What AI implementation actually costs

Why no credible figure exists

"AI implementation" is not one thing with one price. It covers a single automation added to an existing tool and it covers a multi-year, enterprise-wide programme, and the two have nothing in common except the word AI. None of the major research bodies cited on this site, RAND, Gartner, MIT, IBM or McKinsey, publish a cost figure for it, because none of them studied spend. They studied adoption, failure and return, and cost turns up only as a named cause, never as a number.

Gartner names escalating costs among the reasons it forecasts that at least 30% of generative AI projects will be abandoned after proof of concept. That is a cause cited in a forecast, not a documented figure, and it is the closest thing to a cost finding that a credible source publishes.

What the vendor ranges actually are

Type this question into a search engine and almost everything that ranks is published by a development agency, quoting a range for what a custom AI build costs. None of them discloses a sample of projects, a definition of what counts as in scope, or a method for arriving at the number. That is not research. It is content marketing for the same agency's own build services, and the range exists to make whatever they eventually quote you look reasonable next to it.

We are not going to repeat one of those ranges here, not even to argue with it. Putting a figure on the page, even inside a sentence that debunks it, is how vague estimates calcify into facts that people later cite as if somebody had measured them. Describing the pattern is honest. Attaching a number to it, for or against, is not.

The costs that are documented

What is actually published sits one level away from a price tag. IBM surveyed 2,000 CEOs and found that only 25% of AI initiatives had delivered the ROI expected of them, and just 16% had scaled enterprise wide. MIT separately found that 95% of organisations were getting zero return from generative AI. Neither is a cost figure. Both say the same thing from a different angle: cost is only half the equation, and the other half is that the return often does not arrive at all, whatever was spent to get there.

Where money actually goes wrong

The documented failure conditions point at three specific gaps: no cost cap, no baseline metric, and no funded route to production. Where those are missing, spend does not fail gracefully, it simply keeps going. Two practitioner accounts, each one person's experience and not a study, show what that looks like from the inside.

One account, posted to Hacker News, describes an unsupervised agent cron job that spent $24.88 in two days, with no cost guards and no human review built in at any point. Small money, and the account is offered as one anecdote about a missing cap, not a finding about typical spend.

A second account, a comment from a consultant who had worked on several large retrieval chatbot rollouts, describes users telling the business plainly that they saw no value over ChatGPT, while the company was carrying five and six figure cloud bills for the system those users wanted replaced. Again, one person's account of one pattern they had seen more than once, not a survey. It is quoted for what it makes concrete, not for how common it is.

What to do instead of asking for a number

Asking "what will this cost" is the wrong question when nobody credible can answer it. Three things you can actually do instead, each one lifted straight from the checklist this site publishes alongside this page: define the metric and its baseline before any spend is committed, cap the spend at the provider with an alert to a named person rather than a monthly bill review, and price the pilot on its own, not the programme it might one day become.

A funded, capped pilot with a named metric is a comparison you can actually make. A vendor's unsourced range is not.

What we would need to publish a real figure

A sample of real projects, with their budgets disclosed rather than estimated after the fact. A definition of what counts as in scope, so a chatbot wired to an existing helpdesk and a bespoke model trained on proprietary data are not averaged into the same number. And a stated method for how the projects were selected and the figures collected. Nobody has published that for AI implementation, on any side of the market, including us.

Until somebody does, the honest version of this page is the one you are reading: what is documented, what is not, and a refusal to fill the gap with a plausible-sounding guess.

Questions, answered from the sources

Related: build versus buy, where the same evidence points fairly firmly in one direction. Back to the reconciled failure figures.