A one-page memo for Investment Committees, boards, and anyone about to approve an “AI investment.” The figures in the example below are made up to protect commercial terms, but they are in the range of the real ones, and the logic is exactly what we ran.
I’ve approved AI initiatives, built them, and priced them. The most useful thing I ever did with one happened before a single line of code was written.
It was a spreadsheet.
But first, the uncomfortable context.
For two years, “we’re investing in AI” was enough. Not anymore. The 2026 numbers are blunt:
Estimating the ROI of an AI investment is not something you do after the pilot. If you postpone it, you’re not postponing a calculation. You’re postponing the decision of whether this is a business or a hobby.
Melissa Perri warned us about the build trap years ago: teams measuring progress by features shipped instead of value created.
We’re falling into it again, with better branding.
AI initiatives today get filed as a cost center: a line item for tokens, infra, and a couple of engineers, justified by vague productivity. That framing is the trap. A cost center’s only future is to be cut. A value lever has a P&L.
McKinsey’s data makes the trap visible. Across their sample, moving from opportunistic AI use to embedding it in the operating model barely moved the revenue multiple. The jump only shows up once AI changes what the company sells, not just how it runs.
Exhibit A
Median revenue multiple by AI maturity level · 471 PE-backed companies, 2023–25 · McKinsey
Productivity is where the multiple stalls. The product bet is where it compounds.
The difference isn’t the technology. It’s whether anyone did the math.
A while back, as CPTO of a B2B SaaS company, we wanted to ship an AI feature that automated a painful document workflow for our clients. Before committing the team, we built the unit economics from our POC’s hard data, not from a vendor’s slide.
Exhibit B · illustrative figures, proportional to the real case
| Line item | Per client / month |
|---|---|
| Token costmeasured during the POC, real usage volumes | €5 |
| Developmentto reach production quality, one-time | €40,000 |
| Maintenanceevals, prompt regressions, model migrations, monitoring. The number everyone underestimated | €2,000 |
| Price chargedthe add-on we’d bill clients | €25 |
| Contribution per client | €20 / mo |
| Break-even: ongoing maintenance | ~100 clients |
| Break-even: development, within a year | ~270 clients |
Against our active base and realistic attach rates, that was ambitious but achievable, so that adoption number became our kill-criterion before a single prompt was written.
The token cost, the thing everyone fears, was a fifth of the real cost. The model wasn’t the expensive part. Keeping it reliable was.
That spreadsheet changed the conversation before anyone touched the product. Pricing stopped being an afterthought. Adoption became the metric, not the demo. And when someone asked “is this worth building?”, there was a number to answer with instead of an anecdote.
First, be clear about what kind of bet it is. They’re all legitimate, they have different P&Ls, and the discipline below applies to each of them.
Product bet
A capability your clients pay for, directly.
Internal productivity
Capacity you free up and redeploy to a named goal.
Enabling foundation
Data, platform, and governance work that makes the next ten bets cheaper.
If I were sitting on your Investment Committee, four things before approving:
Exhibit C · IC approval checklist
None of this requires a data science team. Ours was a spreadsheet and an afternoon of honesty.
So, next time an AI initiative asks for budget, whether it’s a feature for clients or a tool for your own teams, ask one question: when does this pay for itself, and how would we know?
If nobody in the room knows, you’re not looking at an investment yet.
David Jordan · Author & Product Leader · growth-engine/artifacts/memo-ic-defensible-ai-roi