On the TBM Council's numbers, tokens are 4.5–15% of AI cost. What that means for your budget

In brief

The TBM Council puts labor at 50–70% of AI project cost. On its numbers, tokens are 4.5–15% of the total. What the figures show and what to budget for.

Steve Elliott

Steve Elliott · · 5 min read


On this page
  1. How much of an AI initiative is tokens?
  2. How solid is the 50–70% labor figure?
  3. What did the TBM Council announce?
  4. Where do AI costs sit in a TBM model?
  5. What should change in next year's budget?
  6. Where do FinOps and tokenomics fit?
  7. Sources

On the TBM Council's own figures, tokens are between 4.5% and 15% of what a typical AI initiative costs. Its strategy paper puts labor at 50–70% of AI project costs, and says tokens are often 15–30% of the non-labor spend. Multiply those through and a budget built around a token dashboard is watching one of the smaller lines.

Two caveats matter as much as the arithmetic. The paper doesn't yet cite a source for the labor range, and the best public evidence on the split points in both directions. So the useful move is not to adopt the Council's numbers but to measure your own.

How much of an AI initiative is tokens?

The paper says token costs "often represent only 15-30% of the non-labor spend for a typical AI initiative." The press release that promoted it says "non-labor token spend represents only 15% to 30%," which is easy to read as a share of total cost. The paper is clear that it is a share of non-labor spend.

The difference is large. Take a $1 million AI initiative and apply the paper's ranges:

Labor shareLaborNon-laborTokens at 15% of non-laborTokens at 30% of non-labor
50%$500,000$500,000$75,000 (7.5%)$150,000 (15%)
70%$700,000$300,000$45,000 (4.5%)$90,000 (9%)

The paper doesn't break down the rest of the non-labor spend. It names data engineering, software licenses and business process change as costs that a token-only view misses.

How solid is the 50–70% labor figure?

It's an informed estimate, not a benchmark. In the paper, the sentence giving the labor range ends in a "[source]" placeholder where a citation should be, and neither the paper nor the press release describes a sample or method. Until the Council publishes its source, treat the range as a starting assumption.

The public evidence supports both a labor-heavy and a consumption-heavy picture, depending on the work:

  • Labor-heavy. Bain estimates that in software engineering, the domain furthest along with AI, token spending is about 1–2% of headcount cost.
  • Consumption rising fast. Bain also found the cost of tokens fell by half between December 2024 and December 2025, while tokens consumed grew 4.5 times. Nvidia's Bryan Catanzaro told Axios that for his team, "the cost of compute is far beyond the costs of the employees." Uber's CTO reportedly used up the company's entire 2026 budget for AI coding tools in the first four months of the year.

The consumption examples come from unusual organizations: a frontier AI research team and a heavy early adopter of coding agents. They show where costs go as usage scales, not where most enterprises are today.

In our view, three factors decide your split:

  • Pilot or scaled. Pilots are mostly people. As usage grows, consumption grows faster than headcount.
  • Buy or build. Seat-based copilots put cost into software licenses. Custom agents put it into engineering time and tokens.
  • Agent-heavy work. Agents that reason over many steps use far more tokens per task than chat does.

What did the TBM Council announce?

On November 17, 2026, the TBM Council repositioned the TBM Framework as "the business operating system for the AI era." The announcement promoted the first paper in a strategy series by executive director Matt Guarini, Beyond Tokenomics: TBM Delivers the Full Economic Picture for Accurate AI ROI, published September 2, 2026.

The paper argues that the debate on AI economics is stuck on tokens, treating AI as a new flavor of cloud consumption. It says AI economics span labor, infrastructure, data, software, business processes and value realization, and that a view limited to consumption misjudges both cost and return.

The press release lists five strategic moves: persona-led guidance for executives on funding, value capture, scaling and governance; the TBM Taxonomy as a shared decision model; a body of practice covering fully burdened and variable AI costs; TBM as the context layer for agentic workflows; and an expansion into business architecture and value streams that the Council calls Enterprise Business Management.

Where do AI costs sit in a TBM model?

Most already have a home in TBM Taxonomy 5.0. Employee and contractor time sits in the Staffing cost pool, which replaced the separate internal and external labor pools. Model APIs and hosted AI services sit in Cloud Services, now a top-level pool. Version 5.0 also added a Data tower, AI sub-towers for compute, storage, network and models, and Artificial Intelligence as a solution type. The gap is usually not the category. It's that AI work isn't tagged, so AI labor disappears into the run cost of whichever team did it.

What should change in next year's budget?

Three moves hold up whether the labor share turns out to be 40% or 70%.

  1. Tag AI labor time. If labor is the largest line, measure it first. Add an AI initiative code to timesheets or allocation rules, even a rough one, so you can see your real split by your next budget cycle.
  2. Track each AI initiative as its own project or service. Roll its labor, data, license and consumption costs up to one fully burdened cost, rather than reporting only the API invoice.
  3. Report cost per task, not per token. A falling token price means little if each task uses more of them. Cost per resolved ticket, document processed or forecast produced is a number the business can weigh against value.

This is the case for running AI through the same operating model as the rest of the portfolio: the funding decision, the work and the cost need to meet in one place to say whether an initiative paid off.

Where do FinOps and tokenomics fit?

Both still matter. The paper puts it this way: FinOps helps you make the right tactical decisions, while TBM helps you decide whether to run the workload at all. The FinOps Foundation's 2026 framework now covers technology categories well beyond public cloud, including SaaS and AI. What neither consumption view models is internal labor and process change, which on the paper's numbers is most of the cost.

Sources

Frequently asked questions

What percentage of AI project costs is labor?

The TBM Council's *Beyond Tokenomics* paper says labor makes up 50–70% of AI project costs on average. The paper does not yet cite a source or method for that range, so treat it as an estimate and measure your own split.

What share of AI costs are tokens?

The same paper says tokens are often 15–30% of the non-labor spend on a typical AI initiative. If labor is 50–70% of the total, that puts tokens at roughly 4.5–15% of the whole. Your own split depends on scale, on buying versus building, and on how much agent work you run.

Is tokenomics the same as AI cost management?

No. Tokenomics usually refers to the price and consumption of model tokens. AI cost management also covers the people, data work, software licenses, infrastructure and business process change needed to build and run AI. On the TBM Council's figures, those other costs make up most of a typical initiative.

How should IT finance teams track AI costs?

Start by tagging AI work in time and allocation data, since labor is usually the largest cost. Track each AI initiative as its own project or service so labor, data, license and consumption costs roll up to one fully burdened figure, then report cost per business task rather than per token.

Free playbooks

In-depth guides for product and operations leaders.

Browse the playbooks
Steve Elliott

Steve Elliott

Published November 22, 2026

More on TBM & IT Financial Management

Cost pools, towers, showback, unit cost.