recognize this organize the budget for it themselves. The operators who don’t treat it as an unfunded mandate that mysteriously grows.
Where the 75% actually goes.
The operating-model layer breaks into four classes of work. Each is real, each is funded one way or another, and each is currently being hand-rolled at every successful agentic deployment in the public record.
The technology layer is finishing its price-discovery cycle. The operating-model layer hasn’t started.
Foundation model prices have fallen ~80% over 24 months. Orchestration platforms are commoditizing. Cloud agent services bundle infrastructure into operating expense at predictable rates. The 25% slice is doing what mature technology categories do — getting more capable and less expensive. The 75% slice is in the opposite phase: capabilities are rising, prices are not visible, and the market has not yet figured out how to charge for the work coherently.
The 75% is where the operator’s competitive advantage actually lives.
The 25% is buyable. Every competitor can buy the same models, the same orchestration platforms, the same Copilot licenses. The 75% — your specific governance design, your talent depth, your decision-rights map, your contracting model — is bespoke to your operation. It is what differentiates a deployment that compounds from a deployment that runs at parity with the technology vendor’s reference architecture. Operators who skip the 75% buy parity; operators who fund it buy advantage.
Skipping the 75% is the dominant pilot-purgatory failure mode.
The pattern is consistent across the public agentic-AI failure record. The platform works. The model works. The agent works in the demo. Three months in, the deployment plateaus — not because the technology fails, but because the governance gate the agent needs to clear hasn’t been built, the constraint engineer hasn’t been hired, the decision-rights conversation hasn’t happened, and the contracting model hasn’t been updated. The operator blames the technology, the vendor blames the change management, and the deployment quietly stalls.
The strongest argument against this position.
The strongest counter is that vendors will eventually package the operating-model work and sell it. The hyperscalers (Microsoft, Google, AWS) and major cloud SIs are already moving in this direction with “agentic AI Centers of Excellence” and “forward-deployed engineer” programs. Under this counter, the 75% gets absorbed into the platform offering over the next 24 months and the operator should wait rather than build.
This is partially right and structurally wrong. Some of the 75% will get packaged: the governance tooling, the audit infrastructure, the standard agent-management primitives. But the load-bearing parts — your specific decision rights, your specific talent overlay, your specific contracting model — are bespoke by definition and not packageable. Vendors will sell you a 75%-shaped bundle that addresses the standardizable pieces; the rest you will still own. Plan for that residual to be 40-50% of total spend, not 0%. Operators who wait will pay the same money two years later, with the platform-shaped constraints baked in.
Three things to do this quarter.
01 · Build a 25/75 budget envelope before you sign the next platform contract. Whatever the technology vendor quotes you, multiply by four to size the full transformation envelope. Get that approved at the board level before the platform decision, not after. The operators who go to the board with the 25% line and discover the 75% later are operating on borrowed runway.
02 · Identify which slice of the 75% is most under-resourced today. For most operators, it is governance gating or talent overlay. For regulated industries, it is governance. For mid-market operators, it is talent. The first dollar of operating-model spend should go to whichever slice will block scale fastest.
03 · Hire one constraint engineer before you hire one more AI engineer. The AI engineering market is saturating. The constraint engineering market is empty. The single highest-leverage hire an operator can make this year is the person who will own the 75% — give them a seat at the operating-model design table, not a desk in the platform team.
SALT’s standing position-review rhythm grades published positions against subsequent reality.