Building the business case for AI on one page of numbers

5 days ago
3 min read

Updated: 4 days ago

Introduction

Most internal arguments about adopting AI are arguments about the technology, which is why they are never resolved. One side describes capability and the other describes risk, neither is wrong, and no decision follows. A business case ends this by changing the subject to a specific process, its current cost, and what would be different.

The discipline required is unglamorous. Name the process, measure what it costs now, state what you expect to change and by how much, count the full cost of getting there, and write down what would make you stop. That is one page, it takes an afternoon, and it is the difference between a decision and an enthusiasm.

1. Building the business case for AI starts with a named process

Not with a capability.

"Reduce the time to produce a quote from three days to one" is a case. "Use AI in the sales process" is not, because it cannot be costed, measured or falsified. Pick the process where the pain is already being complained about, because the baseline evidence is easiest to gather and the change is easiest to notice.

2. Measure the current cost properly

The number everything rests on.

Hours per week, at a realistic loaded cost, plus the consequences: work lost to slow response, errors, rework, overtime. Most cases fail because the baseline was guessed and could not be defended. Spend a week actually measuring rather than an hour estimating; the measurement is what makes the rest of the page credible.

3. State the expected change as a range

Honesty improves credibility.

A likely case, a conservative case and what has to be true for each. A single confident figure invites scepticism, and a range with stated assumptions invites a conversation about the assumptions, which is the conversation you want.

4. Count the whole cost

Licences are the smallest part.

Setup, integration, data preparation, training, the time to review output, ongoing administration, and the productivity dip during transition. Cases built on subscription cost alone are routinely wrong by a factor of several.

5. Separate saved time from saved money

The distinction that decides credibility.

Freed hours are only a saving if they are redeployed or if headcount changes. If the honest answer is that people will simply be less pressed, say so, and make the case on quality or capacity instead. Cases that overstate this are the ones that fail their review at renewal, because the saving never appeared anywhere.

6. Include the cost of doing nothing

Frequently the stronger half.

Slow quoting losing work, errors costing customers, an owner spending twelve hours a week on administration instead of selling. The status quo has a cost and it is rarely stated.

7. Propose a pilot with a decision point

The right shape.

A limited scope, a fixed period, a stated measure and a date on which you decide to continue or stop. This converts a large uncertain commitment into a small certain one and is far easier to approve.

8. Write down what would make you stop

The clause nobody includes.

A threshold and a date, agreed in advance. Without it, the review at renewal is decided by whoever is most invested, and sunk cost wins reliably.

9. Name an owner

Cases without one do not happen.

Somebody accountable for the implementation, the measurement and the recommendation at the decision point. This is the single strongest predictor of whether anything actually changes.

Be careful about cases built on vendor-supplied figures. Savings quoted from other customers were achieved in other circumstances, and the only number that matters is the one from your own baseline.

Conclusion

Name a process, cost it, and put the case on one page.

Measure the current cost including its consequences rather than guessing, express the expected improvement as a range with stated assumptions, count setup, training, review and transition costs alongside the licence, be explicit about whether saved time becomes money, include the cost of doing nothing, propose a pilot with a fixed period and a decision date, write down in advance the result that would end it, and give the whole thing a named owner.