The AI costs nobody puts in the business case

The hidden costs of AI projects start after the licence is signed: staff reworking AI output (US$186 per affected employee a month, BetterUp and Stanford), training (5+ hours a person, BCG), a post-launch productivity dip, model migrations, a US$670,000 shadow-AI breach premium (IBM) and cancellation. On a 1,000-seat rollout, training and verification alone cost about 3.3 times the licence.

• The named list: the Below-the-Line Register, seven costs with an order-of-magnitude figure each (table below).
• Largest recurring line in the register: the verification tax. About 2 hours per bad AI deliverable, and 40% of US desk workers received one in the past month (BetterUp Labs and Stanford Social Media Lab, n=1,150, September 2025).
• Biggest one-off exposure: a high level of unapproved AI use added US$670,000 to the average breach cost (IBM and Ponemon, 2025).
• The contingent line: cancellation. Gartner forecasts over 40% of agentic AI projects will be cancelled by the end of 2027.
• Worked example: US$360,000 of visible licence cost carries about US$1.19m of quantifiable hidden cost in year one.

What are the hidden costs of AI projects? The Below-the-Line Register

The Below-the-Line Register is a list of seven AI costs that sit outside the build budget. Each one comes with the most defensible published figure we could find and the scope of the study behind it. Build-side costs (licence, build, integration, data remediation, run) are covered in what AI implementation costs, invoice by invoice. Everything in this register happens after that budget is approved, which is why it is missing from most business cases.

Lines 1, 3 and 5 are recurring or probabilistic costs for the whole organisation. Lines 2, 4, 6 and 7 attach to a single project. A business case that lists none of the seven is a budget for the purchase, not for owning what you bought.

What costs did I leave out of my AI business case? Start with the verification tax

The largest recurring line in the register is the time other people spend fixing AI output that looked finished and wasn’t. BetterUp Labs and the Stanford Social Media Lab call it “workslop”. In their September 2025 survey of 1,150 US desk workers, published with Harvard Business Review, 40% had received some in the past month. Each incident took about two hours to resolve. Priced at the salaries respondents reported, that is US$186 per affected employee per month, or about US$9m a year for a 10,000-person company.

The cost is hidden because it lands on the receiver, not the person using the tool. So it never shows up in the adopting team’s time savings. It shows up as a slower review queue two desks away. To measure it in a pilot, have the receiving team log every AI-assisted item they send back, with the minutes spent. That one column tells you whether the saving you are claiming is real or has just moved to someone else.

How big is the productivity dip after AI goes live?

Large-sample evidence that AI adoption lowers output before it raises it comes from US Census Bureau data. The CES working paper by McElheran, Yang, Kroff and Brynjolfsson (April 2025) compares manufacturing plants in 2017 and 2021. After controlling for size, age, capital and IT, a one-standard-deviation rise in AI use goes with 1.33 percentage points lower productivity. Once the authors correct for selection (firms expecting big gains adopt first), the short-run estimate grows to around 60 percentage points.

The authors attribute the dip to adjustment costs: more work-in-progress inventory, new robot investment and labour shedding. The losses were concentrated in older establishments, and earlier adopters grew faster later, conditional on surviving. The boundary matters: this is industrial AI in factories, not a chatbot in a sales team. The transferable lesson is to budget for the months when the old process is gone and the new one isn’t working yet. That period has a cost, and it belongs in the model.

How much training time should I budget for an AI rollout?

Budget at least five hours of training per user. That is the threshold at which BCG’s AI at Work 2025 survey found regular usage “sharply higher”, especially where in-person coaching was available. The same survey, of more than 10,600 leaders, managers and frontline staff, found only one-third of employees say they have been properly trained. For 1,000 users, five hours each is 5,000 hours of paid time, most of it spent away from normal work. Pricing training as the vendor’s onboarding webinar undercounts it by roughly that amount.

What happens to my AI project when the model is retired?

If you build on a model provider’s API, the provider sets your re-work schedule. OpenAI’s deprecations page promises at least six months’ notice before retiring a generally available model, unless safety or compliance requires faster. Preview models get less. gpt-4.5-preview was announced for retirement on 14 April 2025 and shut down on 14 July 2025, three months later. The whole Assistants API was given 12 months: announced 26 August 2025, shut down 26 August 2026.

Each retirement means re-running your test suite, re-checking prompts and guardrails, and re-approving output quality on the new model. This cost is per model dependency, not per project. An agent that calls three models inherits three retirement schedules. Seat-licensed SaaS moves this line onto the vendor, and so do outcome-priced services. In AI appointment setting for corporate sales teams, for example, LeadsNow is paid a 5–20% share of the sales it helps generate, so model migration and run cost sit with the provider. Training, governance and reviewing what reaches your own team still sit with you under any contract.

What does a shadow-AI data breach cost?

A high level of shadow AI, meaning staff using unapproved AI tools, added US$670,000 to the global average breach cost of US$4.44m in IBM’s Cost of a Data Breach Report 2025, researched by the Ponemon Institute across 600 organisations. Of breached organisations that had an AI-related incident, 97% lacked proper AI access controls. Across the study, 63% had no AI governance policy at all. This line belongs in an AI business case because banning AI tools without offering an approved one is itself a cost decision. The written policy that closes the gap is covered in the AI governance framework for business.

What does an AI error that reaches a client cost?

The clearest priced example is Deloitte Australia’s assurance review for the Department of Employment and Workplace Relations. The contract was originally valued at about A$440,000. After fabricated references and a fabricated quote from a Federal Court judgment were found, Deloitte refunded just over A$97,000, according to a DEWR spokesperson quoted by CFO Dive. That is about 22% of the original contract value. DEWR’s report page records a corrected version replacing the September 2025 original. The cash refund was the small part. The re-work, the public correction and the client relationship were not priced anywhere. Any AI output that goes to a paying client without human review needs its own row in the risk register, with an owner.

Worked example: the hidden costs of a 1,000-seat AI rollout

This is a worked example you can rerun with your own inputs. The licence is Microsoft 365 Copilot’s published enterprise list price of US$30 per user per month, paid yearly. The US$60 loaded hourly cost is an illustrative placeholder, so replace it with your own.

Two honest caveats. First, the BetterUp rate covers all AI use at work, not a single rollout, so your share could be lower. Second, the $186 is priced on US salaries. The example still leaves out the adoption dip, migrations, breach exposure and cancellation, because those are probabilities you have to estimate for your own organisation. Even at half the measured workslop rate, the two measurable hidden lines cost more than twice the licence. Carry those lines into the four-number AI business case a CFO will test as part of whole-life cost.

Frequently asked questions

What is the biggest hidden cost of an AI project?

Of the costs with a published figure, the one that recurs every month is the verification tax: time spent checking and redoing AI output. BetterUp Labs and Stanford found 40% of US desk workers received low-quality AI work in the past month. Each incident took about two hours to fix, costing US$186 per affected employee per month (BetterUp workslop research, n=1,150, September 2025).

How much should I add to an AI budget for hidden costs?

No published study gives a single multiplier, so build one line by line. In the 1,000-seat worked example above, training and verification alone came to about 3.3 times a US$30-a-month licence, and 2.1 times at half the measured workslop rate. Your own figures will depend on headcount, salaries and how much AI output moves between teams.

Why do so many AI projects get cancelled?

Gartner forecasts that over 40% of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls (Gartner, 25 June 2025). It is a forecast, not a count, but the money spent before cancelling belongs in any honest risk-weighted budget.

Does AI make a business less productive at first?

In US manufacturing, yes. A US Census Bureau working paper found short-run productivity losses before longer-run gains, concentrated in older plants (McElheran et al., CES 25-27, 2025). Comparable measurements for office and sales AI have not been published, so treat the dip as a risk to budget for, not a known size.

Who pays when an AI model is retired?

If you build on a provider’s API, you pay, in re-testing and re-approval. OpenAI promises at least six months’ notice for generally available models, and preview models have been retired three months after announcement (OpenAI deprecations). With seat-licensed or outcome-priced services, the vendor carries the migration.