Most AI initiatives do not fail because the technology does not work. They fail because the business never agreed on what "working" actually meant.
Over the last few years the cost of capable AI has collapsed, yet the success rate of enterprise AI projects has barely moved. The pattern is remarkably consistent. A team gets excited about a model, builds something impressive in a sandbox, and then watches it quietly die on contact with real operations. The failure is almost never about the algorithm. It is about the strategy around it.
The real reasons projects stall
When we are brought in to rescue a stalled initiative, the root cause is usually one of four things, and often all four at once.
- The project started with a technology, not a business problem.
- No single person owned the outcome, only the build.
- Success was never defined in numbers the business cares about.
- Nobody planned for the messy last mile of adoption and change.
Each of these is a strategic failure dressed up as a technical one. A flawless model that solves the wrong problem is still a failure. A brilliant prototype that nobody in operations trusts is shelfware.
Start with the problem, not the model
The companies in the 20% do something deceptively simple. They start with a sharp, expensive business problem and work backwards to the smallest intervention that moves it. They resist the urge to build the most sophisticated system and instead build the most useful one.
Before we write a line of code, we insist on a one sentence problem statement that a non technical executive can repeat from memory. If you cannot say what changes in the business when this works, you are not ready to build.
"The best AI strategy is the one that survives contact with your operations team on a Monday morning."
Four checks before you build
We run every potential deployment through the same filter. It takes an afternoon and saves months.
- Value: if this works perfectly, what is it worth in ringgit or hours saved?
- Owner: who in the business is accountable for the result, not the build?
- Data: do we already have what we need, or are we hoping it appears?
- Adoption: who has to change how they work, and will they?
Measure what the business feels
Model accuracy is a vanity metric until it shows up in a number a leader already tracks. Tie every project to an existing operational measure: cost per transaction, hours per case, defect rate, time to decision. When AI moves a number the business was already watching, the conversation about value ends and the conversation about scale begins.
That is the whole difference between the 80% and the 20%. It is not better technology. It is better decisions about where to point it.
Back to insights