Home / Insights

What enterprise AI really costs: the line items nobody budgets

Ask what an enterprise AI project costs and the answer is usually the licence or the model, because that is the number the vendor quotes. That line is the tip of the budget. The costs that overrun a programme sit underneath it, and most business cases leave them out.

The short answer

  • The model or licence is one line of many, and often not the largest.
  • The costs that get missed are data preparation, integration, change management, monitoring, security and compliance, and the running cost of inference.
  • A large share of the total is recurring, so AI behaves like an operating commitment rather than a one-off capital purchase.
  • Escalating cost is one of the named reasons AI projects get cancelled, precisely because the full bill was never estimated at approval.
  • A business case that lists only the visible costs will be approved on the wrong number and judged on the real one.

Why the visible cost is the misleading one

The licence and the model are quoted because they are easy to quote. They are also the part of the system a vendor is selling, so they are the part that gets attention. The rest of the cost is work your own organisation has to do to make the purchase useful, and because it is spread across data, engineering, operations and risk functions, no single quote captures it. Gartner has repeatedly named escalating and unanticipated cost among the reasons generative and agentic AI projects are abandoned, which is what happens when the full bill arrives after the approval that assumed the small one.

The point is not that AI is unaffordable. It is that the affordable number on the slide is not the number the programme will actually spend, and a board that approves the first is committing to the second without seeing it.

The line items that get missed

Line itemWhat it coversType
Data preparationSourcing, cleaning, labelling and joining the data the model needs, and fixing the quality problems the project exposes.Mostly upfront, some ongoing
IntegrationConnecting the model to the systems and workflows it has to sit inside, in production rather than in a demo.Upfront, with maintenance
Change managementGetting people to actually use it: training, process redesign, and the adoption work that decides whether value appears.Upfront, then sustained
Inference and computeThe cost of running the model on real volume, which scales with usage and can dwarf the licence at scale.Recurring
Monitoring and maintenanceWatching for drift, retraining as data changes, and the engineering to keep it working.Recurring
Security and complianceControls, impact assessments, audit evidence and the governance the system needs to be run responsibly.Upfront, then ongoing
Human oversightThe people who review, correct and stand behind the model's output where a decision matters.Recurring

Read the type column. Most of these are recurring, which is the structural point about AI cost: it behaves like an operating commitment rather than a capital purchase. A project costed as a one-off build, with a small ongoing line for the licence, has mistaken the shape of the spend, and the running costs will keep arriving long after the launch that was supposed to end them.

The cost that scales with success

Inference cost deserves singling out because it moves the wrong way for comfort. A pilot serving a handful of users is cheap to run. The same system serving the whole organisation, on real volume, can cost more to operate than it cost to build, and that cost rises with adoption. A business case that assumes usage will grow, as the value case requires, but budgets inference at pilot volume, has a hole in it that widens exactly as the project succeeds. The more the tool is used, the larger the surprise.

The cost of running it responsibly

Security, compliance and human oversight are treated as optional in the budget and mandatory in reality. A system touching personal data needs the lawful-basis and impact-assessment work set out in the rules on using your own data. A system informing a material decision needs someone accountable to review it. These are not add-ons to be cut when the budget tightens. They are the difference between a system you can defend and one you cannot, and leaving them off the estimate does not remove them from the bill; it defers them to the moment a regulator or an incident forces them.

The constraint

The binding limit on getting the cost right is an honest, fully loaded estimate of the recurring lines, rather than a sharper price on the licence. Negotiating the model cost down while omitting inference at scale, maintenance and oversight optimises the smallest number and leaves the largest ones unbudgeted, because the deciding question is what the system costs to run for years, and only the total answers whether the value case survives.

How to build the number properly

Cost the system across its whole life, not to launch. List every line above, mark each as upfront or recurring, and put a multi-year running total next to the build cost. Where a line is genuinely uncertain, say so and bound it, rather than leaving it out because it is hard to price, since the pattern behind why pilots fail is a value case tested against a cost that was never complete. A board given the fully loaded number can make a real decision about whether the value justifies it. A board given the licence price is being asked to approve a fraction and own the whole.

Sources

  1. Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, June 2025, naming escalating cost among the drivers.
  2. RAND Corporation, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, 2024, on cost against delivered value.