A Week in the Life of an AI Back Office

Here is a week. Not a thought about a week, the actual grid, the way an AI back office in a small business would look pinned to the wall of a two-chair salon in Sydney's Inner West.

Four sittings with a computer. Ninety minutes across seven days. Three days where nothing runs at all.

The direct answer. An AI back office in a small business is not software that runs the place while you sleep. It is four or five jobs written down once, each one reading a file you exported or already had, each one producing a report or a draft that a person reads before anything leaves the building. The decisions arrive already made. A human still makes every one that matters.

Take Steph, the owner-groomer above. She is a worked example we follow through this series, and her numbers are the sort you would recognise rather than a real client's books.

Priya grooms four days a week. Deb does Saturdays and the front desk. Two groomers, around 40 dogs a week, and an admin evening of two hours, twice a week, that Steph is trying to get back.

What an AI back office in a small business actually is

This is the top rung of the ladder in this series, the one we call Running on rails, and the name is doing precise work. Not a driverless business. A track, so the same journey stops being a new decision every week.

Everything in that grid is a skill: a job written down as a file, in plain sentences, once. Anthropic's own advice on when to write one is the same test: the second time you find yourself explaining a job you have already explained. The mechanics belong to the guide to building a skill, the finished set Steph runs comes from the twelve-skill library, and the order they run in is the Monday morning routine.

Twenty-five posts taught the parts. This is one week of them running together with the friction left in, because a week with no friction is fiction.

Monday, 6:40am: the numbers, then the overdue list

First dog at eight. Nothing is running yet except the kettle.

Steph pulls two exports out of her booking system, drops them in a folder, and types one line: last week's numbers, then who is overdue. Two jobs, in that order, each printing its summary before the next starts.

The numbers come back as one page. Takings, dogs seen, no-shows, busiest and quietest day, and the single figure that moved most against the week before. It also comes back with a section headed with what could not be counted: three completed grooms with no price recorded against them. It refuses to average them or substitute the price list, and hands her three names instead.

That section is the reason she trusts the rest of the page.

Then the overdue list, worked out from each dog's own cycle rather than putting the whole book on six weeks. Eleven dogs are past due. Four are removed before a single message is drafted: one has a booking already on the calendar, one is marked do not contact, and two have been in twice, which is not a pattern. Every removal is printed with its reason.

Seven drafts. She reads all seven, deletes one because that owner rang on Saturday anyway, and sends six from her own phone. Then she presses Team Recap on her reporting overview, reads what it wrote, and pastes it into the team group chat, so Priya and Deb see it without a meeting.

Wednesday, over lunch: the questions you have already answered

Three messages came in overnight and two ask the same thing: what a full groom includes, and whether that changes for a doodle.

Steph answered that question properly once, in her own words, in a file. So the reply is not written fresh at 1:10pm; it is pulled from the answer she already gave and fitted to the dog in front of her. That move belongs to the twenty questions you answer every day, and the file it reads from is the prices, policies, and voice document.

Then next week's captions, seven of them, from a short list of the dogs she groomed and the same voice file. Two of the seven are flat. Not wrong, not risky, just the kind of sentence anybody could have written about any dog. She rewrites those two and keeps five.

Worth saying out loud rather than hiding: a week of captions gets you most of the way and never all of it, and the design half of that job sits with Canva AI.

Priya's whole contribution is 40 seconds long: a voice note in the car park after a hard dog, which comes back as a written handover note with a section for what needs a vet rather than a groomer. She never opens a laptop, which is the point of her being in this story.

Thursday, 5:30pm: Deb, the camera roll, and the receipts

Deb runs two jobs a week and no others.

The first matches the times on a folder of phone photos against the finish times in the appointments export, then copies them into per-dog, per-date folders. It never moves, never renames the originals, and never empties the inbox, so nothing can be lost by running it. Four photos this week landed in a folder called _ask-me, because Ziggy and another Collie finished a minute apart and a coin flip is not a match. Deb names them in one line.

The second reads the receipts, files a copy of each under a name carrying the date, supplier, and total, and writes the CSV the bookkeeper wants. Two receipts could not be read, one folded and photographed at an angle. They went to _needs-a-look with the reason beside them, rather than into the spreadsheet with a guessed total. The high-volume version of both jobs, the shoebox, is its own post.

One structural detail, because it decides whether a routine survives. Deb does not pull the exports. In ours, exporting is limited to Admin and Owner Admin, so it stays Steph's Monday job.

Friday, 4pm: who has not paid

Fifteen minutes, and most weeks nothing happens at the end of it.

The job reads the outstanding payments export and ranks it by days past due, oldest invoice at the top. It keeps apart two numbers everyone tends to add together: the total balance outstanding, and the total that is actually past its due date.

This week the list had four names on it. One had paid by bank transfer that morning, which the export did not know yet. That's not the software being wrong, it's the export being nine hours old, which is the most common reason a chase goes to the wrong person. Steph checked, crossed her off, and drafted two of the three.

Friday is looking, not chasing. Chase every week and you lose the client instead of the $95. The full close-off, the month's takings and the note for the bookkeeper, happens once, on the first Monday of the month, and nowhere else in the grid.

Saturday and Sunday: the two days nothing runs

Saturday is the busiest day of the week. Ziggy has a standing booking and the phone does not stop. Nothing runs, because there's nobody free to read what it would produce.

Sunday is empty on purpose, and that's the more interesting one. A job producing a report nobody reads is worse than no job, because the folder keeps filling and the trust quietly transfers to a number nobody has checked. So the week is built around a person's attention rather than a clock.

What does happen at the weekend is not AI at all. The confirmations and reminders go out on rules Steph set once, which is ordinary automation. If that is the layer you actually want, the plain automation guide is a better starting point than any of this.

Who does what, written out

Five rows, and the useful thing is how small the software's column is. Every verb with a consequence stayed with a person.

Nothing in that week was a clean run

Count the failures. A folded receipt nobody could read. Four photos that could have been either Collie. An invoice already paid. Two captions that read like a template. Three grooms with no price on them.

Three of those five are the same event: the job stopped and told her rather than filling in a plausible number. That's the property worth paying for, and the one to test in your first fortnight.

The invoice is a different shape. Nothing inside the job went wrong; the file was true when she downloaded it and nine hours out of date when she read it. No instruction fixes that, and only a person who can check the bank does.

The fifth, the flat captions, is the unfixable one. Arithmetic can be checked against a calculator. Writing cannot, so anything ending in a written judgement needs your eye on it every time, not just the first three runs. Why these things state a wrong thing confidently, and the ten-second habit that catches it, is its own post.

Nothing in that week sends itself

Every client-facing thing in the grid was drafted and read before it went. That habit is two lines of the never list, quoted here as post 4 wrote them:

And never let it:

• Send something you have not read.
• Delete something you have not backed up.

Six messages went out on Monday from Steph's own phone. Seven captions were posted by Steph. Two payment reminders were sent by Steph. The word doing the work in all three sentences is Steph.

There is a setting on her laptop that would let a job act with fewer questions, and the loosest option on it is one she will never touch. Post 19 names every one and says plainly which to leave alone. Anything about a dog's health, coat, or behaviour stays out entirely, which is where AI must never go in pet care.

What the software in this week can and cannot see

This part gets misdescribed constantly, so here it is flatly.

Nothing on Steph's laptop is connected to her booking system. There is no live link. Every job in the grid reads a CSV she downloaded or a folder already on her computer. A pet business owner cannot connect Claude, ChatGPT, or any other external AI to a Petboost account, and there is no token, key, or setting anywhere in our product that would let you try.

The export-and-then-analyse route, done properly, is asking your business a question.

Inside Petboost, three things in this week are ours. The confirmations, reminders, and payment notifications are rules, not AI.

Grazi AI is the assistant panel that answers questions about your own business data by querying it rather than from memory, included on every plan, Calendar Lite as well, at no extra cost. It is read-only: it cannot book, cancel, reschedule, charge, or send. It drafts, and you send.

It is designated experimental in our published terms, it can be wrong, and it is not instant; a typical question comes back in about 13 seconds and the slow ones closer to 50. And Team Recap, the button that wrote Monday's group-chat message, is the one AI writing surface in our reporting.

If you would rather ask out loud than read a page, call Emily on 1800 291 005. She is our own inbound AI phone agent and she books demos. She is not a receptionist for your salon and cannot answer your clients' calls.

What this actually reclaims for you

Add up the grid. Twenty minutes, 25, 30, and 15 is 90 minutes a week. Across 48 weeks, allowing yourself four off, that is 72 hours a year of back office running on rails instead of in an evening.

Be honest about what that is not. Steph's admin evening is two hours, twice a week, so 90 minutes is under half of it. The half that goes is the countable half: the figures, the overdue list, the photos, the receipts. The half that stays is the half that needs her to decide something, which is the correct half to keep.

And the number that paid for the whole week was found in its first 20 minutes. Three completed grooms with no price recorded, at her $95 casual full groom, is $285 that was never going to be invoiced, because nobody reads that far down a spreadsheet on a Sunday night.

That's one week. The routine did not grow her business. It stopped her leaking it.

If you want to see the reporting, the reminders, and the Grazi AI panel doing their part of a week like this one, Book a demo and we will run it on your diary rather than Steph's.