How to get budget approved for an AI project

The project is sensible, you can explain it, and it has been sitting with finance for five weeks.

The usual diagnosis is that the business does not understand AI. That is rarely it. The request is being assessed the same way every other request is assessed, and against that standard it is missing the things an approval needs.

Why the AI framing works against you#

A proposal that opens with the technology invites a debate about the technology. Now the approval depends on whether a committee believes in large language models, which is not a question any of them can settle and not one your project needs them to.

A proposal that opens with the job invites a debate about the job, which is a thing the company already has opinions about and can decide.

The same project, described as "AI-powered claims triage" and as "the thing where Priya and two others spend a day and a half a week rekeying claims between two systems", gets different treatment. The second one is a cost everybody already agrees is real.

The three numbers#

What it costs today. Hours a week, across everyone involved, times a loaded cost. Not a salary, a loaded cost, because finance will convert it anyway and doing it yourself shows you know the difference.

If nobody has measured it, measure it. Two weeks of somebody noting what they did is enough and it is the single highest-return thing you can do before asking for anything.

What it costs afterwards. The honest version, including the ongoing cost of running it. An AI system has a bill that arrives every month, and a proposal that omits it reads as either naive or evasive. A model's inference cost at your volume is knowable; our own calculator will get you to an order of magnitude.

What happens if you are half right. This is the number that gets approvals over the line and almost nobody includes it.

If the saving is forty hours a week and you are half right, it is twenty, and the project still pays back inside the year. Saying that yourself does two things. It shows the case does not depend on the optimistic figure, and it takes the obvious objection off the table before anybody raises it.

Make it small enough to say yes to#

A fifteen-thousand-pound approval and a hundred-and-twenty-thousand-pound approval are different processes in most companies, often with different people in the room and a different number of weeks.

If the full build is the second, the first request should not be. Ask for the smallest piece that produces a real number, get it, and come back with evidence rather than a forecast. The second request is assessed by people who have seen you deliver the first.

This is the whole argument for buying a small fixed piece of work first. Ours is seven days and $2,500, which is below the approval threshold at most companies that would ever hire us, and that is deliberate.

What finance will ask, and the answers that work#

"What if it does not work?" Name the point at which you would stop, and what it costs to have stopped there. A project with a named kill point is a smaller risk than one without, and saying so converts an objection into a reassurance.

"Why now?" The strongest answer is a cost that is growing. The weakest is that competitors are doing it, because that invites the response that competitors are wasting money, which they frequently are.

"Who owns it afterwards?" Have a name. A system nobody owns becomes a system nobody maintains, and finance has funded enough of those to recognise the shape.

"Can we do it in-house?" Sometimes yes, and you should have done the arithmetic rather than being surprised by the question. What an in-house AI team really costs is the comparison, and it is not the salary.

The short version#

Lead with the job, not the technology. Carry three numbers, one of which is the pessimistic case. Ask for a piece small enough to approve quickly, and come back with evidence.

If the blocker is that nobody has measured the job, that is the first piece of work rather than a reason to wait. Book the call and we will tell you whether there is a case here, including when there is not.