What did the AI actually buy us?
Happy Q4!
For the last few months I've been jotting down the AI ROI stats I come across in articles, reports, and LinkedIn. They are hilariously contradictory. Here are a few examples:
Back in July, EY asked senior leaders whether their AI investments had paid off, and nearly all of them (98%) said yes. Two months later, KPMG asked a similar group whether they consistently weigh AI value against what it costs, 12% said yes.
The practitioner surveys tell a different story again. In the TBM Council's latest State of TBM, about 14% of TBM practices use TBM to measure the ROI of AI. The FinOps Foundation's practitioners named AI value as the thing they most need to get better at.
Meanwhile, on LinkedIn...
My feed has been full of posts about Uber and Esker blowing through their AI budgets. Uber used up its entire 2026 AI coding budget in four months, and Esker's AI costs ran about 4x over plan. Accenture recently found that one in three companies has already used up its annual token budget, with a full quarter still to go.
The problem at hand...
Because ROI is largely unknown and budgets are being blown, CFOs are now fielding questions from analysts on earnings calls and from their boards.
That puts CIOs and AI leaders on the spot. They need to tell a credible story about where the dollars went, and then advise on which bets to double down on, which to refine, and which to stop.
How Mike and I are thinking about it...
This isn't perfect, but it's the rough direction we've been working through:
Understand the entire AI estate, beyond the foundation model labs. AI spend isn’t just what’s coming from OpenAI, Anthropic, and the like. Copilot, Now Assist and Einstein are showing up in renewals, and data platforms now bill for AI consumption too. Snowflake's CFO credited a "meaningful step-up in AI revenue" last quarter, and their AI revenue is somebody's AI spend.
Attribute tokens and spend to enterprise capabilities. A token gets used by a person, inside an application, supporting a business capability. Trace that chain and the spend starts to mean something.
Align usage to the targeted outcomes. Every AI dollar should be pointed at something: new revenue, increased productivity, reduced errors or risk, or something else you can name.
Sort the portfolio: protect, accept, right-size, or kill. Decide what's clearly working and should be protected before anyone starts cutting. Cost programs have a habit of cutting the things that were working.
Look at policy, training and automation to optimize. Automation gets bought first because it's easy. Training probably moves the number most and gets budgeted least.
Where we’re leaning...
We think organizations with TBM models have a leg up here. They've already done the hard work of allocating labor spend. If you can attribute AI spend to the individual, you can attribute token consumption to the things people are actually working on.
Would love to hear how everyone is tackling this and answering these questions about AI spend.
Gboyega Adebayo
Founder & Managing Director, Falconbridge