Already running today
1:21 a.m. A customer writes in.
An hour later, the work is done.
These are real jobs our AI employees did recently, mostly for a subscription research service we run for a client. Every step happened and is on record. The agents do the work at any hour; a person approves what goes out.
Business and customer names are withheld.
The work log
- 1:21 a.m.October 14
Email sorting + Customer service
A subscriber asks to cancel in the middle of the night
A monthly subscriber replied “I wish to unsubscribe” to a report announcement. About an hour later the email-sorting agent had passed it to the customer-service agent, which switched off the renewal, confirmed the subscriber keeps access to the end of the paid month, checked that the site’s own record had updated, and drafted the confirmation reply.
The human partCancelling followed the manager’s standing rule that “unsubscribe” means cancel and stop charging. The manager only had to read the draft and press Send.
About an hour from the email to the cancellation
- MorningOctober 9
Customer service + Technical agent
A paying customer can’t sign in
They had paid under one email address and were signing in with another. The customer-service agent found the cause in the payment records, moved the account to the address they actually use, and drafted the reply. The same day the technical agent improved the support chat so the next customer is guided to the right address, after two independent AI reviewers checked the change.
The human partThe manager sent the reply, approved the chat changes and chose how the address hint looks.
Account fixed and reply sent the same morning
- 7:33 a.m.October 9
Customer service
“No customer should have to write three times”
After that case the manager asked that it never happen again. The agent built an hourly check of the support queue: a new message is handed over for a draft, the manager is told on WhatsApp when the draft is ready, and if a customer has waited six hours the manager gets a reminder until it is handled.
The human partThe manager approved it. It was running seven minutes later.
Live 7 minutes after the manager’s OK
- 7:24 a.m.October 9
Billing
A card dispute, answered with evidence
A cardholder disputed a $10 monthly charge. The agent pulled the timeline from the payment records and the support history, built a one-page evidence file with the emails and the site’s terms, and set up a daily check that reports the bank’s decision. A second dispute two days later got the same treatment.
The human partThe manager said yes before anything was submitted. The second time, the evidence went in 3 minutes after that yes.
About 50 minutes from the dispute to the evidence
- 3:00 p.m.October 11
Billing + Technical agent
A bot floods the newsletter form
While preparing the weekly report, an agent noticed newsletter sign-ups had jumped from about 8 a day to about 80, around the clock. The technical agent confirmed it was a bot and built two fixes: a hold so the fake addresses get no email, and an invisible bot check on the form. Two independent AI reviewers went three rounds on it and caught a problem that would have blocked every real sign-up.
The human partThe manager’s rule: ship only if both reviewers say so. They did, and it went live before the next morning’s email.
About 3 hours from noticing it to the fix being live
- AfternoonOctober 11
Billing
The accountant’s tax question, answered from the invoices
The accountant had found four different figures for one country’s European sales tax. The agent rebuilt all 189 European payments for the quarter from the invoices themselves, explained why the payment screens disagree, and drafted the reply with the spreadsheet attached.
The human partTwo AI reviewers checked it twice. The reply waits for the manager to send it.
The same afternoon
Running quietly in the background
Email sorting, every hour
An agent reads the manager’s inbox every hour and hands each email to the agent that owns that kind of work. In three weeks it handed over 243 emails to 14 different agents. It never replies, deletes or sends anything itself.
A WhatsApp assistant for the manager
The manager gives instructions by text or voice note on WhatsApp. The assistant passes each one to the right agent, word for word, and brings back the questions only the manager can answer. About 1,000 messages since October 1.
Month-end statements that draft themselves
The monthly revenue statement for a partner was drafted automatically on the 5th, checked against the payout records the same day, and sent by the manager three days later.
A post every weekday
An agent writes one original post for our company account each weekday, has it reviewed by a second AI, and posts it. Twelve went out in the first half of October.
Articles in a foreign language
For a content website, agents write and upload new articles in a foreign language. Twelve went up between October 11 and 15.
A watchdog for our own assistant
Every 30 minutes an agent reads the conversations of the WhatsApp assistant we run for friends and testers, and reports any place it let someone down. In its first week it found 24; sixteen are already fixed. One evening, four problems it reported at 7:29 p.m. were fixed and live by 7:45 p.m.
Your turn
What would your agents
handle tonight?
Tell us the work that piles up. We’ll show you which parts an AI employee can take, and where a person stays in charge.
Get a free AI assessment →