Everyone wants to automate the wrong thing first. Usually it is the thing that looks most impressive in a demo, or the thing the salesperson suggested. The right first project is usually boring, and you can find it in about ten minutes with what we call the invoice test.
The test
Pick a repetitive task that runs on a schedule. Ask four questions:
- Is it the same steps every time? If the process changes shape from week to week, it is a judgment task wearing a repetitive costume. Automating it means automating your judgment, which is a different and harder project.
- Does it touch data from more than one system? If someone is typing or pasting between tools, that is the friction. The automation's job is to move the data without the typing — not to do anything clever.
- Can you see the result? You need a before-and-after you can count: invoices sent, hours spent, errors found, leads followed up. If you cannot measure it, you cannot tell whether it worked.
- Would a mistake be survivable? A wrong email draft is a ten-minute embarrassment. A wrong payroll run is a crisis. Start where mistakes are cheap, and put approval gates on anything that touches money, customers, or legal exposure.
Score each candidate. The task that passes all four — same steps, multiple systems, countable, low blast radius — is your first project. In our experience it is almost always invoicing, intake, reporting, or follow-ups. Boring, yes. That is the point.
What the first project is really for
The first automation is not about the hours it saves. It is about proving the model in miniature:
- that the process can be mapped as it actually runs, not as it is described,
- that the pipeline can move data between systems without anyone typing,
- that human approval can sit in the loop without slowing it down,
- and that the whole thing can run on your side, with your data staying yours.
Once that holds, the second and third projects get cheaper, because the plumbing is proven. Teams that start with the impressive project often end up with an expensive demo that nobody trusts. Teams that start with the invoice test end up with a machine that runs quietly, and a list of the next three machines.
What to do with the test results
If a task failed the test, it is not a failure — it means the task is a candidate for process redesign before automation, or a judgment task that should stay human. Deciding that something should not be automated is a legitimate outcome of the exercise, and we have told clients exactly that.
If one task passed, that is your starting point. The cheapest way to start is a written assessment: we map the flow as it runs, tell you what is worth automating, what is not, and what it would cost — before anyone builds anything. That is the whole engagement if you want it to be.
And if you would rather talk it through first, the first call is free and the advice is straight. You can read more about how we build AI automation if you want the details.