Most teams that ask about AI automation start with the most impressive idea they have. The best first project is usually the most boring one.
Here is the test I use with clients. Score each process you’re thinking about on four questions.
1. Does it happen every day?
Volume beats complexity. A five-minute task done forty times a day is worth more than a two-hour task done once a month. It also gives you lots of real examples to test against before you go live.
Good sign: you can name the person who does it, and they groan when you mention it.
2. Does it start with a document or a message?
Invoices, application forms, ID cards, purchase orders, customer emails, photos of handwritten sheets. AI is now very good at reading these and turning them into clean, structured data. That step is where most of the retyping, and most of the mistakes, happen.
Good sign: someone is copying details from one screen into another.
3. Can a person tell when it’s wrong?
You want a process where a mistake is easy to spot and cheap to fix. Then you can put a review screen in front of anything the AI isn’t sure about, and the system learns from every correction.
Bad sign: errors only show up weeks later, in an audit or a complaint.
4. Is the “right answer” written down somewhere?
If the rules live in a policy document, a price list or a spreadsheet, AI can follow them and show its source. If the rules live only in one senior person’s head, start by writing them down. That work is useful whether you automate or not.
Put it together
| Process | Every day? | Starts with a document? | Easy to check? | Rules written? |
|---|---|---|---|---|
| Customer data sheet to record | Yes | Yes | Yes | Yes |
| Supplier invoices to accounts | Yes | Yes | Yes | Mostly |
| Daily sales summary | Yes | No, data | Yes | Yes |
| Pricing a complex custom deal | No | Partly | Hard | No |
Four yeses: start there. The last row is a real use for AI too, but as an assistant that drafts a quote for a person to finish, not as automation.
What a first project looks like
A good first project fits in a few weeks:
- Collect 50–100 real examples of the documents.
- Measure how often the AI gets each field right on your examples, not on a demo.
- Build the smallest version: documents in, a review screen for anything uncertain, data out to where it already goes.
- Run it alongside the current process for a week, then switch.
If you have a process that scores four out of four, write to me with a few lines about it. I’ll reply with how I’d automate it.