Forrester expects open-weight artificial intelligence (AI) models to carry half of enterprise AI tokens next year. The firm predicts just five percent of the spend for them, because they are good enough for the work that burns the most tokens [2]. The same analyst house found that 45 percent of AI buyers will not call a plan a success unless it pays back inside a year. Only 22 percent report a positive effect on EBITDA (earnings before interest, tax, depreciation and amortization) over the past twelve months. Together, these findings point to a quiet but real trap. The number most teams track is about to detach from the bill you pay.

The big picture:

The spending side is settled already. The firm’s 2027 predictions put AI at 20 percent of all business technology spend [1]. That sits inside a global tech budget the same call puts at $6.1 trillion. No earlier wave of corporate tech claimed that share so fast.

The return side is not settled at all. The firm expects more than half of chief information officers (CIOs) to miss their AI return targets next year. Its reason is about how companies work, not about the tools. In its words, “organizational structures, processes, and ways of working need to change to drive measurable, firmwide results with AI.”

I see this miss coming because of where the money sits. Open Future Forum’s October finance report drew on 389 people. It found that 53 percent expect payback inside six months, yet 31 percent report no clear budget for AI and 13 percent cannot name one owner of the decision. A deadline that short, set against a budget line that absent, is a promise made by whoever sits closest to the project.

The gauge is the weak point. In the same report’s AI Leaders group, 44 percent report a full real-time view of AI running costs, 39 percent a partial one and 17 percent none. That answer rests on 77 people, in a nearby group where 80 percent run production agents. The FinOps Foundation’s October guide adds a scale problem. Token and model charges are, in its words, “only 10 to 25 percent of total AI spend once energy, capital, infrastructure, software, labor, and process costs are included.”

By the numbers

  • 50% — Tokens on open weights: open-weight models will carry half of AI tokens in 2027 against five percent of spend, which breaks the link between volume and cost [2].
  • 45% — One-year payback rule: share of AI buyers who need a positive return within a year before they call a plan a success [2].
  • 22% — EBITDA gain: share reporting a positive AI effect on EBITDA over the past twelve months [2].
  • 53% — Six-month hope: share of finance people who expect payback inside six months, with 26 percent allowing six to twelve and 14 percent unsure [3].
Half Your Tokens: 5% of Your Bill Figures as stated in this article's own numbers section (verified figures, %) Tokens on open weights 50% ([2]) One-year payback rule 45% ([2]) EBITDA gain 22% ([2]) Six-month hope 53% ([3])

What I’d watch:

What I am watching is the gap between moving work and removing cost. A six-month test rewards anything that cuts the invoice this quarter. The cheapest lever on that invoice is a routing change, where the same request goes to a smaller or open model, and that lever is real but it is not the same thing as a saving. This desk’s own record on it is a model priced 50 percent cheaper per token that came out 57 percent pricier per finished task.

  • The missing owner: 13 percent of firms name no single owner and 31 percent have no clear AI budget. The payback promise then belongs to whoever owns the project. It does not belong to whoever owns the profit line it is meant to lift.
  • The blind spot: token and model spend is a tenth to a quarter of the whole. A real-time cost screen measures the smallest part and calls it the total, while power, hardware, software, review labor and process change sit outside it.
  • The early results: a survey of 300 executives, reported by CFO Dive, found that 68 percent ran at least some AI projects over budget last year [5]. Only 9 percent said more than three-quarters of their projects delivered a clear financial return.
  • The next unit: whether the 2027 budget round starts quoting actions, tickets or finished tasks per unit of spend instead of tokens. That call is the first strong case that the token count will come apart from the invoice. Pricing teams tend to follow whatever number their buyers can defend.

The catch

I could be wrong about the timing, but these numbers show a drift, not a precise map. The half-of-tokens, five-percent-of-spend split is a call for 2027, not a reading of anyone’s bill today. The firm has not published the routing choices behind it [2]. The 45 percent figure is a rule and the 53 percent figure is a hope. They come from two groups of people answering two sets of questions, so they sit side by side and never add up. And that 44 percent figure rests on 77 people from a nearby group, not on the 389-person finance base [3].

The larger caution here is about cause and effect. The firm blames the coming miss on company change lagging the tech spend, and nothing in either survey shows that better measuring would have closed that gap. The FinOps Foundation’s band is also an estimate from the Tokenomics Foundation’s framing, not a count of any reader’s own estate. It reads as a warning about what such a screen covers, not as a figure to plan against [4]. What both sets of data do support is narrower and still useful. The volume count and the money count are drifting apart, and the deadline has not moved with them.

At a glance

  • The Big Shift: open-weight models will carry half of AI tokens for five percent of the spend in 2027, while 45 percent of AI buyers demand payback inside a year. Only 22 percent have seen a gain in EBITDA.
  • Why It Matters: the six-month payback test that 53 percent of finance people expect is priced on token spend. That is the smallest and most routable part of the AI bill, so a team can pass it by rerouting work without any business process changing.
  • What I’d Watch: I am tracking three things — whether an owner is named for the AI number, whether the budget line exists at all, and whether buying teams start quoting a unit that survives the shift.
  • The signer: who signs the AI case, and whether that person also owns the profit line the case is meant to lift.
  • The unit: whether requests are priced per action, per ticket or per finished task rather than per token, so the measure keeps tracking the bill.
  • The coverage: whether the cost screen includes power, hardware and review labor, or only model spend.
  • The Catch: the split is a call, the two payback figures come from separate groups, and better measuring alone will not close a gap the firm blames on company change.

Related reading

Sources

[1] Forrester, “Forrester’s 2027 Technology And Security Predictions: AI Will Account For 20% Of All Business Technology Spend,” October 21, 2026 — https://investor.forrester.com/news-releases/news-release-details/forresters-2027-technology-and-security-predictions-ai-will [2] The Register, “AI will eat a fifth of tech budgets in 2027, but don’t expect it to pay for itself,” October 21, 2026 — https://www.theregister.com/ai-and-ml/2026/10/21/ai-will-eat-a-fifth-of-tech-budgets-in-2027-but-dont-expect-it-to-pay-for-itself/5304466 [3] Open Future Forum, “CFO AI Leverage Report, October 2026,” Edition 4, published October 2, 2026 — https://openfutureforum.com/research/cfo-ai-leverage-report-october-2026 [4] FinOps Foundation, Tammy Burnitt, “AI Spend Budget Busters: Expanding the Budget Planning Lens,” October 6, 2026 — https://www.finops.org/insights/ai-spend-budget-busters [5] CFO Dive, “Nearly 7 in 10 firms report AI cost overruns,” October 13, 2026 — https://www.cfodive.com/news/7-10-firms-report-ai-cost-overruns/825961