Most AI projects in SMEs run aground on three things:
- they started too big
- no KPIs were agreed up front
- the knowledge stayed with an external party instead of the internal team
A project that starts small, has a clear measurable goal and leaves its knowledge behind rarely fails outright. At worst it delivers less than hoped, and you know that within weeks.
Starting too big is the most common. An eighteen-month programme with a seven-phase roadmap sounds solid, but you only find out in a year and a half whether it was worth anything. A four-week project on one process tells you now. Cut the plan back until it fits on one page; what does not fit, you keep for later.
The second mistake is everywhere: no KPIs up front. Without a number fixed in advance (the lead time of a quote, say, or how quickly a customer gets an answer) almost any outcome can be presented as a success in hindsight. Everyone talks the project up and nobody learns anything. So agree before you start what has to improve, and measure the same thing at the end.
The third point is the most expensive over time: knowledge that stays with the supplier. If it works but you cannot change anything without calling the builder, you have bought a subscription to dependency. Dependency becomes part of how you operate. Make sure your own people understand the approach and start innovating themselves. That is the difference between a project that stays alive and one that stalls the moment the invoice is paid.