What an AI analytics project costs, and how to budget for one

AI project budgets tend to be wrong in the same direction: undercosted at the start and expanded mid-delivery. Getting a realistic number before you begin is not pessimism, it is the thing that decides whether the project survives long enough to deliver anything.

It is the question every managing director has and almost no vendor answers straight. What does an AI project cost? The honest answer starts with a fact that surprises people: the part everyone pictures as expensive, the clever model, is now the cheap part. The money goes somewhere else entirely, and budgeting well means knowing where.

The model is the cheap part.

The algorithms that do the learning, and the large language models behind the assistants, are largely a commodity now. Microsoft and others supply them, improve them, and price them low or build them into tools you already pay for. You are not funding the invention of intelligence. You are funding everything around it that turns it into something useful on your data, and that is where the real cost sits.

Where the money goes.

Four places, in roughly descending order. Getting the data ready, which is the largest and most underestimated, because a model needs clean, joined-up, defined history and most businesses do not have it lying around in that state. Building and validating the model itself, which is real work but smaller than the data preparation that precedes it. Embedding it into a decision, the report, the alert, the process change that means someone acts on the output rather than admiring it. And running it afterwards, which almost every budget forgets. If a proposal is heavy on the model and light on the data and the embedding, it has the proportions backwards.

Budget for a proof, not a programme.

The expensive way to do this is to commit to a large programme before you know whether the first use case even works. The sensible way is to fund a small, scoped proof on one question and one slice of data, see whether it pays, and only then decide how far to take it. A fixed price for an unscoped AI project is a guess with your money on it. Scope the decision first, prove it small, and let the proof earn the programme rather than the brochure promising it.

The running cost nobody quotes.

A model is not a thing you build once and own forever. The world it learned from keeps moving. Customer behaviour shifts, products change, and a model trained on last year quietly drifts out of date. So there is an ongoing cost: monitoring that it is still accurate, retraining it as the data moves, and maintaining the pipelines that feed it. It is modest if planned for and painful if discovered later. A proposal silent on the running cost is hiding part of the bill, in exactly the way an analytics estate left untended quietly rots.

A sensible shape for a first budget.

As a rough guide for a first project, expect the bulk of the effort to go on getting data ready and embedding the result, a smaller share on the model, and a standing line, small but real, for running it afterwards. Keep the first commitment small enough that a disappointing result is a lesson rather than a loss. The aim of a first budget is not to deliver the whole vision. It is to find out cheaply whether the vision is real.

So this week, before you ask anyone for a price, write down the single decision the project is meant to improve and what better would be worth. A price quoted without that is a number pulled from the air. Scoping the decision, proving it on a small slice, and being honest about what it costs to run is how every Analytics Acceleration Programme and machine learning engagement we run begins, and it is the same discipline we set out for buying any analytics work. The model was never the expensive part. Knowing that is most of budgeting well.