Stack Shift

Your first AI project should bore the room slightly.

Not a little bit ambitious. Slightly underwhelming.

Most companies pick the wrong starting workflow for a very human reason: they want something worth announcing. A result that justifies the investment in the original pitch meeting. An AI project that sounds good when the MD presents it to the board.

That instinct is exactly what produces pilots that fail quietly.

The right workflow is high-volume, genuinely resented by the team, expressible as a single number, and low enough risk that the sceptic in the room agrees to run it. That last point matters more than any of the others.

Gambling cannot be measured. Experiments can.

The boring win is not the consolation prize. It is the fastest route to the proof that funds everything else.

I wrote up the full thinking, including the four-part filter for finding the right first workflow, over on the Stack Shift blog

➜ link in the comments.

What is the most tedious, repetitive workflow your team would happily hand over tomorrow?

74% of AI pilots fail before they deliver anything measurable.

Most businesses assume the problem is the software.
It is almost never the software.

Here is where it actually breaks down.

Swipe through to see the full breakdown.

(The most common failure reason is one that barely anyone talks about publicly. It is also the one that is entirely fixable before you spend another dollar.)

If you have hit any of these, comment the number.
I am tracking which failure patterns are most common in SMBs right now.

NZ and Australian businesses spent over $4.2 billion on AI tools in 2025.

Most of them cannot tell you what they got for it.

Not because the tools did not work.
Because nobody agreed on what "working" would look like before they signed the licence.

No baseline. No KPIs. No agreed definition of success.
Just a budget, an announcement, and a hope it would become obvious.

Measuring ROI on a baseline you never captured is like weighing yourself after a diet without knowing how much you started at. You have a feeling something happened. You cannot prove it.

The businesses actually pulling ROI out of AI right now did one boring thing before they started. They wrote down what "normal" looked like. Hours per task. Cost per output. Error rate per week.

Twenty minutes of documentation.

That twenty minutes is the difference between "we think it helped" and "here is exactly what changed, and here is the number."

What metric do you wish you had captured before your last AI experiment?