The token is the wrong unit for AI ROI

Benefit doesn't arrive per token, so token cost against return isn't the ROI question. Why the initiative is the unit, and where token cost belongs.

McKinsey thinks the conversation about AI is about to turn to ROI. We agree on the timing, but tokens are not the unit.

What did McKinsey predict about AI ROI?

That within three to six months, companies that have moved from experimenting with AI to scaling it will start asking what it returns. In a September 2026 interview with WIRED, Dan Swan, who leads McKinsey's Tech & AI group, framed the question as what a token costs versus "the benefit you get from investing in that token."

The cost side of that question is easy to answer. The benefit side, measured per token, can't be answered at all.

Why is the token the wrong unit for AI ROI?

Because benefit doesn't arrive per token. A token is an input price, like a kilowatt-hour. The return comes from a piece of work changing: a support ticket closed without a person, a contract reviewed in one hour instead of four, a deal that closed sooner.

The same tokens can buy a great deal or nothing. A retry loop, a draft nobody used, and a prompt that is mostly pasted context all consume tokens, and none of them changes the business. Divide benefit by tokens and you get a ratio that moves for reasons that have nothing to do with the return:

  • Switch to a cheaper model. Token cost halves, so the per-token ratio doubles. If the work comes out the same, nothing in the business changed.
  • Switch to a more capable model that resolves more tickets. The per-token ratio falls, while the return went up.
  • Your provider cuts its prices. The ratio improves, and nobody on your side did anything.

A unit cost is what one unit of an input costs. A return is measured on the work that input changes, which is why the two need different units.

What is the right unit for AI ROI?

The initiative: one piece of work the AI is meant to change, with an owner, a budget, a metric, and a baseline for that metric from before the AI. Every cost lands on it, and its value is the change in the metric against the baseline.

Tokens are one line of that cost, and often not the largest. Seats, cloud services, integration work and the people running the initiative all count.

Question Per-token view Per-initiative view
What counts as cost? Model usage Model usage, seats, cloud, integration and people's time
What counts as benefit? Nothing it can measure directly The change in a business metric against its baseline
What happens when you switch to a cheaper model? The ratio improves Cost falls; value moves only if the work does
Who can say the benefit isn't real? Nobody — there is no benefit figure to challenge Someone in the business who approves the value
What decision does it support? Which model to send work to Whether to scale, fix or cut the initiative

Doesn't knowing your token spend get you most of the way?

No. Seeing what AI costs and showing what it returned are different problems, and solving the first doesn't solve the second.

The Tokenomics Foundation, a Linux Foundation project, asked 472 organizations about both. Among those most confident they could see their AI spend, only 25% were also confident they could connect it to a business outcome their CFO would accept. Across all respondents, 39% were not confident they could do that at all.

Those figures come from State of Tokenomics (September 2026, CC BY 4.0), published as a release candidate. Respondents were self-selected rather than randomly sampled, and the publisher doesn't say how they were recruited, so read it as where large AI adopters say they stand, not as a probability sample.

Where does token cost still matter?

Inside an initiative's cost, and as an operating metric for the work. Routing simple requests to a smaller model, caching prompts and trimming context are real savings, and an initiative whose whole purpose is cutting AI spend should be measured on exactly that.

Even then, the useful figure is the cost per unit of work, not per token: what a resolved ticket or a reviewed contract costs now against what it cost before. That puts Swan's question in a unit a CFO can check. And when that cost falls, it is worth separating what your team earned by changing how it runs the work from what your provider handed you by cutting its prices. The second is real money, but nobody should be credited for it.

How do you weigh cost against benefit for one initiative?

Write down what the initiative is meant to change and how you will measure it, then compare its full cost with the approved change in that measure. In order:

  1. Name the work and the metric. Cost per resolved ticket, hours per contract review, days to close.
  2. Record the baseline for that metric from before the AI, with the months it covers and how it was measured. Label it as an estimate if it is one.
  3. Attach every cost to the initiative, from token bills to seats to the people running it, split where a bill serves more than one initiative.
  4. Measure the change against the baseline under the same scope, and have someone in the business approve it before it counts.
  5. Compare cash value with cost, and report hours freed beside it rather than inside it. Freed time is real capacity, but it isn't cash until a budget line moves.

If that sounds like more work than dividing by tokens, it is. It is also the version that survives a CFO asking how the number was made.

Where Roiva fits

Roiva measures AI ROI per initiative. Once you link a connection to an initiative as a cost source, the part of its bill you assign to the initiative is booked as the initiative's cost. Value is measured against the initiative's baseline: calculated by a versioned formula, synced from a connected platform, or entered by hand. It counts only once it is approved: by a person, or by an auto-approve setting the account can turn on, which approves a formula's entries only after a person has approved one of them. Where Roiva can read a provider's cost by model and token type, a saving in AI spend shows how much of it came from the provider's prices and how much from what your team ran.

The method is published at /methodology, and how costs reach an initiative is documented at /docs.

See it against your own numbers

The free AI ROI Exposure Assessment runs Roiva's published math on your inputs — five minutes, no account, no card.

Get the next one

New field notes by email as they're published. Or subscribe to the RSS feed.