Budget season arrives with a pile of AI projects, each with a champion and none with a clean number. Somebody has to decide, and the honest position is that a good share of them should not be funded.
Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, largely over cost, unclear value and inadequate controls. Better to cancel yours deliberately in October than accidentally in June.
The Three Questions That Sort Everything
- Can you state the outcome as a number? Not improved efficiency. Cost per resolution down from X to Y. Projects that cannot answer this after a year of work are not going to answer it after two.
- Did it reach production? Roughly 78% of enterprises have agent pilots and fewer than 15% reach production scale. A pilot that has not shipped in twelve months is telling you something.
- Who owns it when you stop paying attention? A named person, in their objectives. Unowned AI projects do not fail, they linger, which is worse.
What to Fund
| Category | Why it earns budget |
|---|---|
| Data quality and pipelines | Every other project depends on it and it is always underfunded |
| Gateway, identity and budgets | Controls make everything else safely fundable |
| Evaluation infrastructure | You cannot improve what you do not measure |
| One or two deep production use cases | Depth beats breadth, decisively |
| Training your existing people | Cheaper and more effective than hiring for it |
What to Kill
- The internal chatbot nobody uses. Every organisation has one. Check the usage numbers before the meeting.
- Pilots without a production sponsor. If no operational leader wants to run it, it will never be run.
- Duplicate efforts across teams. Three departments building the same document assistant is common and rarely noticed.
- Anything needing a data foundation you have not funded. Sequence it or cancel it, but do not fund the second floor before the first.
- Projects justified only by not being left behind. That is a feeling, not a business case.
The Line Items People Forget
Three costs reliably missing from AI budgets. Evaluation, which is ongoing work rather than a one-off. Model migration, because providers deprecate and specifications change, as MCP demonstrated this year with a deliberate breaking revision. And the retry and failure cost, since failed runs still bill and a third of them failing means your unit economics are materially worse than the sheet says.
Add a line for each. A budget that assumes none of these is one you will be revising in April.
How to Present It
Lead with cost per successful outcome rather than adoption numbers. Show what you are cancelling before what you are funding, because a plan that only adds looks like enthusiasm while a plan that also subtracts looks like judgement.
Then set a review gate at six months with a pre-agreed kill criterion. Deciding in advance what failure looks like is the single most useful thing you can do for next year, because in June nobody will want to be the person who says it.
Conclusion
Fund the foundations, which are data quality, controls and evaluation, then one or two use cases with real depth. Kill the unused chatbot, the sponsorless pilots and the duplicates. Add explicit lines for evaluation, migration and retry costs. Set a kill criterion now, while everyone is still calm. The teams that look competent in 2027 will be the ones who cancelled early and went deep on something narrow, not the ones with the longest project list.
Frequently Asked Questions
How much should we budget for AI overall?
Less on tools and more on the data and controls around them than feels natural. Most organisations overspend on capability and underspend on the plumbing that makes capability usable.
What if leadership wants more projects, not fewer?
Show the production rate. Fewer than 15% of pilots reach scale, so a longer list mostly raises the number of things that will quietly fail. Depth is the argument, and it is a strong one.
When should we revisit the plan?
Every six months, with the kill criterion agreed in advance. This field moves fast enough that an annual cycle leaves you funding something for months after it stopped making sense.