This post used to open with a headline ROI figure from our AI programme. I have removed it. When I went back to reconstruct how it was calculated, I could not do it to a standard I would defend in front of a CFO, so it should not have been published as a fact. What follows is the version I can stand behind: how the costs actually break down, and how to build a business case without inventing a number.
Most enterprise AI business cases are built backwards. Someone picks an outcome, works out what it would be worth, and presents the result as a forecast. It survives exactly until finance asks how the baseline was measured.
So here is the honest structure of what enterprise AI costs, what it returns, and which parts of that you can actually prove.
The costs nobody puts in the slide
Licences are the visible line, and usually the smallest one.
- Seats and usage. Predictable, and the easiest to control. Do not buy seats ahead of adoption; unused seats are the purest form of waste.
- Integration work. The model is a small part of the system. Connecting it to ERP, finance and operations data, with permissions intact, is where the engineering goes.
- The governance layer. Access control, logging, redaction, evaluations. Real work that produces no demo.
- Your people's time. The largest cost in most programmes, and the one that never appears. Every hour a business team spends testing, training and changing how they work is part of the price.
- Running it. Monitoring, incident response, version upgrades, re-running evaluations when a model changes underneath you.
- Compliance. Processor contracts, data mapping, policy work, audit evidence.
Infrastructure deserves its own warning. On-premise inference is attractive right up to the point you price the hardware against actual usage. We looked hard at a dedicated AI server and did not buy it: the proposal assumed a workload we did not have. Checking demand against the real numbers before the spend was worth more than the purchase would have been.
What returns, and what you can prove
Benefits fall into three tiers, and they are not equally defensible.
Tier one: measurable, with a baseline. Cycle time on a documented process. Error rates in a system that logs them. Time from request to answer. These are provable if — and only if — you measured before you started. This is the whole game, and it is the step most programmes skip.
Tier two: real but attributable only in part. Faster decisions because the data is visible. Fewer escalations. You can show the change; you cannot honestly claim all of it was the technology.
Tier three: genuinely unmeasurable. Better judgement, less frustration, people spending their day on work that matters. It is real. Do not put a number on it.
Building the case without a fake number
- Baseline first, build second. If you cannot state today's number, you have no case later — only a story.
- Agree the counterfactual up front. What would have happened anyway? Volumes grow, teams change. Say so before the result arrives.
- Report a range, not a point. A point estimate to a decimal place is a claim about precision you do not have.
- Separate cost avoided from cash saved. Boards know the difference. Blur it once and everything you present afterwards is discounted.
- Publish corrections. When a number turns out to be wrong, say so in the same place you said it originally. It costs a little credibility once and buys a lot permanently.
Could I hand this figure to an auditor with the workings attached? If not, it does not go in the board pack — and it certainly does not go on my website.
What I would tell a CIO starting this year
Instrument before you automate. Pick processes that already produce data, because those are the only ones where you will be able to show what changed. Keep the governance work in the budget from the start; it is not a phase two.
And be careful with your own numbers. The temptation to lead with a big percentage is strong, especially when the work is genuinely good. But a claim you cannot reconstruct is a liability with a delay on it. The work stands up better without it.