The number you have heard is that about 95 percent of corporate AI efforts return nothing. It is a real number with real caveats, and we will get to both. But a percentage is a place to hide. It lets everyone assume they are in the other 5 percent. So here is one failure with a name attached, told from the pitch to the bill, because the wreck teaches more than the average does. The instructive part is not that it failed. It is that the post-mortem reads almost word for word like every other one.
The pitch was completely reasonable
Air Canada put a support chatbot on its website. The pitch writes itself, and it is a good pitch: customers ask the same questions all day, a bot can answer instantly at any hour, deflect volume from human agents, and lower the cost per contact. This is the most defensible automation case there is. It is the same case behind every support bot, including the eval-first version we walked through that ships safely. Nothing about the goal was wrong. Almost nothing ever is. Failed AI projects rarely start from a bad idea; they start from a good idea with a step quietly skipped.
The skipped step
In November 2022, a customer named Jake Moffatt asked the chatbot about bereavement fares after his grandmother died. The bot told him he could book now and apply for the reduced fare retroactively, within 90 days, by submitting a form. That was wrong. Air Canada’s actual bereavement policy, on a page the bot even linked to, said the discount could not be claimed after the flight. The chatbot had not retrieved the policy. It had produced a fluent, plausible, made-up version of it.
That is the skipped step, and it is worth naming precisely because it is invisible in a demo. Nobody built the boring layer that constrains what the system is allowed to assert. There was no grounding of the answer in the real policy, no confidence gate that escalates a benefits question to a human, no evaluation set of real customer questions checked against real policy pages before launch. The bot was shipped to do the one thing these models do most confidently and least reliably: state a specific fact it was never given. In our triage walkthrough, the single most important component was the part that taught the system when not to trust itself. This system had no such part. It trusted itself completely, in writing, to a grieving customer.
The slow failure, then the bill
Failure here was not a crash. The bot ran fine. It answered fast, it deflected volume, the dashboard was probably green. The failure was latent, sitting in the gap between what the bot said and what the company would honor, waiting for someone to act on the wrong answer. Moffatt booked the full fare on the bot’s promise. When he applied for the refund afterward, a human agent told him the bot had been wrong and the airline would not pay.
He took it to the British Columbia Civil Resolution Tribunal. Air Canada’s defense is the part operators should tape to the wall. The airline argued, in the tribunal’s words, that the chatbot was “a separate legal entity that is responsible for its own actions.” Tribunal member Christopher Rivers called this “a remarkable submission” and held that “it should be obvious to Air Canada that it is responsible for all the information on its website,” whether it comes from a static page or a chatbot. The airline owed a duty of care, the bot’s answer was a negligent misrepresentation, and Air Canada paid: C$812.02 in total, C$650.88 for the fare difference, C$36.14 in interest, and C$125 in fees (Moffatt v. Air Canada, 2024 BCCRT 149).
The C$812 is not the cost. The cost is the legal proceeding, the precedent that now sits in every Canadian liability memo about chatbots, and a brand story that traveled the world under the headline that an airline blamed its own software. A skipped check that would have taken an afternoon became the most-cited corporate AI failure of the year.
The one story is the whole pattern
A single case proves nothing on its own, so put it against the research. The figure in circulation comes from a 2025 MIT study reporting that about 95 percent of enterprise generative-AI efforts showed no measurable bottom-line impact. Read it with care: it leans on roughly 150 leader interviews, a 350-person survey, and 300 disclosed deployments, companies are reluctant to volunteer their failures, and “no measurable impact” is partly a story about not measuring. Treat it as directional, not gospel. But it does not stand alone. S&P Global’s 2025 survey of more than a thousand enterprises found the share of companies abandoning most of their AI initiatives before production had jumped to 42 percent, up from 17 percent a year earlier, with roughly 46 percent of proofs of concept scrapped before they ever reached broad use. Two different instruments, the same shape: a lot of AI work dies on the way to production.
The most useful evidence is not the body count but the autopsy. RAND interviewed 65 data scientists and engineers and found AI projects fail at more than 80 percent, twice the rate of ordinary IT projects, for five recurring reasons: leaders misunderstand or miscommunicate the problem, the data is not ready, the team chases the latest technology instead of the real problem, the infrastructure to deploy and monitor is missing, or the problem is simply too hard for current AI. Notice what is not on that list. Not “the model was not smart enough.” Every one of those causes is a discipline that was skipped, and Air Canada’s bot managed to hit several: a problem framed as “answer questions” instead of “answer questions correctly or hand off,” no data grounding, no infrastructure to check answers against the source of truth. The story and the statistics are the same fact at two zoom levels.
What survives, and the one rule to take
The projects that return nothing are not unlucky. They are recognizable in advance, because they all declined the same unglamorous work: scoping the job narrowly enough to grade, building the evaluation before the build, putting a human at the edge where the model is unsure, and owning the output instead of disclaiming it. The teams that get return on AI and the agents that survive in production do the boring layer first. The failures do it never, or do it after the tribunal date.
So carry one rule into your next project. Before any AI system is allowed to speak or act in your company’s name, write down the ten things it must never get wrong, then build the test that proves it does not, and the gate that catches it when it does. If you cannot ship that test, you have not built a smaller version of the system. You have built Air Canada’s chatbot, and the bill is in the mail.