AI projects rarely fail because of the technology. They fail because a solution is looking for a problem that is not big enough, or because the organisation cannot use the outcome. Four questions filter that out in advance.
Does it happen often enough?
Automation pays off at volume. A task that recurs ten times a day is interesting; a task that comes up twice a month almost never is, however annoying it may be. Multiply the number of occurrences per year by the time each takes, and put the build cost next to it.
Is a mistake survivable?
Language models do not offer one hundred percent certainty and never will. So the question is not whether the model makes mistakes, but what happens when it makes one. For pre-sorting email that is fine. For a medication decision it is not. In between, the answer usually lies in a human reviewing the doubtful cases.
Is your data in order?
A model drawing on outdated or contradictory documents produces outdated or contradictory answers — only with more confidence. For many clients, cleaning up the source is the real project, and the AI layer afterwards is a matter of weeks.
Who is going to use it?
The most common cause of a failed AI project is that it sits next to the existing way of working instead of inside it. If staff have to open a separate screen, it will not happen. Build it into the system where the work already takes place.
The best AI implementation we ever delivered was one we halved, after it turned out the other half could be solved with a simple rule.