Published
Why AI Projects Should Start With a Fixed Price and a Definition of Done
Most AI projects don't fail because of technology — they fail because of unclear goals, rising costs, and undefined outcomes. Discover why starting with a fixed price, clear acceptance criteria, and measurable business value can turn AI experiments into investments that deliver real results.
- AI Transformation & Enablement
- AI Value & Readiness Sprint
The gap shows up in the numbers, not just in experience
McKinsey's latest global survey confirms what many leaders already sense: almost nine out of ten companies say they now use AI regularly in at least one part of the business. But using AI regularly is not the same as getting results from it. Only about a third of companies say they've actually scaled AI across the whole organization- usage keeps climbing, but real value at scale stays rare.
The gap between companies that "use AI" and companies that get measurable results from it is large- and two independent research teams, looking at different data, landed on almost the same split:
Source | What was measured | The result |
|---|---|---|
McKinsey, global survey | Companies regularly using AI in some part of the business | ~88% |
McKinsey, global survey | Companies that have scaled AI across the whole organization | ~33% |
McKinsey, 1,933 companies surveyed | Companies where AI drives more than 5% of profit | 109 companies- about 5.5% |
MIT Project NANDA, ~300 AI deployments reviewed | Corporate pilots with no measurable effect on profit | ~95% |
MIT Project NANDA, ~300 AI deployments reviewed | Pilots that delivered real, extractable value | ~5% |
The pattern is the same every time it's measured: a large majority of companies are "using" AI, and a small minority- somewhere between 5% and 10%- are the only ones actually getting paid back for it. That's not a fluke of one survey. It's a sign of how these projects get started, not bad luck for a few unlucky companies.
Where the money goes, and why it's not really a tech problem
MIT's research also shows where the budget actually goes. Roughly half of generative AI spending went into sales and marketing, while some of the clearest returns actually showed up in quieter back-office automation work- companies spent where it was visible, not where it paid off. Researchers also noticed a pattern of companies trying to build their own AI tools in-house, and those homemade tools tended to underperform compared to proven, ready-made approaches.
More than 40% of agentic AI projects will be canceled by the end of 2027, because of rising costs, unclear business value, or weak risk controls. The problem isn't the technology itself.
The same pattern shows up in a newer category: agentic AI. Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027, because of rising costs, unclear business value, or weak risk controls. The problem isn't the technology itself. Most agentic AI projects today are still early-stage experiments or proofs of concept, driven more by hype than by a clear plan- which hides the real cost and complexity of running AI agents at scale from the people funding them.
Three separate sources point to the same root cause. Projects don't fail because the models aren't good enough. They fail when nobody defined the task, the boundaries, and what counts as success- before the money went out the door.
What this means for how an AI project should start
Open-ended, pay-by-the-hour consulting- "let's start and see where it goes"- hands that exact risk to the client: unclear value plus rising cost, the same combination Gartner names as the reason projects get canceled. The client gets billed for hours long before anyone knows whether the whole thing will pay off.
That's the thinking behind our own rule for starting some AI projects: a fixed price and agreed acceptance criteria before work begins, not after. Before a client signs anything, they see a costed plan and a clear definition of "done"- not an open-ended budget with no clear finish line.
This is exactly what our AI Value & Readiness Sprint is built for. It's a fixed-price sprint that doesn't start with "what's trendy to build"- it starts with "where in this business would AI actually pay off, and what needs to happen to get there." The sprint is a diagnostic step: it either finds a specific place with a measurable return, before the company becomes one of the 95%, or it honestly shows that now isn't the right time for a bigger investment. Either outcome is useful, and either one comes at a price the client knew about from day one.
Why this works better than "let's just start"
All three sources above point to the same diagnosis: the gap isn't about the technology, it's about how the project gets set up. The companies that win ask the value question before they ask the tooling question. A fixed price with acceptance criteria isn't just a safer deal for the client. It's a direct fix for the exact problem McKinsey, MIT, and Gartner all describe from different angles: projects with no clear definition of success get expensive precisely because nobody can point to the result.
Engineering discipline is what keeps a demo from staying a demo. It's what turns "we regularly use AI" into a number that actually shows up on the P&L.
None of this requires more AI. Most organizations already have enough of it sitting unused. What they're missing is the discipline to notice which parts are actually working, and the willingness to stop funding the parts that aren't.
