Why companies aren't shipping the AI they already bought
Most enterprise AI pilots don't reach measurable business value, and the reasons are rarely the model. The common causes: nobody defined the value as a number before starting, the data wasn't ready, the team lacked people who can build and run AI in production, the tool didn't fit an existing workflow, nobody owned the outcome, and there was no feedback loop to improve it after launch.
What it looks like
Budgets were approved, licences bought and pilots demoed. Six months later, usage is low and nobody can say what changed. Some pilots are still running because nobody decided to stop them, and the company can't name every AI project it's paying for.
What the research says
- In Gartner's survey of 644 organisations, published in May 2024, 49% named estimating and demonstrating AI value as the top barrier to adoption.
- MIT NANDA's 2025 report "The GenAI Divide", as reported by Fortune, found about 95% of generative AI pilots had little or no measurable impact on profit and loss, and attributed this to a learning and integration gap rather than model quality.
- The same reporting says purchased tools and partnerships succeeded about 67% of the time, internal builds about a third as often, and the highest returns came from back-office automation even though most budgets went to sales and marketing.
- Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027 because of rising costs, unclear business value or inadequate risk controls.
My reading of the 95% figure: it's real, and it's also what an early period of experimentation looks like. The fix is to experiment with a defined metric and a feedback loop, so the experiments teach you something.
Six causes
| Cause | What you see | Page |
|---|---|---|
| Value never defined | "It's a strategic initiative", no baseline | ROI and cost estimation |
| Data not ready | Pilot used clean samples, production data is messy or locked | AI readiness assessment |
| No one to run it | Built by a vendor or one person, nobody maintains it | Structuring an AI team |
| Outside the workflow | Separate tool, people forget to open it | POC to production |
| No owner | Engineering owns the model, nobody owns the business number | AI strategy on one page |
| No feedback loop | Quality unknown after launch | Measuring AI adoption |
One-line test
For every AI initiative you own, complete this sentence: "This will move [metric] from [baseline] to [target] by [date], verified by [an owner outside engineering]." If the owner can't complete it, the pilot isn't ready to be funded, not because it's a bad idea, but because it hasn't been designed to produce a measurable outcome.
Pilot graveyard check
Answer five questions. Three or more "no" answers mean you have a pilot graveyard.
1 - Can you name every active AI pilot right now? 2 - Does a business owner, not engineering, own a metric with a baseline for each one? 3 - Has any pilot been officially stopped in the last 12 months because of its results? 4 - Does your CFO know the full AI spend, including engineering time? 5 - Is there a single named person accountable for each pilot's outcome?
From the floor
At MasterDexter, the line we use for this is "Nobody's shipping the AI they already bought." The boring work is where the value usually is, document processing, internal workflows, compliance checks, while pitch decks full of agentic, autonomous and transformative often aren't solving a real problem.
In the founding session of the AI leadership cohort, I walked through killing ideas with a value and complexity view: a competitor monitoring tool that produces reports nobody reads, or a training content generator with no usage data and no learner outcome metric, should be stopped so the budget goes to projects that move a number.