Before you approve the 2027 AI budget
Four questions that separate an AI budget from an AI price list
Most 2027 AI budgets are built on the vendor price, and many are funded by cutting the people who check the output. Both mistakes come from counting the invoice and missing the review. Ask four questions of every AI line before you approve it. Each takes a few days to answer, and none needs a consultant.
The budget grows on the cheapest line
CFOs are raising AI budgets 15 to 16 percent on average for 2027, and 92 percent plan increases, according to Gartner research published September 28 [1]. Gartner traces the overruns to data preparation, integration, and consulting. Token prices get the headlines. Human review adds to all three and rarely gets a budget line of its own.
The invoice is the smallest cost an AI workload carries. In the illustrative workflow from The Price Is Not the Price, inference costs $8,000 a year out of roughly $868,000. Review, errors that get past review, and evaluation make up the rest. A budget line priced at the invoice covers under 1 percent of that workflow.

Ask: What does this line cost per year once evaluation, review, and rework are counted?
Cutting headcount to fund AI cuts the reviewers
More than half of enterprises (54 percent) are funding AI by cutting SG&A headcount and costs [1]. Gartner warns that cutting staff before AI workflows mature risks operational bottlenecks and lost institutional knowledge.
The people being cut are often the ones checking the output. In a Workday survey of 3,200 employees at companies with $100 million or more in revenue, nearly 40 percent of the time AI saved went to rework: fixing errors, rewriting content, and verifying output [2]. Stanford and BetterUp researchers found that each piece of low-quality AI work handed to a colleague takes nearly two hours to resolve [3]. Both studies rely on self-reported time, and Workday sells AI tools. They still agree: review is real work, and someone has to do it.
Cut review before the workflow is proven, and its cost moves into errors that reach customers, filings, and invoices. In my experience, an error that reaches a customer costs more than the review that would have caught it. One that reaches a filing adds compliance exposure no budget line priced.
Ask: Which roles does this line assume away, and who reviews the output after they're gone?
Usage pricing needs an owner and a stress test
AI spend rises with use. A workload that succeeds gets used more, and usage is what the bill counts. Yet 67 percent of enterprises hold contingency buffers of only 1 to 10 percent against these consumption costs [1].
Someone also has to own the overrun. Gartner's 2027 CIO survey found 73 percent of enterprises have no plans to set rules for who owns technology costs and solutions, even as business units buy AI on their own [4]. Without that rule, the overrun lands on whoever notices it first.

Ask: What does this line cost at twice today's usage, and who owns the overrun?
Few companies measure what they bought
The average AI kill switch trips at a 2.5 percent budget overrun [1]. That trigger stops a good workload that runs over and keeps a bad one that comes in on budget. The right test is whether the workload returns what it was approved to return. Gartner found 88 percent of organizations demand full ROI within two years [1]. Most skip the measurement that would tell them whether they got it.
In my experience they skip it for two reasons. A return that falls short embarrasses whoever sponsored it, and nobody volunteers to publish that. And when people do want to measure, they rarely agree on how: which baseline, which costs, whose time counts.
Measurement pays anyway. At BP I took over Well Advisor, a $128 million program that was failing to deliver. Six months on a single North Sea rig paid back its full cost, and we knew that because we measured it. By the end, it ran on every offshore rig and many onshore ones.
Gartner's survey data points the same way. Among 1,303 respondents, high performers tracked returns continuously, managed AI as a portfolio, and stopped underperforming projects. They reported positive returns on 81 percent of their AI initiatives. Low performers didn't know the return on 29 percent of theirs [5].
Settle both problems at approval, before anyone's reputation rides on the result:
Agree on the measure before money moves.
Track operating numbers that are cheap to count and hard to dispute: volume, review minutes per output, error rate, and cost per completed task. Derive ROI from those.
Write the stop rule into the approval, so stopping a workload means the plan worked.
Have finance read the numbers each quarter, in place of the sponsor.
Ask: What will you measure, who reads it, and what number stops this workload?
Run the questions before the vote
Four questions per line fit the budget cycle you already run. For a whole AI portfolio, the two-week inventory in The Price Is Not the Price runs them line by line.
If some of your 2027 AI lines are priced on the invoice, I can help you find them before the vote.
Notes
[1] Gartner, "2027 AI Budgeting Assumptions: CFOs Misjudge Cost Volatility and Stifle ROI," Nauman Abbasi, September 28, 2026. Client research; public summary at https://www.gartner.com/en/documents/8426313.
[2] Workday, "New Workday Research: Companies Are Leaving AI Gains on the Table," press release, January 14, 2026, reporting the study "Beyond Productivity: Measuring the Real Value of AI." Survey fielded by Hanover Research in November 2025: 3,200 full-time employees, all active AI users, at organizations with $100 million or more in annual revenue. https://newsroom.workday.com/2026-01-14-New-Workday-Research-Companies-Are-Leaving-AI-Gains-on-the-Table
[3] Niederhoffer, K., Kellerman, G. R., Lee, A., Liebscher, A., Rapuano, K., and Hancock, J. T., "AI-Generated 'Workslop' Is Destroying Productivity," Harvard Business Review, September 22, 2025. Research by BetterUp Labs and the Stanford Social Media Lab. https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity
[4] Gartner, "CIO Planning for 2027: How to Close the AI Ambition Gap," August 31, 2026, drawing on the 2027 Gartner CIO and Technology Executive Survey. https://www.gartner.com/en/articles/cio-planning-for-2027
[5] Gartner, "Gartner Survey Finds Only 22% of Organizations Have Successfully Scaled AI Across Multiple Business Units," press release, September 1, 2026. Survey of 1,303 respondents, January to April 2026, at organizations with at least $50 million in annual revenue. https://www.gartner.com/en/newsroom/press-releases/gartner-survey-finds-only-22-percent-of-organizations-have-successfully-scaled-ai-across-multiple-business-units
© 2026 David Nix. Licensed under CC BY-NC-ND 4.0.
