We've built around those jobs in our own business. The question we start with is what someone no longer has to do by hand. Buying a tool comes later.
Start with one job that eats into your evening.
Try taking one repeated task off your hands, then count the work left to check.
Choose a job
Define its input and expected result.
Run a trial
Use synthetic examples and inspect the output.
Review the evidence
Keep unclear work with a person.
Expand only after reviewing the trial and how unfinished work is handled.
Find the job that follows you home
Think about your last busy week. Were you writing quotes after dinner? Copying enquiry details into a spreadsheet? Searching for documents someone had already sent?
Choose one of those jobs and describe what you want ready when you sit down. For example: a short reply to a routine enquiry, with the original message beside it. You still decide what to send, but you aren't starting with a blank page.
Be specific about what stays with you. Agreeing a price, making a promise or accepting a booking may need your judgement. The tool's permissions should enforce those limits. Asking it to be careful in a prompt doesn't prevent it from taking an action.
Check what your existing tools can already do. A rule may be enough to label or move a message; Gmail filters are one example. AI is useful where understanding the request or preparing the writing adds something.
Before building, we investigate how your team does the job now and confirm that understanding with you. We look at existing tools and local files, then decide what to keep, connect or rebuild. One system built for your business may make sense where several disconnected tools keep sending the work back to you.
Count the checking as part of the job
A draft that takes longer to correct than to write hasn't given you your evening back. During a trial, count the time spent opening the result, finding the source, checking it and putting mistakes right.
Start in a separate test environment with made-up examples. Include a request with missing details and a repeat of something already processed. Check that an uncertain answer is left for you rather than completed with a guess.
Keep a short record of the corrections. If the same price is wrong every time, look at where that figure comes from before changing the model or adding another subscription. The source information may need fixing.
We repeat tests after a change and use daily operation to find further problems to check. Compare the cost of running the system with the work it removes. Include the subscriptions and the time spent correcting it.
You should be able to step away without losing track
An unfinished task needs somewhere visible to wait and someone responsible for it. Agree how you will find the jobs that stopped, and how to finish them manually.
Ask what happens if a stopped job runs again. It should recognise work already finished rather than create a second enquiry or another task for someone to chase. Test that before depending on it.
The result you're looking for is a smaller pile of work waiting for you, with a clear view of anything still needing your attention. Before expanding the system, compare that pile and its checking time with what you had before.