STRUCTAIt reads your business. It asks before it changes anything.
Structa AI answers from the live rows in your own workspace — today’s bookings, last month’s takings, what a client has spent. Ask it to do something and it does not do it: it draws a card naming the exact thing, and waits for you to press the button. It has 15 of those, and no others.
It answers from your data, or it does not answer
This is not a general chatbot that happens to know your business name. Before your question reaches it, the assistant is handed the state of your business today — the rows your role can already open, and nothing else.
What it is holding when you ask
- Today’s takings
- Today’s and upcoming bookings
- Your clients
- Open requests
- Tables
- Open orders
- The kitchen queue
- Stock that is low or out
- Menu items marked unavailable
- This month’s best sellers
- Your house rules
Each line only if your business runs that module and your role carries it. A salon is never handed a kitchen queue, and is never told one is missing.
For anything older than today, it looks it up
get_revenueTakings for any range of dates — the total and the number of sales.get_top_itemsBest sellers for a range, ordered by what they actually brought in.find_clientOne client’s record — membership, visits, last visit, lifetime spend, stamps, notes, phone, email.get_laborLabor for a range — hours, cost, and how many people were late or absent against the schedule. Managersget_bookingsBookings for a range — booked, attended, no-shows, cancellations, still to come.get_clientsClient counts — new ones in a range, the total now, the members now.get_staff_salesSales per person — how many, net of refunds, and tips. Managersget_revenue_trendRevenue by day of the week, so the slow day is a fact rather than a feeling.get_retentionWho has stopped coming, repeat against one-time, and the clients worth the most.find_app_actionThe action list of an app you have connected, with that app’s own field names.
These run as you, under your own permissions, and change nothing — so they need no confirmation. The list on screen covers today and forward; a question about last March is answered from the books rather than guessed from what happens to be on screen.
It knows exactly as much as you do
The permission rule is the one the whole product runs on, and the assistant gets no exemption from it: you reach a module when the owner has turned it on and your role carries it. That intersection decides what the assistant is shown, and which actions it is allowed to offer you at all.
- A front-desk person who asks what labor cost last week is told that labor answers to managers. The refusal is in the database, not in the wording of a prompt.
- An action for a module your business does not run is never proposed — and would be refused again at the moment you pressed the button.
- Nobody learns anything through the assistant that they could not have opened themselves.
It proposes. You press the button.
There is no autonomous mode, and no setting that takes the card away. The card names the real thing — this client, this appointment, this amount — because a card that says “do the thing” is not a decision anybody can make.
Book a client in
A name, a service, a day and a time. Today unless you say otherwise.
BookingsOpen a request
A job for whoever is on shift, normal or urgent, on the board the team already works from.
RequestsSet a standing reminder
Once, every day, or every week, at an hour in your own time.
AI AssistantWrite a letter
A supplier about a late order, a customer who asked in writing. It lands unsent in the Drafts of the mailbox your business connected.
AI ReceptionistPost a notice to the team
News, or a task people tick off. Signed by whoever presses the button.
ManagersMove a request along
Open, in progress, done — found by its title and pinned to one row before the card appears.
RequestsAdd a client
A name, and a membership if they have one.
ClientsAdd an inventory item
Quantity, unit, and the level you want to be warned at.
InventoryRecord a payment
Who paid, what for, how much, card or cash.
Payments & POSClose a table’s order
Paid, with the total worked out on the server; or closed unpaid, which is a walkout and is recorded as one.
OrdersTablesCancel an upcoming booking
Future appointments only. It never touches one that has already happened.
BookingsMark a client as arrived
Today’s bookings, and it records the visit on the client at the same time.
BookingsMark a no-show
Today’s bookings, on the record where it belongs.
BookingsMove a booking
The card shows the old time and the new one. The new time has to be free for that staff member.
BookingsRun an action in an app you connected
The one action whose effect lands where no change of ours can undo it, so it takes manager rights and writes an audit entry before you are told it worked.
Open APIManagers
How a card behaves
- A booking action is narrowed to one appointment before the card is drawn. If your words fit two, it tells you and proposes nothing.
- Press the button twice and the booking is still made once. Each proposal carries its own identity, so a double tap or a lost reply lands on one row.
- Cancel, and it says so plainly: nothing was changed.
- Every action that runs is written to the log with the name of the person who confirmed it. An action in a connected app is written to the audit ledger as well, before you are told it worked.
- A notice to the team is signed by whoever confirmed it, never by Structa. Posting one takes manager rights, and so does running an action in a connected app.
The things no table holds
Who approves a refund over fifty dollars. Where the spare key lives. Which jobs are never promised for the same day. Every business has twenty of these, staff learn them by being told twice, and no amount of reading your database produces one. So a manager writes them down — a title and a few lines — and from then on they are part of the answer.
- A manager writes them. Everyone whose role carries the AI Assistant reads them, which is the point: the new hire is the person who most needs the answer.
- A rule outranks whatever the model would otherwise assume, and the answer names the rule it followed — so a rule that has gone out of date is visible instead of quietly wrong.
- If the rules cannot be read on a given day, it says that, and refuses to answer a question about your policy from what businesses usually do.
A memory, a clock, and a morning brief
Many conversations, and they are yours
Conversations are kept and titled by the first thing you asked, shelved under today, yesterday, the last week, older. Each belongs to one person: the rule in the database is a single line, so there is no state in which a colleague — or the owner — can open yours. They are left out of the owner’s data export on purpose. Clear all chats deletes them; the actions they took stay in the log, because those had consequences.
Remind me every Monday to count the stock
Said as a sentence, not filled into a form. It becomes a task on your own board at the hour you named, in your business’s time. It reminds you. It never performs the thing it is reminding you of — a run at six in the morning has nobody there to press the button.
Yesterday, in five sentences
The daily brief reads yesterday’s numbers and writes the operator’s summary, plus at most two actions the data plainly supports — each one a card, like every other. It never fires by itself — you press for it, and it costs what one question costs.
When it could not read something, it says so
A database hiccup used to come back as “you have no appointments today”. That is wrong in the most expensive way, because it reads exactly like an answer. So a failed read is not a number here. Every section handed to the model is one of three things — present, absent because you do not run that module, or named on a list of what could not be read this time — and the instruction that a named section is missing rather than empty is written once and shared by the chat and the morning brief, so the two cannot drift apart.
- If today’s takings could not be read, it says so. It does not say you took nothing.
- If it could not read what day it is, it refuses to work out what you meant by tomorrow, and asks you for the date.
- If your house rules could not be read, it will not answer a question about your policy from what businesses usually do.
- If most of the morning’s figures were unreadable, the brief is not allowed to call it a quiet day. It says which numbers are missing and that the brief is incomplete.
Every demo of every assistant is a confident answer. The question worth asking is what the thing says on the morning it cannot see.
What it costs to run
- A question costs 20 credits, and that is the whole answer — one that looks up three separate date ranges before replying still spends one question’s worth.
- The daily brief costs the same as one question.
- An action inside Structa costs nothing on top: a booking, a payment, a task, a reminder, a notice to the team.
- An action in an app you connected costs 40 credits, charged whether that app accepted it or refused it — the attempt is what we are billed for.
The software is free, and a new business starts with 2,000 credits that never expire. At zero it stops and says so — in a sentence that costs nothing to produce — rather than running up a bill. Top up and it carries on.
What’s actually true today
- Every answer is built from rows in your own workspace, filtered to the modules your role can already open.
- 15 actions, each behind a card a person presses. There is no automatic mode to switch on.
- Every action that runs is logged with the name of the person who confirmed it.
- House rules are written by a manager and read by everyone whose role carries the assistant.
- A read that failed is named as unreadable, and never reported as a zero.
- A full export any time and permanent deletion on request — with your conversations left out of the owner’s export on purpose.
What it will not do
- It never acts on its own. Every write waits on a card you press, and there is no setting that removes the card.
- It drafts letters; it does not send them. The draft lands in the Drafts of the mailbox your business connected, and without a connected mailbox it says so rather than pretending to have written one.
- It does not read the team chat. No tool of its touches those tables, and a test in our build stops one being added by somebody building something else.
- It cannot open a colleague’s conversation with it, and neither can the owner.
- It does not answer your phone or your public email. That is the AI Receptionist, a separate thing on a separate line.
- Labor and per-person sales answer to managers only, and the refusal comes from the database rather than from the prompt.
- There is a ceiling on questions per business per day, and a short pause if one person asks a great many at once. It tells you which one it hit.
- There is no model picker. Answers arrive whole rather than word by word.
- It writes no house rule of its own. A manager writes them.
- A reminder reminds. It does not do the thing at the hour it fires.
Questions owners ask
Can it invent a number?
It is given your data and told to use nothing else, and a section it could not read arrives named as unreadable rather than blank — so what it is instructed to say is that the figure could not be read. The failure this guards against is not invention; it is a broken read arriving as a confident zero.
Can an employee read what the owner asked it?
No. A conversation belongs to one person, by a rule in the database that is the same for reading and for writing. There is nothing to share and no setting that shares it. What a manager can still see is what the assistant DID: every executed action is in the log, business-wide.
What if it proposes the wrong thing?
You do not press the button, and nothing is written. Cancel and it says nothing was changed. A booking action that cannot be narrowed to exactly one appointment refuses to propose anything at all.
Does it work the same for a salon and for a restaurant?
The tools it is offered are assembled from the modules you actually run, so a salon is never offered a table to close and a restaurant is never offered a membership it does not sell. Nothing about a vertical is baked in.
Where do I use it?
A drawer beside your work, on every screen in the workspace, and a full page in the Business Console for the owner — conversations down the left, the one you are in on the right. Same assistant, same confirm card, either way.
Is this the thing that answers the phone?
No. The AI Receptionist answers guest email and live calls; this one works inside the business, for the people who run it. They are turned on apart and priced apart.
Looking for the one that answers guests? See the AI Receptionist →