Issue 015: The Wrong Number

Last week this column argued that AI spend should be governed like any metered resource: owned, budgeted, measured. This week is about what you measure, because most return-on-investment (ROI) conversations about AI are built on the wrong question. They ask: what did the AI return? The right question is: what did the team return because of the AI?

The difference sounds semantic. It is not. The first question treats the model's output as the asset. The second treats the team's growing capability as the asset, and the output as a by-product. This week's news shows both framings in the wild. Aviva is measuring outcomes: £230m (about $305m) of fraud stopped, a number a board can audit. The Copilot revolt shows the opposite: teams that priced AI by output per pound watched metered billing break their business case overnight, because the case was built on a number the vendor controls.

First, outputs are infinite and capability is not. A model will generate drafts, code and analysis at near-zero marginal cost, and anything infinite falls in price. The scarce asset is a team that knows which outputs matter, which to discard, and what to ask next. Scarcity is where value lives; measure the scarce thing.

Second, the boards that win in 2026 measure cost per outcome, not adoption. Adoption tells you people opened the tool. Cost per outcome tells you whether the fraud was stopped, the claim settled, the customer kept, and what it cost to do it. Aviva reports the second kind of number, which is why nobody asks about its first.

Third, capability survives the model swap and output does not. Models will be replaced, repriced and retired; this week's IPO filing makes repricing a quarterly certainty. A team that has learned to decompose work, brief a model and verify results carries that skill to whatever tool comes next. The output you bought last quarter is already gone.

The honest counter-argument: in some functions, raw output ROI is exactly the right metric. Customer support deflection and document classification are bounded, repetitive, and the output genuinely is the product; if the model deflects the ticket or sorts the file at acceptable quality, cost per unit is the whole story, and Aviva's claims-screening sits partly in this category. Concede it, and ring-fence it. Those functions are real, but they are the minority of most AI budgets, and they are precisely the workloads where falling model prices will keep doing the work for you.

So the recommendation: take your three largest recurring AI expenses and write two things next to each. The outcome it is supposed to move, in one sentence with a number in it. And what your team can now do that it could not do a year ago, because of it. If the second column is empty, you are renting output, and output is getting cheaper every quarter, so you are overpaying for it. The Prompt Pocket below turns those two columns into a fifteen-minute audit.

Runway is a financial planning and analysis (FP&A) platform that lets non-finance leaders build budgets, scenario forecasts and runway models without a spreadsheet apprenticeship. To be clear, this is Runway the finance platform, not the AI video-generation company of the same name. It earns its place in this issue because it answers the editorial's question directly: once you decide to measure AI cost per outcome, you need somewhere to model the spend, and Runway makes that a drag-and-drop scenario rather than a fragile spreadsheet. Models read in plain English rather than cell references, every seat is free so the whole leadership team can work in the live model, and an AI assistant analyses variance and explains what changed in your numbers. It connects to more than 750 sources, including Xero, QuickBooks, NetSuite, Stripe and payroll systems.

The weaknesses are commercial rather than technical. There is no free tier, no self-serve signup and no published price list; you book a demo and get a quote, which puts it beyond casual trial for a small team, and implementation takes weeks rather than minutes. Reviewers also note a real, if manageable, learning curve. This is a tool for a funded startup or mid-sized business ready to commit, not a Saturday experiment.

Runway does not publish prices. Plans are quoted to your setup, tiered by the number and type of integrations, with unlimited seats on every plan. Independent buyer guides report entry pricing from about ~£375/~$500 a month, with mid-market deployments commonly £22,000 to £60,000 (about $30,000 to $80,000) a year. Treat those as reported figures, not vendor ones, and get a quote for your own stack. Verify before committing; software-as-a-service (SaaS) pricing moves more often than newsletter cycles.

See Runway in action →

Last week I predicted that "AI adoption" would quietly leave serious board packs. This week I am putting dates on it. First: by June 2027, "AI cost per outcome" replaces "AI adoption rate" as the lead AI metric in serious board packs. Aviva just showed everyone what the outcome version looks like, and metered billing is forcing the cost half onto every invoice. Adoption was a metric for the era when usage was the goal. Usage now costs money that scales.

Second: the first wave of AI budget consolidation lands by the fourth quarter of 2026, and it lands hardest in the enterprises that scaled fastest. Metered pricing plus a public-market repricing of model costs will push every chief financial officer to ask the question this issue hands you the tools to answer first. Both predictions carry dates. Check me against them.

If you ran the first-line-item audit in this issue, hit reply and tell me which one you metered first. I read every response.