AI spend in 2026: what earns a place in the 2027 budget?

AI does not need another impressive spending headline. The more useful question is what happens when a company has to approve next year’s bill.

A tool can be popular and still have no clear renewal case. Another can look expensive and be doing work the business would struggle to replace. A budget review has to tell the difference.

For teams planning 2027, the task is not to keep pace with the AI market. It is to decide which commitments have earned another year, which need a bounded test, and which should stop before they become permanent.

The market can justify paying attention. It cannot justify your renewal.

The market is growing. That is not your budget benchmark.

Gartner’s May 2026 forecast puts worldwide AI spending at approximately $2.60 trillion in 2026 and $3.49 trillion in 2027.[1] Both are forecasts, not completed-year totals. These rounded figures come from its published table, rather than its $2.59 trillion headline.

That market includes infrastructure, software, services and other categories. It is not a count of AI subscriptions bought by ordinary businesses. Adding hyperscaler capital expenditure to it would also risk double-counting activity already inside the market estimate.

Keep three questions separate: how much suppliers are investing, how much buyers are paying, and what useful work buyers receive. A server investment announcement answers the first question. An invoice helps with the second. Neither, by itself, answers the third.

Recent transaction evidence is less tidy than an ever-steepening growth curve. Ramp’s September update reports continued but slowing paid AI adoption in its business-spending data.[2] That is a signal from Ramp’s observed population, not proof that every industry or small business follows the same pattern.

The practical conclusion is modest: prepare for AI to remain a budget discussion, but do not copy a market growth rate into your own spending plan.

Useful to a person is not the same as proven on the P&L.

McKinsey’s 2026 survey reports that 80% of respondents report that AI has improved their individual productivity, while 37% attributed a positive enterprise-level EBIT impact to AI.[3] EBIT means earnings before interest and taxes. These are different questions, not stages in a conversion funnel.

The survey does not establish that the remaining respondents received no value, or that their spending was waste. Its results are self-reported, not an audit of financial returns.

For a buyer, the important distinction is between capacity created and money actually released. Finishing a task sooner can improve service or make room for more work. It does not automatically reduce a payroll expense, remove a contractor invoice or cancel a subscription.

Before renewing, ask what changed after the tool arrived. Was more work completed? Did quality improve? Did a bill fall? If the answer is only that people like using it, that is a starting point for a discussion, not a complete financial claim.

The uncertainty reaches finance leaders too. In Deloitte’s Q2 2026 survey of 200 North American CFOs at companies with at least $1 billion in revenue, 46% named cost uncertainty or transparency as their biggest internal AI concern.[4] That is a large-company sample, not a benchmark for smaller businesses.

Compare the cost of finished work, not just the token rate.

OpenAI’s July guidance recommends evaluating cost per accepted outcome: what it takes to reach a usable result, including attempts and human review.[5] A cheaper model can create a more expensive workflow if it repeatedly fails or needs correction. This is provider-authored guidance, not independent proof of any model’s return.

Apply that idea to a task your team can actually inspect. For a proposal-writing workflow, count approved proposals rather than generated drafts. Use comparable work, the same quality standard and the same period when comparing options.

Keep cash costs and staff time visible as separate components. If you convert hours to dollars, record the rate and assumptions; do not relabel that capacity as cash savings. Include failures in the cost and count only accepted work in the output. If nothing meets the quality bar, the metric is undefined, not zero-cost success.

This does not mean every experiment needs a finance project. Start with the commitments large enough, close enough to renewal, or uncertain enough to deserve attention. A short sample of real work can be more useful than a confident score with no supporting evidence.

Give the 2027 budget room to change its mind.

There is no settled answer for how quickly AI capabilities, prices and adoption will change. Anthropic’s September economic scenario explorer models conditional outcomes for 2030, not a single prediction or a 2027 spending target.[6]

Our recommendation for a company budget is simpler than an economic model: write a base case, an expansion condition and an exit condition for each meaningful commitment. These are planning choices, not forecasts about the market.

This makes the annual number easier to defend. You are not promising that demand will double or that a cheaper model will arrive. You are agreeing on what the business will do if circumstances change.

For a tool still being tested, that might mean a short commitment and a review date before the cancellation deadline. For a dependable workflow, it might mean continued funding with a limit and an owner. Contract terms and operational dependencies should determine what is actually possible.

Start with one renewal, not an attempt to fix everything.

Choose a subscription coming up for renewal and get the person responsible for the work into the conversation. Ask:

1. What would stop or get worse if we did not renew?
2. Which completed work supports that answer?
3. What are we paying in total, and which parts could vary?
4. Could something we already buy do the same job at an acceptable standard?
5. What must happen before the cancellation deadline, and who will do it?

Then make the decision explicit. Keep when the case is supported. Review when evidence is missing, with someone responsible for getting it. Consolidate when a replacement has been tested and the transition is workable. Cut when the case no longer holds and the business can exit safely.

Record the result after the action. Flagged spend is reviewable spend. A recommendation to cancel is not captured. Confirm the later billing change and account for replacement costs before counting anything as captured, and do not count the same reduction twice.

The objective for 2027 is not the biggest AI budget or the smallest one. It is a budget your team can explain.

What this note does not claim

Research cutoff: September 12, 2026, and every source was re-checked on September 23. Market forecasts, transaction observations, survey answers and provider guidance measure different things, and this note does not combine them into an estimate of company waste or of savings. The budget framework is our recommendation, not a measured result.

No customer data was used. Both charts are original renderings of the cited figures.