What an AI Mistake Actually Costs Your Business

The refund policy the bot described did not exist.

Air Canada refused to pay. When the dispute reached British Columbia's Civil Resolution Tribunal, the airline argued that its chatbot was "a separate legal entity, responsible for its own actions." Tribunal member Christopher Rivers called that "a remarkable submission" and ordered the airline to pay Moffatt C$650.88 in fare difference, C$36.14 in pre-judgment interest, and C$125 in fees (2024 BCCRT 149). It was the first ruling of its kind in Canada, and the case is worth ten minutes of any business owner's time.

The payout was small. The cleanup was not. Air Canada spent staff hours disputing a fare difference, lawyers on an argument it could not win, and a news cycle explaining that its own website had misled a grieving customer. That is the real shape of an AI mistake. The tool bills by the seat, tens of dollars a month. The error costs whatever the cleanup runs, and the cleanup is where budgets actually break.

The subscription was never the price

Owners price AI the way they price software: seats, tiers, tokens. The failure modes don't live on that invoice.

A wrong answer has to be found first, and finding it costs staff hours. Somebody has to re-check every output the system touched, because after one bad answer you can't trust the rest. Records get re-done. Customers get remediated, sometimes with compensation. If data left the building, you are into disclosure conversations and, increasingly, legal exposure.

IBM and Ponemon's Cost of a Data Breach research puts the global average breach at USD $4.99 million in their 2026 report, a record high, up 12 percent in a year, driven mostly by detection, escalation and lost business rather than the technical repair itself. A ten-person shop will not average $4.99 million. The anatomy scales down faithfully anyway: the tool bill is trivial, and the trust damage is the line item that hurts.

The four failures we get called in to fix

We have been doing rescue work for fifteen years, and the AI calls all sound alike on the phone. A tool was supposed to save a few hours a week. It worked for a while. Then something went wrong, and nobody inside the company can say exactly what, when it started, or how far it has spread.

Almost every engagement maps to one of four patterns.

The confident wrong answer

The Air Canada pattern. A customer-facing bot answers policy questions from its general knowledge instead of from your actual documents, and it does this with complete confidence. It invents a refund window, a warranty term, a delivery promise. The customer screenshots it. That screenshot is now evidence.

We see the small version too: a quoting assistant that under-prices custom work because it helpfully ignored the rush fee, a support bot that promised a feature the roadmap never included. The amounts are smaller than an airline's. The customer still has the screenshot.

The guardrail: the bot answers only from source documents you control, and a human approves anything that commits money or policy. Log every conversation, so that when someone eventually asks what the tool told people, the answer exists.

Your data walks out the door

An office manager pastes a customer contract into a free public chat tool to get a summary. Bookkeeping drops financials in to build a variance explanation. Nobody means harm. The speed is real, so the habit spreads quietly, and one day you discover half the company runs on tools you never approved, sending data to servers you have never audited.

The same IBM research logged a 56 percent increase in AI-driven attacks, led by deepfake impersonation and AI-enabled malware. You don't control the attack side of that ledger. You do control how much of your data is sitting outside your walls when an attack, a vendor breach, or a plain policy change comes for it.

The guardrail: give people a sanctioned tool, with a plain-language allowlist of what may and may not be pasted into it. Shadow AI exists because the official path was slower than the unofficial one. So make the official path the fast one.

The pilot that never lands

The demo dazzles in week one. Six months later it is still a demo: the real data was messier than the sample, the workflow had exceptions the prototype skipped, and the person who championed the project got pulled onto something else. The licenses renew anyway.

The hidden cost is belief. A team that watched one AI project die in pilot purgatory will quietly sandbag the next three. We inherit that skepticism on rescue engagements, and it costs more to unwind than the original pilot did.

The guardrail: pilot against real, dirty data with one success metric and one hard date. Decide before you start what "good enough to keep" means, and name the business-side owner who makes that call.

The automation that quietly rots

This one is the most expensive because it is silent. An AI writes product descriptions, classifies support tickets, adjusts inventory records. Outputs get reviewed carefully for the first month. Then attention fades. Errors compound in records nobody looks at until a customer complains, or until someone asks why so much of the catalog now says the same wrong thing.

There is no dramatic failure event. There is drift: a widening gap between what the system says and what is true, and a cleanup bill that grows every week nobody looks.

The guardrail: every automated change carries an audit trail, a sampled human review on a fixed schedule, and a rollback path that takes minutes rather than a developer-week. If you can't undo it cheaply, don't let it run unattended.

The short version

Failure pattern	What it costs you	The guardrail
Confident wrong answer	Refunds, disputes, screenshots that become evidence	Source-grounded answers plus human sign-off on anything that commits money
Data walks out the door	Breach exposure, disclosure, lost trust	Sanctioned tool, plain-language data rules, fast official path
Pilot never lands	Sunk licenses and a team that stops believing	Real dirty data, one metric, one date, one named owner
Automation rots quietly	Compounding record errors and a growing cleanup	Audit trail, scheduled sampling, cheap rollback

One first step covers all four: list every AI tool actually in use at your company, approved or not. Most owners we ask can name one or two. The real count is usually higher, and you can't guardrail what you can't see.

What a repair looks like

When we get called in, the sequence is always inventory, contain, rebuild. Inventory: what tools exist, what each one touches, and which records they have already changed behind you. Contain: shut the unguarded paths, freeze the automated writes, pull the exposed data back inside. Rebuild: the same workflow, but with the guardrails above designed in rather than bolted on.

Notice what is missing from that list. No panic, and no rip-and-replace. Most AI mistakes are repairable, and the workflow usually survives. The expensive part is the weeks between the first wrong output and the day someone finally looks, because every one of those weeks adds records to fix and customers to win back.

One rule before you switch it on

AI is safe in a business exactly where two questions have answers: who checks this work, and how do we undo it? If a specific person reviews the output on a schedule you could name out loud, and a rollback exists that costs minutes, the tool can run. If either answer is "nobody" or "we'd have to figure that out," you are carrying the Air Canada pattern with your company's name on it.

If one of these four failures already sounds like your Tuesday, our AI development and integration work is where the rebuild starts. We will inventory what is running, contain what is exposed, and put the guardrails in place so the tool earns its keep without betting your records and your customers' trust on it.