I would start with work where you can explain the help you need and recognise a useful result. A first trial should make a decision easier, whether that leads to using AI or improving the task another way.
Look for a specific difficulty. A long document that needs a first summary is different from a recurring calculation. AI may help with the first; a formula may already handle the second reliably.
Compare the work, not just the ideas
These are illustrative candidates, not recommendations for every business:
Task | What AI might contribute | What to check first |
|---|---|---|
Drafting a follow-up from meeting notes | Organise varied notes into readable prose | Can the attendee check every decision and promise? |
Summarising published industry material | Produce a first outline of a long source | Can someone verify the important points against the source? |
Calculating a standard discount | Little advantage over a known formula | Does existing software already calculate it? |
Replying to customers automatically | Handle some questions without waiting for staff | What happens when an answer is wrong or the question falls outside its knowledge? |
A draft checked before use is often a manageable starting point. It still needs suitable information and enough time for review. “Internal” is not a guarantee of low consequence: an internal record can influence payments, employment or a customer decision.
Find out where the effort really goes
Watch someone do the task. If writing takes two minutes and finding the right information takes twenty, a writing tool may leave the main problem untouched.
Estimate the complete effort at the task’s usual volume. A frequent task may already be quick; a monthly task may consume a day. Include preparation and checking when estimating the possible benefit.
Check the information the tool would have. A customer agreement known only to a colleague cannot be reliably used by a system that has never received it. Record necessary facts in an approved source or leave that decision with the person.
Choose a trial you can learn from
Pick a task with suitable material, a clear output and someone able to judge it. Compare a template, search improvement or ordinary automation at the same time. Do not build a complex AI process to reproduce what existing software already does well.
If a candidate is worthwhile, use the trial guide to test it. If none is suitable yet, the work has still identified where better information or a simpler process would help.
About Jay
I’m Jay Janes, a Practical AI Advisor and founder of Reformat Labs. I write about choosing useful work for AI, checking the results and supporting the people using it.