Start with a job you know
To see what an agent can do, give it a job you already do every week.
Pick something small and repeatable. You already know what a good result looks like, so you can judge the first run in minutes.
Once one job runs well, the next one takes a fraction of the time.
Pick the right job
Good first jobs repeat often, happen in apps you already use and end with something you can check.
- You do it at least once a week.
- It happens in web apps like Stripe, Gmail or HubSpot.
- You can check the result in under 5 minutes.
- A mistake is easy to spot and cheap to fix.
Chasing failed payments, sorting a shared inbox and cleaning up a CRM are all good places to start.
Write the ask
Write the job the way you would ask a new teammate. Say what to open, what to do and what you want back.
Name the apps
Tell the agent where the work lives. “Find failed payments in Stripe” is clearer than “find failed payments”.
Say what done looks like
Describe the result you want. Drafts in Gmail, a row in a sheet or a message in Slack. The agent stops when it has that.
Say what needs your OK
If a step sends, pays or deletes something, say that it should wait for you. The agent stops and asks before that step.
“Find last week's failed payments in Stripe and draft a follow-up email for each one.”
Watch the first run
Open the chat and watch the screen while the agent works. You see every page it opens and every button it clicks.
It asks when it is stuck. If something is unclear, the agent stops and asks you in the chat instead of guessing.
You can take over. Take control of the screen at any time, fix the problem and give control back.
Fix it once
The first run is rarely perfect. Tell the agent what to change in plain words, like you would tell a person.
- “Use the customer's first name.”
- “Skip customers who already have an open ticket.”
- “Save them as drafts. Don't send anything.”
Make the fix stick. Ask the agent to remember the fix. It saves it to its memory and uses it next time.
Do this next
Once the job runs well twice in a row, it is ready for a schedule.
Before that, connect the rest of the apps the job needs. The next guide shows how to do it safely.