How do you know an AI project is working before you spend more?
An AI project is working when it runs on real work, the people using it can direct and correct it without the builder, and the leader who owns the result would say out loud that it is useful. Usage charts and hours saved are early signals; the test that matters is whether a decision got better, and whether that would survive the builder leaving.
Most guidance measures AI projects by adoption and time saved. Those numbers can climb while every decision stays the same, so the better test is whether the person who owns the result will vouch for it.
Most AI projects report good numbers early. More people log in, more tasks run through the tool, and someone estimates the hours saved. Then the money is approved for the next phase, and a year later it is hard to say what changed.
The early numbers were not wrong. They were measuring use, and use is easy to grow. What a leader needs to know is whether the business is making better decisions, and whether it could keep doing so on its own.
Why adoption numbers mislead
Usage rises for many reasons: a launch email, a mandate from the top, a novelty that wears off. Time saved is usually an estimate made by the people with a stake in the project. Neither tells you whether the work got better or what the saved time was spent on.
Both can look healthy while every important decision is made exactly as before.
Four checks
Before you spend more, ask four questions about the project.
- Does it run on real work? Demonstrations use tidy examples. Real work is messy, late and full of exceptions. A project that has only seen tidy examples has not been tested.
- Can the people using it correct it? If a wrong answer needs the builder to fix it, the project depends on the builder. Good systems let users correct them in plain words and keep the correction.
- Would it survive the builder leaving? Imagine the team that built it leaves next week. If the system would quietly stop being used, you have bought a service.
- Will the owner of the result vouch for it? Ask the person accountable for the outcome whether they would recommend it to a peer, out loud. Hesitation is data.
A project that passes all four is worth scaling. One that fails two or more needs fixing before it gets more money.
Make the first version small enough to prove
The easiest way to pass these checks is to start small. Pick one decision or capability, build a working first version, use it on real work and keep what works. A small system that a leader will vouch for teaches you more than a large one nobody owns.
Where Mindmake fits
Mindmake is a small AI and commercial strategy practice, and this is how its work is built. Each piece of work starts with one paid proof: one decision or capability, built as a working first version, used on real work and kept by the leader when the work ends. It can start from the leader's own judgement, with Build your AI brain, or from the business's pricing, positioning, product and people, with Build your AI revenue engine. The verified client stories are on Success stories.
Where we stand on this
- A proof counts as working when the client will stand behind it.
- When the work ends, something is running that was not running before, it is used on real work, and the leader can direct it without us.
The questions that follow
What should we measure first?
Three things: which real decisions or pieces of work it touched, what the people using it had to correct, and whether the owner of the result would use it again tomorrow.
When should we stop an AI project?
When nobody who owns a result will vouch for it, or when it only works while the builder is in the room. Either one means you are paying for a demonstration.
Is the test different for a business backed by private equity or venture capital?
The test is the same and the clock is shorter. Investors judge a business on what it can prove, so a working first version used on real work is worth more than a plan for a bigger one.