How to choose the first process to automate with AI
· 6 min read
When a company decides to "do something with AI", the first question is usually which tool to buy. The better first question is which process to fix. The first project sets the tone: if it saves real time and people like using it, the next ten projects get easier. If it fails quietly, AI becomes a line item nobody wants to defend. Here is how we help teams choose.
Look for repetition, not complexity
The best first candidates are boring. They happen many times a week, follow roughly the same steps each time, and have a clear input and a clear output. Think of intake forms, routing incoming requests, copying data between systems, sending reminders, or preparing a weekly status report.
Complex, judgment-heavy work is tempting because it feels important, but it is the hardest place to prove value quickly. Start where the pattern is obvious.
- It happens at least several times a week
- The steps are mostly the same each time
- You can describe the input and the expected output in a sentence
- Mistakes are annoying, not catastrophic
Measure the pain before you pick
Before choosing, spend a week counting. How many times does the process run? How long does each run take? How often does something go wrong or get stuck waiting for someone? These numbers become your baseline, and without a baseline you cannot show that anything improved.
A simple spreadsheet is enough. The goal is not precision, it is having an honest before-picture you can compare against later.
Talk to the people who do the work
Managers usually know which processes are slow. The people doing the work know why. Twenty minutes with the person who handles the process every day will surface the workarounds, exceptions and unwritten rules that make or break an automation.
Those conversations also build buy-in early. People support changes they helped shape.
Score the candidates
Once you have three to five candidates, score each one on a few simple questions. You do not need a complex model; a rough 1–5 score on each is enough to make the trade-offs visible.
- Impact: how many hours or errors would this remove each month?
- Effort: are the systems involved easy to connect to?
- Risk: what happens if the automation gets it wrong?
- Visibility: will people notice and feel the improvement?
Start with a pilot, not a rollout
Pick one team, run the new process for a few weeks, and compare the results to your baseline. Keep a person in the loop for approvals and exceptions at first. When the numbers hold up and the team is comfortable, expand.
A small, visible win does more for AI adoption than an ambitious project that never quite ships.
