A 20-person design and fabrication shop in Fort Collins came out of a staff meeting with 11 ideas for where AI might help. The owner picked the one everybody was most excited about: a chatbot on the website to answer customer questions.
Six weeks later there was a half-configured tool, an unhappy office manager, and no answer to the question of who owned it. Nothing shipped.
Then they picked a different one off the same list. Proposal first drafts, built from the 30 proposals already in the shared drive. Writing a proposal went from about 5 hours to about 90 minutes, and it landed in 2 weeks.
Same company, same month, same tools. The only difference was which item got picked. That choice is most of the outcome, and it is a choice you can make well with 5 questions.
1. Does it happen at least weekly?
A task you do 3 times a year will never repay the setup, no matter how much you hate it. A task somebody does every Tuesday will repay it in a month. Frequency beats intensity every time.
2. Does one person own it today?
If you can name the person, you have a pilot, a tester, and someone who will notice on day 4 that the output has gotten worse. If the answer is “it depends” or “the team,” you have a coordination problem sitting in front of the AI problem, and the coordination problem goes first.
3. Is it text going in and text coming out?
Notes to summary. Examples to draft. Thread to answer. Documents to a list of fields. These work today, they work well, and they need no integration work. Anything that has to reach into your accounting system, your CRM, and your calendar at the same time is a real project. It might be worth doing later. It is a bad first one.
4. Can somebody check the output in under 5 minutes?
The review step is not optional, so it has to be cheap. A 2-page proposal draft can be read and corrected in 4 minutes by the person who would have written it. A 40-line reconciliation cannot be checked at a glance, which means either nobody checks it or the checking costs more than the drafting did.
5. What happens if it’s wrong and nobody catches it?
Rank the damage. A clumsy internal summary is an annoyance. A wrong number in a quote that goes to a customer is money. A wrong number in something regulated is worse. Start where being wrong is survivable, and you will have earned the right to try the sensitive one later.
Do this with your own list
Get your list of candidate ideas on one page. Ten or fifteen is normal. Score each one 1 to 3 on the first four questions, then cross out anything that fails the fifth badly.
Nearly every list sorts itself. You will end up with 1 or 2 items scoring 10 or better, and they are almost never the exciting idea. They’re usually the one where somebody writes the same kind of document over and over from material that already exists in a folder.
Start there. Give it 30 days, one owner, and a number you measured before you began. The exciting idea will still be on the list, and by then you will know a lot more about what your business can actually absorb.