What an AI assistant should actually do for a booking business

Every booking tool has an AI feature now. Most of them are the same thing: a chat window in the corner that will cheerfully tell you what your dashboard already told you, at greater length. Ask it how last month went and it writes four paragraphs about how last month went.

That isn’t useless, exactly. It’s just not worth anything. You didn’t need a language model to read a number off a screen you were already looking at, and the novelty of asking it wears off in about a week.

The useful question isn’t whether the thing can hold a conversation. Modern models all can. It’s whether it saves you something you actually spend — a trip through four screens, a decision you’d have put off, twenty minutes at the end of a day when you’re already done.

Answering questions is the easy half

The bar most AI features clear is fluency. The bar that matters is whether the answer arrives without you going and getting it.

There’s a specific kind of question a small business owner has constantly and almost never answers, because answering it means opening the calendar, then the reports page, then the customer list, and cross-referencing three things while someone waits at the counter. Which service actually makes me money? Who used to come in every month and hasn’t been back since winter? Is Thursday quiet again, or does it just feel that way?

None of those are hard questions. They’re just ten minutes each, and ten minutes you don’t have is the same as impossible.

It has to know your numbers, not the internet’s

An assistant that can only talk in general terms about the beauty industry is a search engine with worse manners. The thing that makes one worth having is that it can read your own bookings, your own revenue, your own customers — and nobody else’s.

The Bookka Assistant has eight ways of looking at your business: booking volumes, top services, customer patterns, team performance, the week ahead, cancellations, revenue trends, and whatever’s sitting waiting on you right now. You ask in plain words and it goes and looks. Press ⌘I anywhere in the dashboard and it’s there.

What it can see is bounded by who you are. An owner asking about revenue gets revenue; a staff member asking about the schedule gets their own schedule and nothing else. That isn’t a setting someone can talk it out of — the assistant is handed your role by the server, not by the conversation.

The useful version can do things — which is the frightening part

Reading is safe and therefore limited. The genuinely useful version of this closes the loop: mark the booking complete, take the cash payment, drop the stock count, draft the email you’ve been meaning to send since Tuesday.

And that should make you slightly uneasy, because the failure mode of a chatty assistant with write access is that it confidently does the wrong thing to a real customer’s real money.

So on Bookka it’s switched off until you switch it on. A new business gets an assistant that can look but not touch, and making changes is a deliberate opt-in under Settings. The obvious way to ship this feature is to have it do things on day one and impress people. We deliberately didn’t.

So it should have to ask, every time

When the assistant is about to change something, it stops and shows you a card: this booking, this customer, this amount, this stock count. Confirm or decline.

The detail that matters is where that card gets its wording. It’s built from the actual instruction being carried out — not from the model’s description of what it’s about to do. Those two things are usually identical, and the entire point of the design is the cases where they aren’t. A summary written by the thing asking for permission isn’t permission, it’s a sales pitch.

If a card goes unanswered and the conversation moves on, it counts as a no. Silence is never consent to spend someone’s money.

And you should be able to check what it did

Approval at the time isn’t much use if you can’t reconstruct things three weeks later, when a customer rings about a payment marked cash that they’re fairly sure they paid by card.

Every action the assistant carries out lands in your audit log reading via Bookka Assistant, next to the person who approved it. You can filter the log down to just its work and see the lot in one list.

That label is issued by the database at the moment the action runs, not attached by the assistant afterwards, and it can’t be claimed by anything else. An audit trail the tool being audited could write for itself would be worth precisely nothing.

Being told before you think to ask

The real limit of any assistant is that you have to remember it exists. The things that quietly cost you money are the ones you never thought to check: four bookings sitting unconfirmed since Friday, a regular who’s drifted off, a shelf that’s about to be empty.

So Bookka checks five of those every morning at seven, in your own timezone, and tells you if any of them are true:

• Bookings still waiting on you — pending more than twelve hours, or due within two days.
• Tomorrow looks quiet — under half what that weekday normally holds, while there’s still time to do something about it.
• Regulars who’ve drifted — three or more visits, nothing in the last eight weeks.
• Stock about to run out — anything you track that’s down to five or fewer.
• Work you did but never got paid for — completed more than two days ago and still unpaid.

None of that involves a model, and that’s deliberate. They’re five plain rules. You want the same answer every morning for the same facts, you want it to be right rather than fluent, and rules cost nothing to run — so there’s no reason to pay a language model to do arithmetic. If one lands while you’re working, it arrives as a notification with a question already loaded, and the assistant picks it up from there.

What it should cost

Most AI features are priced as though every question is expensive. Most questions are not: a typical exchange with the Bookka Assistant costs about a cent.

It runs on the same monthly AI credit as the rest of Bookka’s AI tools — one pool, on the Executive plan, rather than a separate meter per feature. The morning checks are free, because rules are free. You can cap what the assistant specifically is allowed to spend, and it will never top up your balance by itself — an assistant that quietly charges your card while you ask it questions is not one you’d ask many questions.

When the credit runs out it says so plainly and stops. It doesn’t degrade into something worse and let you keep going without mentioning it.

That’s the whole test, really: does it know your actual numbers, does it ask before it acts, can you check it afterwards, and does it tell you things before you think to ask. Everything else is a chat window.

If you’re setting up, the quick start guide covers getting your page live. Reminders that actually get people through the door and Getting found locally cover the two jobs most worth automating before you automate anything clever.