We've just put the first part of an Agentic Virtual Assistant live inside an Australian service business.

Not a chatbot on the website. Not another content generator. It checks the work that quietly falls between systems and people.

The first job was simple to describe: find completed appointments that should have been invoiced but hadn't been.

Simple is doing a fair bit of work in that sentence.

The business uses different systems for different parts of its operation. There are individual appointments, group attendees, cancellations, no-shows, prepaid services and account credits. A report that gets most of that right is worse than useless if someone trusts it.

Our first version was wrong.

It checked individual appointments but missed attendees attached to group appointments. The API worked perfectly. The answer didn't.

Connecting the software is not the same as verifying the answer. A lot of AI projects stop at "the API worked" and call the job done.

We went back to the business owner's manual report and compared the result properly. His answer was three exceptions. Once we'd corrected the missing record type, our answer was three as well.

Then we ran it silently for a week.

Across seven runs, the assistant examined 295 appointments and returned 26 items that needed attention. A fresh live count of five also matched the owner's report.

Only then did we let it post into the team's existing Slack channel.

Even there, it doesn't dump customer information into a chat room. Staff see the total and a breakdown by practitioner. No patient names. No appointment details. No clinical notes. No automatic invoices.

The system finds the exception. A person still handles the judgement and the money.

We've since added two more jobs.

One finds clients who were attending, haven't returned and have nothing booked. The business owner checked the first 12 candidates directly in his system. All 12 were genuine follow-up opportunities.

The other checks whether required treatment notes are complete. It reads the completion status, not the note itself, and reports aggregate counts to the team every Friday with a Monday catch-up.

Most business owners don't care about the technical distinction between an AI agent and a chatbot. They care whether it can take recurring work off someone's list.

The value isn't having another place for staff to type questions. Most businesses already have enough chatter.

The value is removing recurring work before anyone has to ask.

Check yesterday's appointments. Find what wasn't invoiced. Identify clients who may have slipped through the cracks. Show the team what needs attention. Do it on schedule, inside the tools they already use, without exposing information they don't need.

That's much closer to replacing part of a VA workload than asking ChatGPT to write an email.

We also learned when to say no.

One requested workflow depended on a part of another platform that its public API doesn't support. We could have built a browser robot to click around the private dashboard and hoped the interface never changed.

We didn't.

It was too brittle. A browser robot that quietly breaks when the interface changes is not progress. The better question may be whether this platform should be replaced with one that has a usable API.

The first pilot has taken real work. We had to learn the business rules, build an isolated environment, test against manual answers, tighten the privacy boundaries and prove the schedules. The next one should be faster because the runtime, testing process and delivery patterns now exist.

So I'm looking for the next business to do this with.

The first step is one job, not a grand AI transformation. We map how it is done now, test our answer against yours and run it silently before anything reaches the team.

The best fit is an owner-led service business with recurring admin spread across a few systems. You probably have a VA list, or a mental list, containing jobs such as:

• checking for missed invoices or incomplete records
• finding customers who need follow-up
• comparing reports between two systems
• preparing a daily or weekly exception list
• turning a recurring manual check into a draft for approval

I'm not looking for someone who wants an AI chatbot because AI chatbots are fashionable. I'm looking for a business with one boring, expensive job that happens every week and has a clear right answer.

If that sounds like your business, reply with NEXT CLIENT and tell me the job you would happily never check manually again.

I'll tell you honestly whether it is a good automation candidate, a bad one, or something a cheap app already solves.