Ask Margaret Ellison on Sunday night whether it was a good week, and she says: busy. Ask whether Thursday went better than Friday, and she says she was on the pass.
Margaret is the chef-owner of Tampa Pasta House. Her week is written down — every order that came through, every message her staff answered, every post that went out, every item in her catalog. The numbers exist. Nobody reads them.
That reading is the job of one AI employee: Miles, AI Business Advisor.
The Monday below is an illustration in our own sample business, not a customer case.
He talks to you, not to your customers
Every other employee on the roster answers somebody: a caller, a buyer, a guest who wrote at nine in the evening. Miles has no channel at all and never turns up in a conversation with a customer. That is the design of the job, not a hole in it. His unit of work isn't a reply sent; it is one weekly review, waiting for Margaret in her workspace once the week has closed.
It starts in her own records
He counts from what her business already recorded: the orders and the items inside them, what sold most, the questions her staff answered and how many were still waiting for a person, the posts that were scheduled and whether they actually published, the post insights where the platform returned any, the catalog — what is active, what has no price, what she added — and the requests made against it. Plus her own decisions from the last review.
Nothing from anywhere else. He cannot see another restaurant's numbers, so he cannot compare her with one, and he will not invent the comparison. There is one baseline: her own week before.
Every number arrives with its window
"Orders went up" means nothing on its own. Beside each figure stands the window it was measured in — this week against the four before it — the source in plain words, and how much of the week it actually covers. A number without its window is an impression wearing the clothes of a fact.
A missing number never becomes a zero
If the platform returned no post insights this week, the review says exactly that. It does not print zero. On a chart an absent value and a measured zero look identical, and Margaret would act differently on each one, which is why this is a named duty and not tidiness.
The decision first, the evidence under it
Three to five facts of the week. At most three decisions, and under each: why it matters, what we don't know, and two or three options. Every option carries its trade-off, a small reversible first step, the number that would show it moving, and a recheck window Margaret picks herself. She never has to read a page to work out what he is driving at.
Then it is hers
Four answers to every decision: take it, rewrite it in her own words, defer it, reject it. What she chose is recorded together with the metric, where it started and when to look again. He will not choose for her even if she asks him to — it is her restaurant that lives with the answer.
A week later, the truth about her own decision
The next review comes back to that exact thing and says what happened: done · the window closed and the number has not moved · deferred by her · missed. Not "good progress." An advisor who flatters is worth exactly nothing.
What Miles never does
- gives general "how to grow a business" advice, or tells her what most restaurants do: he knows no industry averages, and a plausible opinion prints in a review exactly like a measured fact;
- forecasts revenue, profit or growth, or promises an outcome — he reports windows that have already closed;
- states costs, margins, ROI, churn or attribution: the product does not record them, so he asks Margaret for the figure and marks it as hers, not as measured;
- answers a tax, legal, investment, insurance or credit question: that stops before the answer, and what she gets instead is the category named, her own facts organized, and the questions to take to a licensed professional;
- spends, signs, sends, publishes or cancels anything, and never says "done" about something he did not do;
- writes to her customers, or reads the text of their messages: a review is built from counts and totals, not from what a guest wrote.
Hire the AI employee first
That is the order we suggest. Before you hire a person to make sense of the week, hire the AI employee who reads it, and see what is left. What is left for Margaret is what should be hers: the room, the menu, and the choice at the bottom of the page.
Setting him up is a conversation
Not a screen to learn. In the ChatGPT, Claude or Gemini app you already use, or in your workspace if you prefer it, you tell him what you count as a good week, what he should never suggest, and how long you want to wait before looking at a result.
See the work before you hire
Miles is the one employee you cannot write to: there is no channel to write to. What you can see is everything that feeds him. Tampa Pasta House and Casa Lista Home Repair are sample businesses we run ourselves, open to anyone at smashone.us/examples — their hired staff answers questions, files requests and posts in the open, and every one of those working days becomes a line in the records a review is made of. The advisor is the one you would add on top. And before you hire him, you can watch Miles describe his own job, including where he stops, in his own words at smashone.us/team/business-advisor.
