AI Workshops for Cafes and Restaurants: What AI Should Draft in Your Venue, and Where It Stops

Train on the writing your venue repeats, and keep allergen wording out of it. Review replies, function and catering enquiries, supplier chasers, roster messages and social posts are all fair game, because you already own those facts and you approve the send. Anything that states what is in a dish is a different kind of job. Book the full-day Sunshine Coast workshop, or private onsite training, for whoever already writes those messages. In most venues that's the owner or the manager, not someone rostered on the floor.

TL;DR

• The two repeated writing jobs that pay back fastest in hospitality are the public review reply and the function or catering enquiry. Both are slow, both carry commercial weight, and both follow a pattern that already lives in the owner's head.
• Allergen, dietary and food-safety wording is the hard boundary. Australian rules require allergen declarations for unlabelled cafe food using specific required names, and "gluten free" has a legal meaning, so a drafting workflow should flag and stop rather than answer. 1 2
• Most published advice about AI in restaurants describes software you buy: voice ordering, table management, inventory forecasting. Training is about the writing layer, which is the part you can change this month without a new system.
• A full day away from the floor is the real price of this training, and it only pays back for someone who owns the policy, the prices and the approvals.
• We teach a team to build and run the workflow themselves, and we also build systems for clients. Those are two different engagements with different scope and price.

Which writing does a cafe or restaurant actually repeat?

Five jobs, in most venues. None of them are cooking.

The first is the public review reply. It's slow because it's public, it usually gets written at the worst possible moment, and a flat apology reply reads worse than silence to the next person scrolling.

The second is the function and catering enquiry. Head count, date, minimum spend, room hire, deposit terms, what happens with dietary requirements, whether the back room can take 30. That reply is a small quote document, and it decides whether a four-figure booking lands.

Third, supplier and ordering messages. Chasing a short delivery, asking for a credit, pushing back on a price rise, changing a standing order before a long weekend. Plain commercial writing that nobody enjoys doing at 3pm between services.

Fourth, roster and team communication. A shift swap, a written explanation of a roster change, the same onboarding note you retype for every new starter. Written instructions carry more weight in a venue than in most businesses, because the person reading them may have started this week. Pay rates and entitlements sit outside this, though. Those are questions for your accountant or the Fair Work Ombudsman, not a drafting job.

Fifth, menu wording and social posts. Specials board lines, a Christmas trading notice, an event post, a rewrite of a tired dish description. This is the one that comes with a catch, and it's the next section.

Every one of those has the same shape: the facts already exist somewhere in your venue, and the writing is the bottleneck. That's exactly the shape an AI workflow built around five inputs handles well.

Allergen and food-safety wording is where the drafting stops

This is the boundary that matters most on this page, so it gets stated plainly. A drafting workflow never states, confirms, implies or amends what is in a dish.

Two reasons. The first is legal. In Australia, for food that isn't required to carry a label, which covers what you serve across the pass, the Food Standards Code requires allergen declarations to be "displayed in connection with the food or provided to the purchaser upon request using the required names". 1 Those required names are specific. Individual tree nuts, molluscs and individual cereals must all be declared separately, so "nuts" isn't the required name and "cashew" is. 1 A "gluten free" claim is tighter again: under the Food Standards Code it means no detectable gluten, and a food containing oats or oat products can't be labelled gluten free in Australia at all. 2 Those are the kinds of words a fluent drafting tool adds without being asked, because they read well.

The second reason is mechanical, and it's the one I'd argue is more important. A model works from the source material you gave it. Your kitchen changed the bread supplier on Tuesday. The chef swapped the crumb on the schnitzel because the delivery was short. Nothing about that reaches the document your workflow is reading, so a draft that sounds confident about allergens is confident about last month's recipe.

Australian food service already names a qualified human for this. Cafes, restaurants, caterers and takeaway shops handling unpackaged, potentially hazardous ready-to-eat food have to implement food safety management tools, and a food safety supervisor must hold recognised, formal certification obtained in the past five years. 3 The regulator's answer to "who is accountable for what's in the food" is a certified person, not a document and not a draft.

So the rule you write into the workflow, in the Rules input, reads roughly like this:

• Never state, confirm or change allergen content, dietary suitability, or ingredient provenance such as organic, free range or locally sourced.
• Where the incoming message asks about allergens, intolerances or a medical dietary requirement, write a flag line and stop. Do not attempt the reply.
• Where a review or enquiry mentions illness, an allergic reaction, injury or a foreign object, flag it to the owner immediately and draft nothing.

That third rule is the one people forget. A one-star review alleging someone got sick is not a review reply task. It's an incident, and it's the owner's phone call.

What you're doing here is deciding in advance which outputs a workflow is allowed to produce. I do the same thing in my own work. This article was researched and written through a workflow that reads a defined set of business sources and returns a draft, a brief and the sources behind it, and then stops, because publishing isn't in its scope. The boundary is what makes the rest of it safe to run quickly. Where a whole team is drafting, that boundary belongs in a written team AI use policy rather than in one person's memory.

Menu descriptions are still workable, by the way. Write the texture, the cooking method, the size, the flavour. Leave every dietary and allergen line to come from the kitchen's own allergen matrix, unchanged.

How do you turn a review reply into a workflow you can run twice?

You write five inputs, then test them on a real review you've already answered, so you can compare the draft against something you know was right.

Say it's a Sunday and a two-star review lands: cold coffee, a 25 minute wait on eggs, and it names a staff member by first name. Here's how that job breaks up.

• Context. Who the venue is, which service period this was, and who the reply is really for. That last part changes the writing. You're not writing to the reviewer. You're writing to the next person reading your reviews before they book.
• Task. Draft one public reply, 40 to 70 words, plus a one-line internal note for the owner.
• Source. Your service recovery policy, your actual trading hours and staffing for that date, the owner's name and direct contact, and three replies you've written before that you were happy with. Those three examples do more for tone than any instruction about tone.
• Rules. Never name a team member in a public reply. Never dispute the customer's account of the wait. Never offer a specific refund or comp amount. Plus the food-safety flags from the previous section.
• Output. One paragraph, plain sentences, signed by a named person, under 70 words, followed by the internal note.

The internal note is the part I'd push you to keep. The public reply is a few dozen words of courtesy; the note is where the workflow says what it saw, so the owner reads "third slow-service review this month, all Sunday breakfast" before approving anything. A month of replies written that way turns your review page into something that tells you what to change, instead of a chore that gets cleared. Checking the draft against your own source material is a skill in itself, and there's a method for checking AI output that we work through in the session.

Function enquiries run the same way, with a different source pack: your function menus, minimum spend, deposit and cancellation terms, room capacities. The rules do the heavy lifting again. Never quote a price that isn't in the source pack. Flag a missing head count or date instead of assuming one. Route every dietary question to the kitchen with a named person attached. That's the same structure we use for AI enquiry management in other industries, and the closest thing I've built to it myself is the enquiry and admin automation we did for Glen House Homoeopathy, where the facts came from the practice and the workflow only handled the writing and the routing.

How is AI being used in restaurants right now?

Mostly as software somebody sells you, rather than as writing you do. That's the honest read of the coverage.

The guides I read for this article describe AI phone answering that takes bookings when nobody can reach the handset, review summarisation across platforms, demand forecasting for rosters, inventory tools that predict when you'll run short, table management algorithms and dynamic pricing. 4 5 Point-of-sale vendors have been adding the same features, including menu generation and marketing copy inside the till system you already pay for. 6 One vendor puts adoption of AI in some form at 65 percent of Australian operators, which is that company's own industry data rather than an independent measurement. 7

Two things follow from that. One, a decent amount of this is a purchasing decision, not a training decision, and it lands whenever your booking platform or POS ships it. Two, the writing layer is the part nobody sells you, and it's the part you can change this month with the tools you already have open.

There's a gap in that coverage that I'd point out plainly. Three of the pages I read list menu copy or menu writing as an AI job, including the competing training offer. 5 6 8 Not one of the five mentions the allergen declaration rules that sit underneath those words. That's not a small omission in a country where the declaration requirement applies to the food you hand across a counter. 1

The full day off the floor decides who should attend

Hospitality runs on thin margins and very little admin time, so I'd rather be straight about the cost of this. A full day is half a shift. For a small venue that's a genuine hole in the week, and it needs to earn its place.

It earns it for an owner, a venue manager or a functions coordinator. Those people already carry the prices, the policies, the tone and the authority to approve a reply. They can build a workflow in the session and use it on Monday.

It doesn't earn it for a floor staff member. Not because they can't learn it, but because the workflow's value sits in its Source and Rules, and a floor staff member doesn't own either. They can't decide what the deposit terms are or who gets flagged when a review mentions illness. Send them and you get an interested person with no authority to finish the job.

Two sensible patterns. A single venue sends the owner or manager to the public workshop. A group with multiple sites, or a venue where three or four people write to customers, runs a private session onsite and builds the shared source pack together, which is most of the value anyway. If you're an accommodation business with a restaurant attached, the same logic applies across both sides of the operation, and our AI training for accommodation businesses covers the booking-side work in more detail. Once more than one person is drafting, you'll also want the shared facts written down properly, which is what building a staff knowledge base for AI is for.

Should you take a public seat or run the session in your venue?

Take the public seat if one person does this writing. Bring us onsite if the writing is spread across a team, or if you'd rather not have your manager off-site during a trading day.

The public workshop date is Wednesday 7 October 2026, and you can join the workshop registration to hear when it's set. For a team session we come to you and scope it first, and private onsite AI training is available across the Sunshine Coast, Brisbane, the Gold Coast, Ipswich, Noosa, Moreton Bay, Logan, Redlands, Toowoomba and Gympie. Team size, session length and travel determine the quote.

Bring one real task and one safe example of it. A real review you've already answered, or a function enquiry from last month with the customer's details removed. The session is a full day on that task, and you leave with the written workflow and the tested example.

What if you'd rather have it built than learn it?

Then say so when you enquire, because we do both.

Training is the entry rung. You or your team learn to build the workflow, run it and check it, and you own it afterwards. Implementation is the managed rung, where the system gets built and connected for you. They're two different engagements with different scope and different pricing, and which one suits you depends on how much of this you want to run yourself.

For most single-site venues I'd start with training, honestly. The writing jobs in a cafe are close enough to the owner's own judgement that learning to direct them is faster than specifying them for someone else to build. Once you're running several sites, or once the social and marketing output becomes a standing commitment rather than a Tuesday job, the calculation changes, and a built social content workflow starts to make more sense than a manual one. Either way, the conversation starts with what you repeat and who currently does it.

I should be straight about one thing. I haven't run a venue. My implementation work has been in other industries: internal processes at Radiant Roof Repairs, where I previously worked before leaving to run Searchmaxxed and Sunshine Coast AI Workshops; blog creation, SEO and custom proposals in my own business; and SEO, enquiries and admin for Glen House Homoeopathy. What transfers is the method and the judgement about where a workflow's boundaries belong, which is why this page spends more words on the allergen line than on prompt wording. What doesn't transfer is knowing your kitchen. That part you bring.

Frequently asked questions