- About 75% of AI transformation spend is operating-model work, not technology.
- Vendor pricing describes the technology quarter, which is why budgets look affordable and then are not.
- Budget for the 75% deliberately, or discover it halfway through and stall.
Here is the number that reframes an AI budget: roughly three quarters of what an enterprise spends getting AI to work lands below the technology layer. Not on models, licences or infrastructure. On process redesign, role changes, governance, data cleanup and the slow work of teaching an organisation to operate differently. Vendors price the quarter. You pay for all of it.
Where the money actually goes
Open any AI programme after twelve months and the spend profile looks nothing like the business case. The licence line is close to forecast. Everything else is larger.
The overruns cluster in the same places: mapping a process nobody had documented, deciding who approves what now, rebuilding reporting because the old measures no longer describe the work, retraining people whose roles changed, and cleaning the data that the pilot quietly worked around.
None of those are technology problems. All of them are prerequisites.
Why vendors systematically underprice it
This is structural rather than dishonest. A software vendor prices what it sells. Operating-model work is not in the SKU, so it is not in the quote.
The effect compounds through procurement. A business case built from vendor pricing looks like a technology purchase, gets approved as one, and is governed as one. Then the operating-model work arrives with no budget line and no owner, and the programme slows to the speed of whoever can absorb it on top of their day job.
The tell is a business case where implementation is a fixed percentage of licence cost. That ratio is a convention, not an estimate.
Roughly 75% of an enterprise's AI transformation spend goes below the technology layer. Vendors systematically underprice it.
How to budget for the 75%
Reverse the arithmetic. Start from the operating-model work, scope it honestly, and treat technology as the smaller line it is.
That means funding process documentation before deployment, naming the people who will own new approval gates, and putting money against measurement so you can prove the outcome rather than describe it. It also means accepting a smaller first scope, because a narrow workflow done completely beats a broad one done to the technology layer only.
Operators who budget this way spend less in total. The work does not get cheaper by being unplanned.
What to do next
- 01Rebuild your next AI business case from the operating-model work up, with technology as a line item rather than the frame.
- 02Fund process documentation as a deliverable in its own right. It is the precondition for everything after it.
- 03Put a named owner against every new approval gate before the agent that needs it is built.