AI SYSTEMSWHAT GOES WRONG
Why Most AI Projects Fail.
Not because the technology does not work. Because it gets pointed at the wrong thing. Here are three well-documented examples, none of them ours, and the single pattern sitting underneath all of them.
Most of What Is Being Sold Is Not What It Says.
40%
of agentic AI projects will be cancelled by the end of 2027, Gartner expects, on escalating costs, unclear business value, or inadequate risk controls.
Their analyst's explanation is blunt: most are “early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied.” And the line worth reading twice before you buy anything: “many use cases positioned as agentic today don't require agentic implementations.”
The same research names a practice it calls agent washing: rebranding ordinary chatbots, assistants, and rule-based automation as agents. Gartner reckons only around 130 of the thousands of vendors selling agentic AI are the real thing.
Gartner (opens in a new tab), 25 June 2025
Two Companies Who Found the Edges.
Neither is our client and we had nothing to do with either. They are public, well documented, and worth more than any testimonial, because both companies paid to learn something you can now have for free.
APPLIED BY COST, NOT BY FIT
Klarna, and the round trip
In February 2024 Klarna announced an AI assistant handling customer service, said it was doing the work of 700 full-time agents, and projected forty million dollars of profit improvement that year. It became the most-cited AI success story in business.
By May 2025 they were hiring humans again. Their chief executive's own explanation is the part worth keeping: “As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality.” They now run both, with AI on the repetitive work and people on anything needing judgment.
AI did not fail them. They asked what it would save before they asked where it belonged.
Klarna's 2024 announcement; the 2025 reversal reported by CX Dive (opens in a new tab) and others, May 2025
YOU OWN WHAT IT SAYS
Air Canada, and the $650 sentence
A man booking a flight after his grandmother died asked Air Canada's chatbot about bereavement fares. It told him he could apply retroactively. That was wrong, and the airline refused the refund.
Air Canada argued, in effect, that the chatbot was responsible for its own statements. The tribunal rejected that outright, found the airline had not taken reasonable care to ensure the chatbot was accurate, and held it liable for negligent misrepresentation.
Whatever you put in front of customers speaks with your voice and binds you like your voice.
Moffatt v. Air Canada, 2024 BCCRT 149, British Columbia Civil Resolution Tribunal, February 2024
The Third Failure Is Quieter.
Almost nothing goes wrong on day one, because day one is when everyone is watching. What kills these systems is month eight: the person who understood it moved on, the prices changed in March and nobody told it, and it has been confidently giving out last year's answer ever since.
That is what most of Gartner's cancelled projects look like. Not explosions. Drift. Which is why the work cannot finish when the thing works.
Read the transcripts
Actual conversations, on a standing schedule, looking for where it fumbled or hedged or escalated something it should have handled. Every one is a gap to close.
Keep it current
Prices move, service areas change, you stop taking a kind of job. An agent that has not been told is telling your customers something that stopped being true months ago.
Widen it slowly
As parts earn trust, it can do more without asking. That should happen because it proved itself over real weeks, never because it was set up that way on day one.
One Pattern, Three Failures.
Nobody in these stories was defeated by the technology. Klarna optimised for cost before fit. Air Canada shipped something customer-facing without deciding what it was allowed to say. The quiet failures were never revisited after launch. In all three the missing step is the same one, and it happens before anything gets built.
Automation applied to an efficient operation will magnify the efficiency. Automation applied to an inefficient operation will magnify the inefficiency.
Bill Gates, The Road Ahead, 1996
Thirty years old, and it still decides the outcome. It is why we work out where AI fits before building anything, and why sometimes the honest answer is that your process needs sorting out first.
The Short Answers.
+What percentage of AI projects fail?
Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing escalating costs, unclear business value, or inadequate risk controls. Their analysts describe most current projects as early stage experiments driven by hype and often misapplied.
+What is agent washing?
Agent washing is rebranding ordinary chatbots, assistants, and rule-based automation as AI agents. Gartner identified the practice in 2025 and estimated that only around 130 of the thousands of vendors selling agentic AI were offering the real thing.
+Is a business liable for what its AI chatbot tells customers?
In the leading case to date, yes. In Moffatt v. Air Canada (2024 BCCRT 149), the British Columbia Civil Resolution Tribunal rejected the airline's argument that its chatbot was responsible for its own statements, found the airline had not taken reasonable care to ensure the chatbot was accurate, and held it liable for negligent misrepresentation.
+Why did Klarna reverse its AI customer service rollout?
Klarna announced in 2024 that an AI assistant was doing the work of 700 full-time agents. By May 2025 it was hiring people again. Its chief executive attributed the reversal to cost having been too predominant an evaluation factor, which produced lower quality. Klarna now runs both, with AI on repetitive work and people on anything needing judgment.
Rather Not Be in That 40%?
Start with the free live look. Thirty minutes on how your business actually runs, and an honest answer about whether AI belongs in it yet.