What AI Can and Cannot Do for an Independent Café

AI can take much of the desk work off an independent café: drafting the rota, suggesting pastry and milk orders from your till history, writing menu and social copy, and replying to reviews and messages. It cannot make the coffee, recognise your regulars, guarantee an allergen answer, or know what happened on the floor today unless someone tells it.

The dividing line is simple once you see it. AI is good at pattern and wording jobs where you read the result before it matters to anyone. It is poor at anything that needs physical presence, today's live stock, or someone who can be held accountable. Most disappointment comes from owners handing it a job from the second group and expecting the results of the first.

Follow me on Instagram@sagnikteaches

The desk jobs AI handles well in a café

These are the jobs that usually eat the owner's evenings. Each one works with a general chat assistant such as ChatGPT, Claude or Gemini, and none of them needs special café software to try.

Connect on LinkedInSagnik Bhattacharya

In a café that also hires out a back room for birthday teas and baby showers, the messages row is often the biggest time sink: each enquiry needs the date checked, the minimum spend explained and a deposit requested. An assistant can draft all three from a single note of your room rules, leaving you to check the diary and press send. In a café with no bookings at all, that row barely matters, and the ordering and rota rows are where the time is.

Subscribe on YouTube@codingliquids

If messages and bookings are the part that eats your day, the detail is in whether a café can use AI to take bookings and answer messages. For copy, writing menu descriptions that still sound like you covers the voice problem properly.

What four of those jobs look like on a real Tuesday

A table makes every job sound equally easy. In practice each one has its own catch, and the catch is usually visible in the first draft if you know where to look.

The specials board, and the word AI slipped in

You type: "Write three short specials-board lines for a roasted squash and red lentil soup, served with sourdough. Friendly, no puns, under 15 words each." An illustrative reply:

1. Roasted squash and red lentil soup, warming, vegan, with thick-cut sourdough.
2. Golden squash and lentil soup with a hunk of our sourdough.
3. Slow-roasted squash, red lentils and a little chilli. Sourdough on the side.

Line 2 is usable. Line 1 says "vegan", which you never told it: if the soup is finished with butter or made with chicken stock, that word is now a false dietary claim on your board. Line 3 invented chilli. This is the most common failure with menu copy: the model fills gaps with plausible ingredients and labels. The fix is a standing rule in every menu prompt, "Mention only the ingredients I list. Never add dietary labels such as vegan or gluten-free unless I state them", and a read-through by whoever made the dish.

Social captions from the morning's photo

Upload the photo of the cinnamon buns and ask for an Instagram caption. A typical first attempt reads something like: "Freshly baked every hour! 🥐 Our famous cinnamon buns are the perfect treat to brighten your day. Come and grab one before they're gone! #cafe #coffee #cinnamonbuns #foodie #instagood". Three things to fix: you don't bake every hour, they aren't "famous", and that croissant emoji is the wrong pastry. Edited, it becomes "Cinnamon buns out of the oven at 8. Usually gone by 11 on a Saturday." Shorter, true, and it sounds like a person who works there. Give the assistant three of your own past captions as examples and it gets much closer on the first try.

A review reply, before and after

The review: "Lovely cake but waited 25 minutes for two coffees on Saturday. Staff seemed overwhelmed." An unedited AI draft often opens "Dear valued customer, we sincerely apologise for any inconvenience caused and we are committed to providing exceptional service." Nobody talks like that in a café. After you tell it what happened (one barista off sick, a queue out of the door) and ask for a reply under 60 words in your voice, you get something closer to: "Thanks for the kind words about the cake, and sorry about the wait. We were a barista down that Saturday and it showed. We've since added a second person on the machine at weekends. Hope you'll give us another go." Only send the last sentence about the extra barista if it's true.

Supplier invoices: the price rise nobody noticed

Photograph a month of supplier invoices and ask: "List every line item with unit price, and flag any item whose unit price changed between invoices." An illustrative output:

Item                     Invoice 3 Sep   Invoice 24 Sep   Change
Whole milk (per litre)   $1.05           $1.18            +12%
Oat milk (per litre)     $2.40           $2.40            none
Takeaway cups 12oz (50)  $6.90           $7.45            +8%
Espresso beans (1kg)     $24.00          $24.00           none

The milk rise matters: at, say, 40 litres a week it adds about $5 a week, or roughly $270 a year, that nobody would have spotted from the total alone. Check every figure against the paper invoice before acting on it, because photo-reading occasionally misreads a digit. Totals should always be checked against your bank statement.

Six things AI cannot do for a café, however good the tool

These limits don't go away with a more expensive plan. They come from what the software can see and who carries the responsibility.

1. Give a certain allergen answer. A model can only repeat what your allergen records say, and café recipes drift: a supplier swaps a brand of oat milk, a new granola has a nut warning, a baker changes the flour. An assistant can point a customer to your allergen information, but the final answer at the counter must come from a person checking the current record. The reasoning is set out in whether an AI chatbot should answer allergen questions.
2. Know today. It has no idea the combi oven failed at 7am, the bread delivery came up short, or it has rained since breakfast. Every forecast it gives is a guess about a normal day, and you adjust for the day you're having.
3. Recognise regulars and read the room. The person who has the same flat white at 8.10 and wants to talk about their dog is part of what they pay for. No tool replaces the barista who notices.
4. Make sense of vague till data. If half your sales go through a generic "Food" or "Cake" button, no model can tell you how many almond croissants you sold on Tuesdays. The data has to exist before any AI can read it.
5. Take responsibility. Food safety records, temperature logs, cash-ups and staff disputes need a named human who signs off. AI can draft the log template; it can't be the person who checked the fridge.
6. Do the physical work. Obvious, but worth saying because it sets the ceiling. In most cafés the labour cost sits on the floor and behind the bar, and AI touches none of it. The time it gives back is admin time, usually the owner's.

The half-and-half jobs: useful only if you prepare the inputs

Three jobs sit in the middle. AI helps with them, but only after you supply things it cannot find by itself.

Forecasting bakes and waste

A chat assistant can average your sales by weekday and spot that Mondays are consistently quieter than Fridays. It needs at least eight weeks of item-level sales, ideally twelve, and it needs you to mark the odd days: the week the road was closed, a public holiday, the local fun run. Without those notes it will treat a freak Saturday as normal. For a deeper method, how bakeries predict demand and cut unsold stock works through the same maths with bread.

The rota

Given availability and your rules, AI produces a tidy draft in seconds. What it won't know is that one part-timer can't do Saturdays during exams, or that two particular people shouldn't close together. Write those constraints down once in a note you paste every week, and the drafts get far closer. A filled-in version for a small café might read:

ROTA RULES - updated 1 Sep
Open 7.30-16.00 weekdays, 8.00-16.00 Sat, 9.00-15.00 Sun.
Openers (must be trained on the machine and the alarm): owner, Staff A.
Two on the bar 11.30-14.00 Sat and Sun; one is enough weekdays.
Staff B: max 16 hours a week; no Saturdays until exams end 20 June.
Staff C: not before 9.00 (school run). Can close any day.
Nobody works more than 6 days in a row.
Owner works the till Monday and does the order Monday afternoon.

Paste that with the week's availability and ask for a rota table plus a list of any rule it couldn't satisfy. The second part is the useful one: a good draft says "Saturday 11.30 to 14.00 has only one person on the bar because Staff B is unavailable" instead of quietly breaking the rule. The approach for larger teams is in building a staff rota with AI demand forecasts.

Pricing when costs rise

Give it your recipe costs (beans per shot, milk per drink, cup and lid) and it will calculate the margin on every drink and show which ones fell below your target after the last supplier increase. With illustrative figures, a takeaway flat white might cost 18g of beans at $24 a kilo ($0.43), 200ml of milk at $1.18 a litre ($0.24) and a cup and lid ($0.15), so $0.82 in ingredients. Sold at $4.00 that leaves about 80% before labour and rent. Run the same sum across the whole drinks list and the oat-milk drinks usually show the thinnest margin, because plant milk costs roughly twice as much per litre. Ask the assistant to show each calculation line by line, because arithmetic is where chat models still slip. What it cannot tell you is what your customers will pay. That judgement comes from you watching the street, not from a model.

A Sunday-evening test on your own till export

Before paying for anything, run this once. It takes about 30 minutes and tells you whether AI forecasting is worth your time with the data you already have.

1. Export sales by item by day for the last eight to twelve weeks from your till system's reports section. A CSV or spreadsheet is ideal.
2. Strip anything personal. You only need date, item and quantity. Delete customer names, card details and staff names before uploading anything to a chat tool.
3. Write down the odd days in a short list: closures, events, heatwaves, the day the machine broke.
4. Paste this prompt with the file attached:

You are helping an independent café plan next week's bakery order.
The attached file has daily sales by item for the last 10 weeks.
Unusual days to treat with caution: [list dates and why].

1. For each bakery item, give average sales for each weekday,
   excluding the unusual days.
2. Suggest an order quantity per weekday for next week that would
   sell out about an hour before closing on a typical day.
3. List any item whose sales are clearly rising or falling.
4. Tell me which items have too little data to forecast.
Show your working in a table. Do not guess at days not in the file.

5. Compare the suggestion with what you would have ordered. If it matches your gut on most items and disagrees sensibly on two or three, it is worth running weekly. If it disagrees wildly, check the data first: the usual culprit is items rung through a generic button.

What comes back should look roughly like this (illustrative):

Item              Mon  Tue  Wed  Thu  Fri  Sat  Suggested Mon order
Croissant          14   15   18   19   22   31   16
Almond croissant    6    6    8    8   10   15    7
Cinnamon bun        9    8   11   12   14   24   10
Too little data: seasonal fruit tart (only 3 weeks on sale).
Rising: cinnamon bun, up about 20% over the last four weeks.

Two things to check before trusting it. First, whether it really excluded the unusual days you listed; ask it to show which dates it dropped. Second, whether the "suggested order" makes sense against your sell-out target; if you want stock left for the last hour, tell it to add a small buffer rather than accepting a figure that sells out at 2pm.

The numbers add up quickly. Take a 35-seat café with the owner and three part-time staff that orders 60 pastries every weekday and throws away about 15 on Mondays and Tuesdays. The weekday averages from ten weeks of data might show Monday demand closer to 42 and Friday closer to 58. Trimming the Monday and Tuesday order by 15 each saves 30 pastries a week. At a cost of, say, $1.20 each, that is about $36 a week, or roughly $1,800 a year, from one prompt a week. Your numbers will differ; the point is to test with your own.

Some till systems now answer this kind of question directly. Square, for example, has an assistant in its dashboard called Square AI that answers questions about sales, staff and customers. Whether yours has something similar, and which features are switched on for your account, varies, so look in your till's dashboard before buying another tool.

What an independent café would realistically spend

For a single-site café, the honest answer is very little in cash and a few hours of your time.

Keep customer and staff personal data out of free consumer tools, and switch off the setting that lets your chats be used for model training if you use a personal plan for work.

Three beliefs that waste café owners' time

"I need a chatbot on my website first"

Most questions a café gets are opening hours, whether you take bookings, dog-friendliness and parking. A complete Google Business Profile and a clear website answer those without any bot. A chatbot earns its place only when you have real booking volume or a function room to sell.

"AI will run my Instagram"

It will write the caption. It won't take the photo of the morning bake in good light, and the photo is what people stop scrolling for. Treat AI as the writer and yourself as the photographer and editor.

"Forecasting needs expensive software"

For one site with one bakery counter, a spreadsheet export and the prompt above cover most of what paid forecasting tools do. Specialist software starts to make sense with several sites or a large, fast-changing menu, where the weekly copy-and-paste becomes the bottleneck.

A sensible order for trying these

If I were starting in a café tomorrow, I'd do the bake-order test first, because waste is money you can count within a fortnight. Then the rota, because it gives the owner an evening back. Social captions and review replies come third: useful, but the savings are smaller and the risk of sounding generic is higher. Leave anything customer-facing and automatic, such as a chatbot or auto-sent replies, until the first three are running smoothly and you know what your answers should say.

Whatever you try, give it four weeks and write down two numbers before you start: the hours you spend on the job each week and, for ordering, what you throw away. If neither moves after a month, stop paying for it.

Further reads

• Where AI Actually Saves Time in a Small Restaurant — Where the hours actually come back in a small food business.
• Your First 30 Days of AI in a Restaurant, Week by Week — A week-by-week rollout you can adapt to a café.
• What Are Diners Really Saying? Using AI to Analyse Your Reviews — Turn a year of reviews into a short list of fixes.
• How to Forecast Covers and Cut Food Waste With AI — The forecasting method in more depth, with waste figures.
• Should a Small Business Use AI to Write Its Social Media Posts? — Decide how much of your social feed AI should write.
• When Not to Use AI in Your Business: 9 Tasks to Keep Human — The general list of jobs to keep human.
• How Cafés Use AI to Fill Quiet Hours With Targeted Offers — Find your café's quiet hours in the till data, design offers for the people free at those times, and measure whether they brought anyone new in.
• Best AI Tools for Cafés and Coffee Shops, Grouped by Task — The AI tools a café can use, sorted by the job they do across the day, with what each costs and where each one lets you down.
• How to Train Front-of-House Staff to Work Alongside AI — Three short sessions, six role-play cards and a one-page counter card for teaching front-of-house teams to work with AI bookings, chat and phone assistants.
• Can a Small Cafe Use AI to Bring In More Regulars? — The regulars sum for a small café, how to capture customers with permission, four AI jobs that raise visit frequency, and an eight-week plan with numbers.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Square AI press releases and product page (squareup.com); OpenAI and Anthropic plan pages for chat assistant prices.