Why AI budgets overrun, and how leading companies prove what they’re getting.
What published industry research says about AI spend, and what we think it means: where the money goes, why it’s hard to control, and what the companies doing it well measure instead.
For finance and IT finance leaders
·
5 min read
·
Published Nov 3, 2026 · XTrace research team
Research brief
From token to task to outcome to value.
In short
Most AI spend is under-managed.
Nearly every company surveyed overspent its AI budget, and a fifth or more of spend typically goes untracked.
Cost alone can’t show value.
The leaders follow spend through to the work done and the business KPI it moved.
Waste recurs unless it’s watched.
Silent failures cost real money and come back after one-off clean-ups.
Untracked spend
AI spend is outgrowing budgets, and much of it is untracked.
4×
Nearly fourfold growth
How much AI spend grows as companies move from isolated pilots to company-wide use.
93 %
Spent over budget
Of surveyed organizations say they spent more on AI than they budgeted.
20–30 %
Not accounted for
Of AI spend, scattered across vendors, tools and pricing models.
20–25 %
Have mature practices
Of companies manage AI spend with mature practices today.
SOURCE McKinsey, “The cost of intelligence: How CIOs can manage AI demand at scale”, July 2026. Survey data (Exhibits 1–2) and McKinsey’s client experience; all four figures.
Why it’s harder than cloud cost management
The same task can consume very different amounts: up to 30 times as much, per McKinsey.
Agents chain many model calls together, so one request can become dozens.
Usage-based pricing removes the fixed ceiling that per-seat licenses used to provide.
Without guidance, people default to the most expensive model.
Tagging, sourcing strategy and an operating playbook are still immature at most companies.
SOURCE Paraphrased from McKinsey, same article.
What this means for you
Before you can say whether AI is paying off, you need one reconciled view of what it costs, by team and workflow. Most companies don’t have that yet, so getting there is the first win.
Cost to value
Measuring cost is the easy part. Proving value takes three more steps.
Most reporting stops at usage, and finance can’t book usage. The companies furthest ahead trace each step: what was consumed, what work it did, what outcome it moved, and what that outcome is worth.
Consumption
Spend by model, team and tool.
Where most reporting stops
Work done
The tasks and tickets that spend served.
Outcome
The workflow KPI the work moved.
Value
The result in terms finance can defend.
What finance can book
How we see it
Visibility isn’t value.
A dashboard of spend is a starting point. What a CFO needs is evidence that the spend produced something worth paying for.
Saved minutes can disappear.
A task AI finishes faster only helps if the rest of the workflow keeps up. If the work still queues for review and release, the time saved never reaches the business.
One data source isn’t enough.
Billing shows cost, AI tools show usage, and business systems show outcomes. Value only becomes visible when all three are joined up.
Measure by business result.
Over time, judge AI by measures such as cost per claim processed or revenue per AI-enabled workflow.
SOURCE McKinsey, “The cost of intelligence”, July 2026.
What this means for you
Agree up front which business KPIs define value for each AI workflow, then measure cost against those, not against token counts.
Hidden waste
Silent failures are a recurring cost, not a one-time fix.
When a tool an AI agent depends on breaks, the agent usually retries and finds a workaround. The task completes, so nothing looks wrong, but the extra attempts are billed. One published audit shows the scale.
7 bugs
Behind the waste
Small tool bugs in one company’s internal agents.
~$499K /yr
Wasted AI spend
Plus roughly 12,000 engineer-hours a year spent waiting.
~$1.2M /yr
Lost productivity in total
Found and fixed in about an hour.
SOURCE Databricks, “How we eliminated $1 million a year of wasted AI agent spend in one hour”, Sept 1, 2026. Databricks’ own annual estimates, extrapolated from one 24-hour window of traces.
The audit’s authors recommend repeating it regularly, because tools and models keep changing and new failures appear. McKinsey’s guidance points the same way: it lists unnecessary agent loops and repeated failed requests among the hidden costs to manage.
What this means for you
Treat repeat errors as a standing line item. Monitor for them continuously, and make sure a fix in one team reaches every team.
Shaping demand
Shape demand. Don’t ration it.
Cutting AI spend across the board slows the work it was meant to speed up. The better approach is to make the cost-effective choice the default one.
Know what you spend.
One source of truth for spend, usage and outcomes across every vendor and tool.
Optimize the systems, not just the model.
Route work to the cheapest model that meets the quality bar, trim bloated inputs, and limit retries and loops.
Assign accountability.
Allocate AI cost to the business units, products and workflows that create it.
Build guardrails into the tools.
Approved models and tools per team, applied automatically, with guidance while people work rather than in separate training.
SOURCE Framework paraphrased from McKinsey, “The cost of intelligence: How CIOs can manage AI demand at scale”, July 2026 (its four actions for managing AI demand).
One caution: hard per-person usage caps tend to backfire. People wait for the next day’s allowance instead of finishing the work.
What this means for you
Prefer routing and guardrails over blunt limits, and make AI cost management a permanent cross-functional capability, not a one-off project.
Where the industry is heading
Expect standard measures for AI cost and value.
In 2026 the Linux Foundation launched the Tokenomics Foundation to define open standards for the cost and return of AI. Early priorities include measuring cost per request rather than per token, and measuring value starting with the share of work AI completes without a person stepping in. Its launch cites a forecast that token consumption will grow 24-fold by 2030.
SOURCE Linux Foundation, launch of the Tokenomics Foundation, Aug 2026. The 24x forecast is attributed there to Goldman Sachs.
What this means for you
Standard measures make AI cost and value comparable across teams and vendors. Choose tooling that already tracks cost per piece of work and how much work AI completes on its own.
Sources
For anyone checking our work.
McKinsey Quarterly, July 2026
“The cost of intelligence: How CIOs can manage AI demand at scale”
Databricks, Sept 1, 2026
“How we eliminated $1 million a year of wasted AI agent spend in one hour”
Linux Foundation, Aug 2026
Launch of the Tokenomics Foundation
See what this looks like for your AI spend.
XTrace ties AI spend to the work it does and the value it delivers. Your team approves budgets, routing and caching in one place, and the XTrace plugin stops known repeat errors inside AI tools.