Where AI actually helps a catering business, and where it does not

There is a lot of noise about AI in hospitality software right now, most of it written by people who have never had to get 180 plates out of a marquee. So here is the version from the other side: what it genuinely does, what it genuinely does not, and what it costs.

Six things it does well today

1. Turning an inquiry email into a draft proposal

This is the strongest use by a distance. Around eighty percent of inquiries are the same shape — a date, a headcount, a rough style, a budget hint buried in a sentence — and pulling those four things out and producing a first-draft proposal is exactly the kind of pattern-matching that works reliably.

It does not replace your judgement about the client. It removes the twenty minutes of typing between reading the email and having something to react to, which means the reply goes out the same day instead of Thursday. In this trade, replying first is a real advantage.

2. Reading supplier documents

A produce invoice as a photographed PDF, and you want the line items and prices in your system. This is tedious, error-prone, universally hated, and something a model does well. Same for a venue's requirements sheet, a client's dietary spreadsheet, or a rental quote.

Check the totals. Extraction is good, not perfect, and an ingredient price that lands with a decimal in the wrong place quietly poisons every recipe that uses it.

3. The morning brief

Everything happening today, what changed since yesterday, what is overdue, what needs a decision — assembled and written as a paragraph you read in thirty seconds instead of four screens you check in sequence. Not clever. Genuinely useful every single day, which is a better property than clever.

4. Drafting the awkward emails

The deposit chase. The 'we need your final count by Friday' nudge. The follow-up to the tasting that went well and then went quiet. These do not get sent, not because they are hard to write, but because they are unpleasant to write. Something that produces a polite first draft in your own tone removes the friction that was actually stopping you.

5. Answering questions about your own data

'Which events next month still have no deposit?' 'Did the Anderson payment land?' 'What did we charge the Harris wedding per head?' Questions you can answer with three clicks, asked in one sentence instead. The value is not the answer; it is that you ask questions you would not have bothered to click through for.

6. Structuring the mess

A menu typed into a text message. Handwritten tasting notes. A client's forwarded email chain with the guest count in the fourth reply. Turning unstructured mess into structured records is dull, constant work, and it is where the hours quietly go.

Four things it should not be doing

1. Deciding your prices

It can do the arithmetic. It cannot know that this client is worth taking at a thin margin because their company books four more events a year, or that this Saturday is worth holding out for. Pricing is a judgement about relationships and risk. Let it compute; do not let it decide.

2. Anything allergen-related, unsupervised

This is the hard line. A model that is right 99% of the time about whether a dish contains nuts is a model that will eventually be wrong about a dish containing nuts, and that is not a category of error a catering business survives. Dietary and allergen information gets entered by a person, checked by a person, and carried to the table by a person who knows whose plate it is.

Use it to flag and to prompt — 'this menu contains an allergen listed on the guest sheet' — never to certify.

3. Writing your contracts

It will produce something that reads like a contract. Whether the cancellation schedule is enforceable where you operate is a question for a lawyer in your jurisdiction, once, cheaply. Do not find out during a dispute that your terms were drafted by autocomplete.

4. Sending anything to a client on its own

Draft, always. Send, never. This is not a technical limitation you should expect to disappear — it is how you keep your voice, and it is the only reliable defence against the confidently-wrong sentence going out under your name.

What it costs to run

The part almost nobody publishes. AI features cost real money per use, and 'unlimited AI' either means the limits are hidden or somebody is planning to change the price.

Costs are charged per token — roughly, per chunk of text in and out. A drafted proposal from an inquiry email might run a couple of cents. A document extraction, a few cents more. A daily brief, a fraction of a cent. For a typical catering business using it across the working day, actual monthly usage tends to land in the low tens of dollars rather than the hundreds.

How we do it, since it is the honest answer to the question: $50 of AI usage is included in the $425/month plan, which covers normal use for most operators. Past that it is billed at cost — the model provider's own token price, no markup — and you can see the running total and set your own cap. We would rather show you the meter than pretend the electricity is free.

Three rules that make it safe

1. 1Draft and retrieve, decide and send. The machine writes and finds. The human chooses and presses send. Every failure story I have heard violates this line.
2. 2Everything it does is logged. If an AI touched a record, you should be able to see what it did, when, and who approved it. In a business with staff and clients, an unlogged action is an unaccountable one.
3. 3It sees exactly what the person using it sees. If a server cannot open payroll, neither can the assistant when that server is asking. Anything less and you have built a very polite way around your own permissions.

The honest summary

AI will not run your catering business and anybody telling you otherwise has something to sell that will not work. What it will do is delete a large amount of typing, reading and retrieving — probably the least valuable hours in your week — and hand them back to you.

That is what Maître is inside Apron Desk: an assistant that drafts, extracts, briefs and answers, that logs everything it does, that inherits the permissions of whoever is asking, and that never sends anything to a client on its own. Included in the plan, metered honestly, and firmly on the draft side of the line.