The widely quoted failure rates are real and almost useless, because they lump together five quite different failures with different causes and different fixes. A model that was never accurate enough is not the same problem as a model that works and nobody uses, and neither is the same as an initiative that was approved without anyone naming the decision it would change.
- 01The headline failure rates are not one phenomenon. At least five distinct failures are being counted together.
- 02The most common is not technical: nobody could name the decision the output would change.
- 03The second most common is a prototype that works being mistaken for a system that can be relied on.
- 04Data quality is usually the real project, wearing an AI label.
- 05Almost every one of these is visible before the build starts, which is why the cheapest intervention is a question rather than an engineer.
Why do AI projects fail?
Because five different things are being counted as one. The reported rates are high and they are not wrong, but a single number covering all of them tells you nothing about which one is about to happen to you, and the remedies have almost nothing in common.
In rough order of how often we see them: nobody named the decision the output would change; a working prototype was mistaken for a production system; the data the model reads was never trustworthy; the thing shipped and nobody used it; and the model was genuinely not accurate enough for the job. Only the last of those is a modelling problem, and it is the rarest.
That ordering matters because it inverts the usual response. Teams facing a failure rate tend to hire for the technical end, and the technical end is where the smallest share of the failures actually sits.
What does "nobody named the decision" look like in practice?
It looks like a project with an output and no consequence. The dashboard gets built, the model scores well, the demo is impressive, and when you ask what happens differently on a Tuesday as a result, the honest answer is that someone looks at it.
The test is one question, asked before anything is built: when this produces an answer, who acts on it, and what do they do that they were not doing before? If that cannot be answered in a sentence by the person sponsoring the work, the project has not failed yet but it is going to.
This is the cheapest failure to avoid and the most frequently skipped, because the question feels obvious enough that everyone assumes it has already been asked.
Why does a working prototype so often fail in production?
Because the two are optimised for opposite things. A notebook is optimised for the author to learn something quickly. A production system is optimised to behave predictably for people who were not there when it was written, on inputs nobody anticipated, at a time when the author has moved on.
What separates them is unglamorous and almost always scoped last: explicit interface contracts, defined behaviour when a dependency misbehaves, observability, an evaluation set you can rerun, and a rollout that can be reversed.
The commercial consequence is that the demo creates momentum the engineering estimate then has to survive. By the time anyone prices the remaining work, a decision has effectively been made.
How much of this is really a data problem?
More than most post mortems admit. A model pointed at records that two systems disagree about will produce confident output derived from an argument, and no amount of model work fixes that.
The tell is simple. Ask which system is authoritative for the field the model depends on most. If the answer takes a meeting, the data work is the project and the AI is a label on it.
That is not a reason to abandon the work. It is a reason to sequence it honestly, because a data project described as an AI project will be judged against the wrong expectations and declared a failure when it was actually a necessary first phase.
What about projects that work and nobody uses?
This is the quietest failure and the one least likely to be reported as one. The system does what was asked, the team moves on, and usage decays until the thing is quietly switched off at the next renewal.
It usually traces back to where the output lands. An answer that arrives in a tool somebody already has open on a Tuesday gets used. An answer that requires logging into something else, at a moment nobody chose, does not.
It also traces back to trust. A system that was wrong early, visibly, and without explanation will not be believed later even after it improves, which is an argument for designing the failure behaviour before the happy path rather than after the first incident.
How do you tell which failure you are heading for?
Four questions, none of which need an engineer to answer. What decision changes when this works. Which system is authoritative for the data it reads. Who owns it in year two, and have they been told. And what would make us stop, stated before we start.
A project that can answer all four is not guaranteed to succeed, but it has ruled out the four most common ways of failing, which is most of the available risk reduction and costs an afternoon.
A project that cannot answer them is not ready to be scoped, let alone built, and that is a far cheaper thing to discover now than after the prototype has created momentum.
Is the honest answer sometimes not to build it?
Often. When the volume is low, the engineering, monitoring and maintenance do not get cheaper because the task is rare, and a rare task rarely repays them however irritating it is.
When the underlying process is undocumented, automating it encodes whatever it currently does, including the parts that are wrong, and makes them harder to see afterwards.
And when the real driver is that a competitor announced something. That is a genuine force in these decisions and worth naming out loud in the room, because it is frequently the one actually deciding, and it produces the projects nobody can defend eighteen months later.
We build AI systems, so this is worth stating plainly: the party telling you whether to proceed is usually the party paid to build it. That does not make them wrong. It does mean the question is worth asking somewhere with nothing riding on the answer.
We design and implement AI and automation for specific, identified work: research, document handling, workflow routing, decision support and internal operations.