Every business we talk to has a list of things they'd like AI to do. Almost none of them have a list of what those things currently cost. That gap is where most automation projects go wrong — not in the model, in the math.
Start with the cost, not the capability
Before we propose an AI feature, we ask three questions: how often does this task happen, how long does it take a person right now, and what does a wrong answer cost if the system gets it wrong. A task that happens twice a month and takes five minutes is rarely worth automating, no matter how impressive the demo looks. A task that happens two hundred times a day and takes ten minutes is a different conversation entirely.
This sounds obvious written down. In practice, it's the step most automation pitches skip, because 'we automated it with AI' is a more exciting sentence than 'we did the arithmetic and it wasn't worth it yet.'
The pattern that actually pays off
Across the workflows we've automated, the ones that consistently pay for themselves share a shape: high volume, low judgment, and a clear definition of 'correct.' Categorising transactions, drafting a first-pass support reply, extracting fields from a form, flagging an invoice that looks late — these are jobs where a model has enough signal to be right most of the time, and where a human reviewing the exceptions is cheaper than a human doing every instance by hand.
- High volume: it happens often enough that saved minutes add up to saved hours.
- Low judgment: the 'right answer' is mostly determined by the data, not by taste or politics.
- Reviewable: a person can check the output faster than they could have done the task from scratch.
Where it doesn't pay off
The inverse pattern is just as useful. Low-volume, high-judgment work — the kind that involves a relationship, a negotiation, or a decision with real consequences — is usually a poor automation target, even when it's technically possible. We've turned down more than one request to 'automate' a task like this, because the honest advice was a better default setting or a simpler rule, not a model.
Build the boring part first
The unglamorous truth is that most of the value in AI automation comes from the plumbing: getting clean data into the system, defining what 'correct' means well enough to measure it, and building the review step that catches the cases the model gets wrong. The model itself is often the easy part. Projects that skip straight to 'add AI' without that groundwork tend to demo well and then quietly stop being trusted a month after launch.
If you're weighing an automation project right now, the fastest way to know whether it's worth building is to measure the current cost honestly first. That number tells you what the AI has to beat — and it usually tells you which part of the workflow to automate first.