What an AI project costs, and where the money goes

Everyone asks the price before the scope. That is fair. You have a budget cycle to work with and somebody above you wants a number before the next meeting.

I cannot give you one from here. The figure depends almost entirely on the state of your data, and nobody can judge that from outside your organisation. What I can give you is the shape of the spend, which is more consistent than people expect.

The first two weeks are cheap and they settle everything else

We call it framing: a short piece of work at the start where we find out what you actually need, before anybody writes a line of software. Two weeks of it costs a fraction of a build. It also fixes most of the eventual price, because that is when you find out whether the data exists, who owns it, and whether anyone is allowed to hand it over.

Skip it and you pay later. One programme I was asked to look at had been running nine months. Four of those were spent waiting for permission to read a single table of records. The model was fine. The model is almost always fine.

Tidying the data is the biggest line and the hardest to show a board

Cleaning records, joining them up, removing the duplicates, working out which of the two customer lists is the real one, and writing down the rule that somebody in accounts has been applying from memory for eleven years. On most jobs this is the biggest single cost.

Finance teams dislike this line because it produces nothing you can demonstrate in a meeting. It also decides whether the rest of the money achieves anything.

The build is smaller than people assume

Building the thing is often the shortest phase. The AI models themselves are close to interchangeable now and you buy them off the shelf. What gets built on top of them is well trodden too: a system that searches your own documents and answers from them, a model trained on your own examples, older and simpler statistical methods, software that can take actions in your systems rather than only answer questions. None of it takes as long as working out what to build did.

If a proposal has the build as its biggest line and barely mentions the data, that tells you what the supplier intends to deliver.

Every answer costs money and nobody forecasts it

Every time the system answers a question, it costs you money. Multiply the number of questions you expect by the cost of answering one. Then multiply again by the number of times a user asks a second time because the first answer was not quite right. Then add whatever extra traffic appears once other teams connect their systems to it, which they will.

I have watched an organisation approve a build without once asking what running it would cost per month. The answer arrived in the second quarter and it was larger than the maintenance budget for everything else they owned.

Then there is owning it

Somebody has to check that the answers are still right, retrain the system when they are not, and be named as responsible for it. Be honest internally about that last one. A system with no owner gets quietly worse while everybody carries on trusting the output.

What actually moves the number

• How much of your data the software can already reach on its own, through a login that belongs to a system rather than a person, instead of somebody fetching it by hand
• Whether any human judgement in the current process has ever been written down
• How many of your existing systems it has to connect to, and who looks after them
• Whether anybody in your organisation will say no to extra requests after the first demonstration

None of those are questions about AI. That surprises people who arrived wanting to talk about models, and it is the single most useful thing I can tell you before you commit a budget.

If you want a real number for your case, those two weeks of framing are the fastest route to one. If you want a rough sense before committing to that, send a description of the decision you are trying to improve and I will tell you what it usually takes to find out.