What Bad AI Actually Costs You

The cost of AI errors and hallucinations arrives on four lines, and only one of them appears on an invoice. First, direct remediation: refunds, damages, and rework when an AI output turns out to be wrong in public. Second, the verification tax: the hours your people spend checking outputs they cannot trust, which quietly erases the productivity the AI was bought for. Third, silent spend: token costs that grow run by run as context bloats, invisible until finance notices. Fourth, trust: the customer who got a confident wrong answer does not file a bug report, they leave. The two best-documented public cases involve a global consultancy and a national airline, and their lesson is not the size of the cheque. It is that liability attaches to whatever your system says, and that in both cases nobody had evidence the output was wrong until someone outside the company checked.

What do the documented cases actually show?

Deloitte and the Australian government. In 2025 Deloitte delivered a report on Australia's welfare compliance system under a contract worth about AU$440,000. Researchers found fabricated citations, and the corrected version disclosed that a GPT-4o based tool had been used in its preparation. Deloitte agreed to refund part of the fee; the OECD.AI incident record describes it as a partial refund and links reporting that puts the repayment at almost AU$98,000. The main recommendations were left standing, which is exactly the point: the content was mostly fine, and it still cost a five-figure sum and a public credibility hit, because the parts that were made up were the parts a reader could check.

Air Canada and one bereavement fare. A passenger asked the airline's website chatbot about bereavement fares, and the chatbot told him he could apply for the reduced rate retroactively. The airline's own policy page said the opposite. When the claim was denied, the case went to the British Columbia Civil Resolution Tribunal. Air Canada argued it could not be liable for the chatbot's statements, treating the bot as a separate entity; the tribunal called that "a remarkable submission" and held the chatbot is "still just a part of Air Canada's website". The finding was negligent misrepresentation, because the airline "did not take reasonable care to ensure its chatbot was accurate". Damages were small, $650.88 plus interest and fees, but the ruling is the expensive part: a wrong answer from your AI is your answer, and a correct policy page elsewhere on the site does not save you.

Why is the real cost mostly invisible?

Because the failures that cost money do not throw errors. Both cases above would have looked healthy on any dashboard: the report shipped on time, the chatbot returned a fluent reply with a 200 status. I cataloged this family of quiet failures in The Five Ways Production AI Agents Fail Quietly; the common thread is that the damage lives in the content, where request-level monitoring never looks.

The verification tax works the same way. Once a team has been burned, people start re-checking everything the AI produces, which is rational and also means you are now paying twice for the same work. Nobody budgets for this line, so it never gets measured, so it never gets fixed.

Silent spend is the one you can at least see in arrears. A prompt that quietly accumulates conversation history grows from hundreds of tokens to tens of thousands with no visible change in behavior. The bill arrives a quarter later, attributed to nothing in particular.

What does bad AI cost a small team?

Put the four lines next to each other for your own deployment. Direct remediation: what did the last wrong output cost to unwind, in refunds, credits, or re-done work? The verification tax: how many hours a week does someone spend checking AI output, multiplied by what those hours are worth? Silent spend: what does one run cost today versus at launch? Trust: how many of the customers who got a wrong answer told you, versus just stopped?

Most teams can fill in only the third line, and only roughly. That gap is the actual finding. You cannot know what bad AI costs you because you have no evidence of when it is bad; the absence of complaints is doing the work that evaluation should be doing.

How do you put a number on it before it happens?

Three instruments, in order of effort. First, read ten real production runs end to end this week; it costs an afternoon and it is where every team I have watched finds its first expensive surprise. Second, attribute cost per run and per customer, so silent spend becomes a visible trend instead of a quarterly ambush. Third, score answers against their evidence with a groundedness check, which is the one metric aimed precisely at the Deloitte and Air Canada failure shape: confident content not supported by the sources. I explained how it works in Groundedness: The Metric That Catches Confident AI Lies, and the relationship-flavored version of all five failure modes is in Your AI Agent Has the Same Red Flags as Your Ex.

Doing exactly this on your traffic, with the evaluations left running afterwards, is my Agent Trust Sprint: 2 weeks at about 4 working days, EUR 2,400 fixed, which is less than either of the cases above cost to discover the hard way.

FAQ

Is hallucination the biggest cost? It is the most visible one, because it produces a quotable wrong sentence. For most teams the verification tax and silent token growth cost more over a year; they are just spread thin enough that nobody owns them.

Does a disclaimer protect us? Not reliably. Air Canada had an accurate policy page and still lost, because the tribunal saw no reason a customer should know which part of the website to trust. A disclaimer is a mitigation, not a defense, and it does nothing for the verification and trust costs.

We are small. Does this apply to us? The amounts scale down; the mechanism does not. A refund on a wrong quote, an afternoon of rework, a customer who quietly churns: at 20 people those are still real money, and they are easier to catch, because reading ten runs covers a meaningful share of your traffic.

How much should we spend on evaluation? Less than the last incident cost you, and before the next one. A handful of evaluations tied to failures you have actually seen is usually days of work, which is cheap against any of the four lines above.