Measure AI by margin, not by demos

Key takeaways

• If no number moves, it wasn’t a project — it was an experiment.
• Decide what success looks like in cost, cash, or revenue before you build.
• Measure the baseline first. You can’t prove a gain you never measured.
• The best AI wins often show up as freed-up cash and time, not just headcount saved.
• Pick two or three things that actually drive the business and measure those — don’t boil the ocean.

The question every AI project should answer

Before anything gets built, one question should have a clear answer: which number is this supposed to move? Cost down, cash freed, revenue up, time saved. If the honest answer is “it’ll be impressive,” that’s a red flag.

Impressive isn’t a metric. A $200,000 reduction in idle stock is. A 25% utilisation gain is. An 18% lift in sell-through with no extra marketing spend is. Those are the terms a business actually runs on.

A demo proves the technology works. A number proves the project was worth doing.

Measure the baseline first

The most common reason AI value is “unprovable” is that no one wrote down where things started. Before deployment, capture the baseline: what does this process cost today, how long does it take, how much cash is tied up in it, where does it leak?

That baseline is unglamorous and easy to skip. It’s also the only thing that lets you say, with a straight face, what changed.

Margin hides in places people don’t look

The obvious win is headcount you didn’t have to add. But the bigger wins are often quieter:

• Cash freed up — inventory that was sitting idle, now working.
• Time recovered — weeks of manual work compressed into hours.
• Leaks closed — slow-moving stock, half-empty trucks, missed delivery windows.

Inventory, for example, is usually treated as a cost. Used well, it’s a lever for cash flow — and that flows straight to margin.

---

Don’t measure everything

Trying to instrument every process at once is its own kind of failure. Pick the two or three things that genuinely drive your business — pricing, inventory, logistics, whatever yours are — get those measured and working, and expand from there.

The companies that get value from AI aren’t the ones running the most experiments. They’re the ones who decided up front what number had to move, measured it honestly, and then made it move.