A quarter of AI spend is slipping to 2027
Published 15 August 2026
Forrester expects enterprises to defer roughly 25% of planned AI spend into 2027. The stated cause is not disillusionment with the technology - it is that fewer than a third of decision-makers can tie AI value to financial growth, so CEOs are routing AI investment approval through CFOs on ROI grounds. When that happens, the CFO asks the team holding the cost data. That is you.
Deferral is reallocation, not contraction
Read it against the growth numbers. Gartner forecasts worldwide AI spending of $2.59 trillion in 2026, up 47%, with enterprise AI spend at $407 billion, up 34.8% from $302 billion. Those facts are not in tension with a 25% deferral - money is being redistributed rather than withdrawn, concentrating on workloads with a demonstrated outcome and draining from workloads without one.
In a growing budget that redistribution is easy to miss until one of your workloads is on the wrong side of it. See Gartner's 2026 AI spending forecast.
What makes a workload easy to defer
Deferral decisions get made fast and on thin information. Four properties do the work:
| Property | Why it makes a line easy to cut |
|---|---|
| No unit economic | Total spend with no denominator can only be argued as a big number. Big numbers move. |
| Unattributed cost | If nobody can say which feature or team caused the spend, nobody can defend the feature. See cost attribution. |
| An unexplained miss | A forecast that missed once with no stated cause reads as one that will miss again. |
| Qualitative benefit only | "Improves developer productivity" loses to a line with a number next to it. |
Note what is absent: size, and whether the workload works. An expensive workload with a credible unit economic is defensible. A cheap one described in adjectives is not.
Why the forecasting record matters here
Most AI workloads arrive at this conversation carrying the third property. A review of 127 enterprise agentic AI implementations found 73% over budget, some by more than 2.4x, averaging $2.3M in unmodelled cost. Uber gave Claude Code to roughly 5,000 engineers in December 2025 and spent its full annual AI budget by April.
None of those were technology failures - they were estimate failures. But from the seat approving next year's budget, an unexplained overrun and a failed project look identical, which is why the explanation matters as much as the number. See why agentic AI blows the budget.
How to defend a workload
- A measured unit economic with a counterfactual. "$0.31 per resolved ticket against $4.10 handled by a person" survives scrutiny; "$180k a month on support AI" does not. One workflow with a real number beats a portfolio-wide estimate nobody believes. See cost per request as a KPI.
- A full-cost numerator. Inference plus evals, retries, vector storage, and human review. A unit cost built on the token line alone is provably understated, and getting caught understating it costs the whole argument. See hidden LLM costs.
- A variance history with causes. "We were 60% over in Q1; the cause was a 22x call multiplier against an assumed 3x; it is now measured and capped" is stronger than never having missed, because it shows the estimate is instrumented. See anomaly detection.
The role change worth naming
AI investment decisions are moving from engineering discretion to CFO approval. That is a different conversation with different evidence standards, and FinOps teams have generally prepared for the wrong one - they have optimization stories, which answer a question that is no longer being asked.
Ranking workloads on cost per outcome requires attribution, full-cost modelling, and a unit metric per workflow as standing capabilities, not as a scramble when a deferral list appears. That capability now has a name and it is the top skillset FinOps teams are hiring for: see AI value management.
What to do this quarter
- List the AI workloads and mark the ones with no unit economic. That is the deferral candidate list whether or not anyone has written it down.
- Instrument the two or three most strategically important properly - tagged costs, full-cost numerator, a counterfactual.
- Write the variance explanation for anything that has already missed, before it is requested.
- Accept that some should be deferred. Volunteering the weak line buys credibility for the strong one.
Related
- What is AI value management? - the capability the review assumes you have.
- Why agentic AI blows the budget - the forecasting record behind the scrutiny.
- Gartner's 2026 AI spending forecast - the growth the deferral sits inside.
- State of FinOps 2026: 98% now manage AI spend - where the practitioner gap is.
- How to audit LLM spend - the mechanics of building the case.
Want this applied to your own LLM spend? FinOps LLM runs a free audit of your AI costs and shows where the savings are. Book free audit →
FAQ
How much AI spend is being deferred to 2027?
Forrester's 2026 predictions expect enterprises to defer roughly 25% of planned AI spend into 2027, as the gap between vendor promises and delivered value narrows appetite and financial scrutiny increases. It is a reallocation inside a growing budget, not a contraction.
Why is AI spend being deferred if the market is growing?
Both are true at once. Gartner forecasts $2.59 trillion of worldwide AI spending in 2026, up 47%, while individual programmes get moved out a year. Money concentrates on workloads with a demonstrated outcome and drains from workloads without one.
What makes a workload easy to defer?
Four properties: no unit economic, cost that is not attributed to a feature or team, a forecast that already missed once with no stated cause, and a benefit described only qualitatively. Any one makes a line easy to move; two makes it likely.
Why does this land on the FinOps team?
Because CEOs are routing AI investment approval through CFOs on ROI grounds, and the CFO asks the team that owns the cost data. A FinOps function that can report spend but not spend-per-outcome supplies only half the answer at the moment it is asked for.
What should we do before the review?
Pick the two or three workloads most likely to be challenged. Give each a measured unit economic with a full-cost numerator, a borrowed business denominator, and a documented forecast variance with its cause. That is weeks of work, not a quarter.
Should we defend every workload?
No. A function that defends everything defends nothing. Volunteering the genuinely weak line buys credibility for the strong one, and it is a better position than having the list written by someone with less context.