What an AI project really costs: TCO beyond the licence

The AI business cases we receive for review share a geometry: the visible line, licences or token consumption plus the supplier's days, occupies the centre of the sheet, and around it there is white space. Then the project starts, and the white space fills up: the dataset to remediate that nobody had priced, the ERP integration estimated at two weeks and lasting four months, the privacy assessment arriving mid-development, the training reduced to a webinar and repaid with non-usage. At final count, in the projects we analyse, the initial line rarely covers more than a third of the total, and the surprise is not in the numbers, it is in the fact that we keep calling it a surprise.

The total cost of ownership of an AI initiative can be broken into five families, and knowing them beforehand turns the business case from an exercise in optimism into a decision instrument.

Prerequisites and integration: the cost of reaching the start line

The first family is the prerequisites, dominated by data. Making sources accessible, measuring their quality, remediating what the use case requires, building the pipelines that keep clean what was cleaned: it is the workstream we have written about repeatedly, and in projects on core processes it can be worth as much as everything else combined. Honest accounting puts it in the bill, with one alleviating note: it is the only line that amortises across subsequent use cases, and a well-made business case spreads it instead of loading it all onto the first project, which otherwise never starts.

The second family is integration, the cost of making the system live inside real workflows rather than next to them: connectors to existing systems, redesign of process steps, handling of the exceptions the demo's happy path never contemplated. McKinsey's State of AI research (2025) identifies workflow redesign as the factor with the largest impact on GenAI returns, which makes this line doubly interesting: it is among the most underestimated in quotes and the most correlated with value.

Compliance, people, operations: the cost of staying up

The third family is regulatory oversight, which for systems in AI Act high-risk territory has nameable items, technical documentation, impact assessments, human oversight design, registrations, and for all the others still includes classification, register updates, possible work on legal bases and DPIAs. Putting it in the bill at the start costs hours of legal and risk time; discovering it at the end costs months of standstill, and it is one of the trade's best-documented asymmetries.

The fourth family is people: training for those who will use the system, the time of domain experts to build golden sets and validate outputs, the change management separating an installed system from an adopted one. It is the line business cases cut first because it is the easiest to postpone, and the one whose absence produces the most mocking failure, the perfect system nobody uses.

The fifth family is operations, the recurring cost starting when the project "ends": quality and drift monitoring, model updates with the retuning that follows, inference costs growing with adoption (the paradox of success: the more people like the system, the more it consumes), maintenance of integrations. The rule of thumb we use in reviews: if the estimated annual run is below 20-30% of the build cost, some line is missing.

The honest account as an instrument, not a brake

The predictable objection to a TCO composed this way is that it inflates the numbers and kills initiatives. Experience says the opposite on both counts. The numbers are not inflated, they are anticipated: the lines exist regardless, and the only available choice is between seeing them in the business case or in the final statement. And the initiatives an honest account stops are exactly the ones that needed stopping, while the right ones come out reinforced, with a budget defensible in front of the board and without the season of supplementary requests that consumes more credibility than money.

There is also a strategic use of the full TCO worth naming: compared across the build, buy and wait options, with the five families filled in for each, it often changes the outcome of the fork, because the buy that looked expensive stops looking so once the build shows its perpetual run, and vice versa.

In our AI Assessment the business cases of the priority use cases come out with this structure, five families filled in, prerequisites spread, run estimated, and it is one of the points where clients measure the difference between a plan and a hope: calendly.com/fabiolalli/zerofive, or hello@zerofive.ai. The exercise to run meanwhile on the initiative you have in approval takes ten minutes: next to the visible line, write the other four families, even just as orders of magnitude. Which of the four, on your sheet, had stayed white?