Your First 30 Days of AI in a Restaurant, Week by Week

A restaurant's first 30 days with AI should cover one tool and one person: week 1 sets up a chat assistant and a baseline, week 2 puts it on reviews and guest messages, week 3 adds one numbers job such as a covers forecast, and week 4 scores each job keep, fix or stop. Phones and allergen answers wait.

The reason for the narrow start is service. Restaurants that try five AI projects at once end up testing them at 7pm on a Friday, and the first bad answer in front of a guest kills the whole thing. Thirty minutes a day on quiet afternoons, by one named person, gets further than a big launch.

Follow me on Instagram@sagnikteaches

Before day one: an owner, one tool and a baseline

Pick the person who will do the 30 minutes a day, and put the time in the diary like a supplier meeting: in the bistro, 3pm to 3.30pm on Tuesday, Wednesday and Thursday, between lunch clear-down and the evening briefing. If the slot keeps getting eaten by service, the project will be too, so choose a time that genuinely survives a busy week. Then pick one assistant. For two users, ChatGPT Business and Claude Team both cost $25 a seat a month on monthly billing ($20 on annual), with a two-seat minimum, and neither trains on business content by default. If you already run Google Workspace or Microsoft 365, check what AI your plan includes before buying anything new; month one needs only one tool.

Connect on LinkedInSagnik Bhattacharya

Then record a baseline, because at day 30 you will want to know what changed. Spend the week before day one tallying these, roughly, as a filled-in example from an illustrative 45-cover neighbourhood bistro shows:

Subscribe on YouTube@codingliquids

Guesses are acceptable where you have no log; write down how you guessed so you can compare fairly later.

Week 1 (days 1 to 7): set up and a house brief

1. Day 1: create the workspace, add the two users, and check the data settings. Agree a rule: no card numbers, guest phone numbers or staff personal details go into the assistant.
2. Day 2: write a house brief, a one-page description the assistant reads before every job (below).
3. Days 3 to 7: use it for the easiest writing jobs, the specials board, a menu-change note to staff, a reply to a supplier. Save the prompts that worked.

Expect week one to feel slower than doing things yourself. The first few drafts need heavy editing because the assistant doesn't know your restaurant yet, and the house brief is how it learns. Each time you correct the same thing twice (the wrong sign-off, too many adjectives, calling the terrace a "patio"), add a line to the brief. By day 7 the brief is usually twice as long as on day 2, and the drafts need half the editing. Keep the saved prompts in one shared document with a short name for each, such as "Review reply, 4-5 stars" or "Supplier chase", so the second user can find them.

The specials board is a good first test because the errors are easy to see. Asked for "tonight's specials from these notes: hake, brown shrimp butter, samphire, 24; pork belly, apple, black pudding, 22", the assistant's illustrative first draft read "Succulent pan-seared hake, bathed in a luxurious brown shrimp butter and paired with vibrant samphire, $24". Three corrections came out of it and went straight into the brief: no "succulent", "luxurious" or "vibrant"; describe the dish in the order it sits on the plate; prices as plain numbers, no dollar sign, as they appear on the printed menu. The next day's draft was "Hake, brown shrimp butter, samphire. 24", which is what the board needed all along. Short is harder for an assistant than long, so say so in the brief.

The supplier chase is the other easy early job, and it shows the tone problem in the opposite direction. A delivery arrived 6 kg short on lamb shoulder. The first draft apologised twice for "any inconvenience" and asked whether the supplier might "possibly be able to look into it". The version that went out: "This morning's delivery was 6 kg short on lamb shoulder: invoiced 18 kg, received 12 kg, delivery note attached. Please credit the 6 kg or confirm it's coming tomorrow before 10am, as it's on Friday's menu." Facts, the number, what you want and by when. Add "for suppliers: direct, factual, no apologies" to the brief.

A house brief, filled in for the bistro:

We are a 45-cover neighbourhood bistro, open Wednesday to Sunday,
lunch and dinner. Seasonal menu that changes every three weeks.
Tone: warm, plain, a little dry. Never gushing. No exclamation marks
in review replies. Sign off as "[first name] and the team".
Facts: bookings up to 8 online; larger groups by email. Private room
seats 14. Dogs welcome at the bar only. Card only, no cash.
Never state allergens or dietary claims unless I paste them from the
kitchen's allergen sheet. Never promise refunds or free meals.

The allergen line is there because of a mistake that happens early and often. In one illustrative first week, the assistant described a new risotto as "naturally vegan" on a draft specials post because the dish sounded plant-based; the kitchen finishes it with butter and parmesan. The house brief rule and a habit of pasting the dish's ingredients from the kitchen sheet stopped it recurring. For menu writing in more depth, see writing menu descriptions with AI that still sound like you.

Week 2 (days 8 to 14): reviews and guest messages

This is usually the fastest visible win, because unanswered reviews and slow replies to group enquiries are both easy to measure. Set aside 20 minutes, three afternoons a week, to clear reviews and enquiries with the assistant drafting and a person sending.

Using our house brief, draft a reply to this review. Thank them for
something specific they mentioned. If they raise a problem, acknowledge
it without arguing and invite them to email [address]. Under 80 words.

Review (3 stars): "Lovely lamb, but we waited 40 minutes for mains
and nobody told us why. Shame, as the starters were great."

What came back (illustrative):

Thank you for coming in, and I'm glad the lamb and starters hit the
mark. You're right that a 40-minute wait for mains is too long, and we
should have told you what was happening. That's on us. If you'd like
to tell me more, email [address] and I'll make sure your next visit
goes better. [first name] and the team

A good draft, with one edit. "I'll make sure your next visit goes better" edges towards a promise of compensation some owners won't want to make publicly. Changing it to "I'd love to hear more" keeps the warmth without committing you. For the harder cases, answering restaurant complaints with AI without escalating has worked examples. One kind of review should never go through the usual prompt: anything alleging illness, a foreign object or an allergic reaction. A draft that sounds sympathetic ("we're so sorry our food made you unwell") reads publicly as an admission before anyone has checked what was served. Hold the public reply, have the manager check the booking, the dish and the kitchen records that day, and contact the guest privately first. When you do reply publicly, keep it short and factual: that you've been in touch directly and take reports like this seriously. Group enquiries follow the same pattern: paste the enquiry, the private-room facts and your set-menu options, and ask for a reply that answers every question the guest asked, in the order they asked it.

An illustrative case shows the change. Before: an email asking about a 40th birthday for 12, with questions about the private room, a set menu, a cake and whether a guest's coeliac disease could be catered for, sat for a day and got a two-line reply that answered the first question and suggested a call. After: a draft within ten minutes of opening the email that confirmed the room seats 14, attached the two set menus with prices, said guests can bring their own cake for a small plating charge, and, for the coeliac question, said the chef would reply personally with the options, because the house brief bars the assistant from dietary claims. The guest booked the same day. The draft took the manager four minutes to check and send.

Week 3 (days 15 to 21): one numbers job

Pick one job that uses your own figures. For most restaurants that is a covers forecast feeding the prep list, because waste is easy to count and the saving is visible in the bins. Export the last eight weeks of bookings and covers by day and service from your reservation system, plus this week's bookings so far, and ask for a forecast by service with the reasoning shown.

An excerpt from the bistro's first run (illustrative):

Fri dinner: forecast 58 covers (booked 41; last 8 Fridays averaged 17
walk-ins on top of bookings made by Thursday).
Sat dinner: forecast 62 covers (booked 49; average 13 walk-ins).
Suggested prep, braised short rib: Fri 14 portions, Sat 15.

The Friday figure was reasonable. Saturday was wrong in a way only a person would catch: 14 of those 49 booked covers were one private party on a set menu without the short rib, so the prep suggestion was four or five portions too high. Add a column to your export for private or set-menu bookings, and tell the assistant to exclude them from à la carte prep. If you take few bookings and most covers walk in, use your till system's covers or transactions by service instead of reservations; the forecast is rougher but still better than habit. Whichever data you use, keep the kitchen in the loop from the start: show the chef the forecast and the reasoning on Thursday, and let them overrule it. A forecast the kitchen ignores saves nothing, and the chef's reasons for overruling it are often the missing column in your export. The full method, including waste logging, is in forecasting covers and cutting food waste with AI.

Week 4 (days 22 to 30): the day-30 scorecard

Repeat the baseline tally in week 4 and put the two side by side. Each job gets one of three verdicts: keep (it saved time or money and nobody had to fix its mistakes publicly), fix (useful, but a rule or input needs changing), or stop (the checking takes as long as doing it yourself).

Borderline verdicts are common. If a job saved time but needed one embarrassing correction, it is a "fix", not a "keep": work out which rule would have prevented the correction and add it before month two. If you can't agree a verdict, run the job for two more weeks and count again. What you should avoid is the silent drift where a tool stays on the bill because nobody decided.

The staff-notice "stop" came from a specific slip rather than a hunch. In week three, a rota message drafted from "Sat: open 5.30, [first name] in for 5" came back as "Saturday: doors open at 5pm, please arrive by 4.30". Nobody spotted the change, and one of the team turned up an hour early. A two-line message typed by hand takes less time than reading an AI version closely enough to catch that, which is exactly what the "stop" verdict means: checking costs more than doing.

A quick sum puts the keeps in perspective. Across reviews, menu copy and notices, the bistro's weekly writing time fell from 160 minutes to 100, an hour a week back for the owner. The short rib waste fell by 12 portions a week; at an illustrative $6 food cost a portion, that's $72 a week, or roughly $310 a month, against about $50 a month for two seats. The hour is the saving the owner feels; the waste is the one that pays the bill.

These are illustrative numbers, but the shape is typical: writing jobs pay back fastest, short messages don't benefit, and the numbers job works once its inputs are cleaned up. The "stop" is as useful as the "keeps". It tells you where not to spend month two.

Month two then builds on what worked: bring a second person onto the review and enquiry jobs, fix the numbers job's inputs, and choose one new job from the list below only if the first three are running without the owner's constant attention.

Kept out of the first month, on purpose

• AI answering the phone. Worth considering, but it touches every caller and needs a connection to your booking system and careful testing. See whether AI can take restaurant reservations by phone before planning it for month two or three.
• A chatbot answering allergen questions. The stakes are too high for a first-month experiment; whether a chatbot should answer allergen questions sets out the case carefully.
• Auto-posting review replies. Drafts yes, automatic publishing no, until you have read a few hundred drafts and trust the pattern.
• Dynamic pricing or menu re-engineering. These need months of clean data, not four weeks.

If you want the longer arc these fit into, an AI implementation plan for a small restaurant lays out a full 90 days.

The Monday twenty-minute huddle

Keep momentum with a short weekly check, before Wednesday's first service in the bistro's case. The agenda fits on a sticky note:

1. Which prompt saved the most time this week? Save it to the shared folder.
2. Which draft needed the biggest correction, and what rule would prevent it?
3. Anything that went to a guest that shouldn't have?
4. What's this week's one job?

The notes from the bistro's fourth huddle (illustrative) fit in four lines: the group-enquiry prompt saved the most time, about 40 minutes across three enquiries; the biggest correction was the rota message with the wrong start time, so rota messages come off the list; nothing wrong reached a guest, though one specials draft called a dish "gluten-free", which the owner caught on the read-through because the brief makes any dietary word a red flag; this week's job is the day-30 scorecard. The whole huddle took twelve minutes, and every line changed something.

Question 3 matters most. A single public mistake, a wrong allergen claim or a reply that promises a refund, costs more goodwill than the month's time saving earns. Catching near-misses in the huddle, and turning each one into a line in the house brief, is how the first 30 days produce a restaurant that trusts its AI rather than one that quietly stops using it.

Questions from restaurant owners starting out

Should the chef or the front-of-house manager own the first month?

Whoever runs the admin that takes most time off the floor, and who will actually sit down for 30 minutes on quiet afternoons. In many small restaurants that is the front-of-house manager, because reviews, enquiries and rotas land on them. The chef should still own anything that describes food, since menu facts and allergens have to come from the kitchen.

Is a free chat assistant plan good enough for the first 30 days?

For testing prompts, yes, but switch before you paste anything about guests or staff. Consumer plans let you turn off model training in privacy settings, while business plans such as ChatGPT Business or Claude Team exclude business content from training by default and add shared workspaces. With two people using it, a business plan costs about $50 a month on monthly billing.

What if staff don't want to use it?

Don't make them in month one. Have the named owner use it, share the drafts that saved time at the Monday huddle, and let others ask for access. Forcing it on a busy team during service is how pilots stall. The scorecard at day 30 gives you evidence to show them, which persuades better than a mandate.

Further reads

• How Much Does AI Cost a Small Restaurant Each Month? — What the monthly bill looks like once you add more than one tool.
• Restaurant AI Tools: 12 Questions to Ask Before You Sign Up — Twelve questions to put to any restaurant AI vendor in month two.
• How to Train Front-of-House Staff to Work Alongside AI — Bringing front-of-house staff in once the first month has worked.
• How to Review an AI Tool After 90 Days: Keep, Fix or Cancel — The keep, fix or cancel review to run at day 90.
• 7 AI Mistakes Restaurant Owners Make With Bookings and Reviews — Seven booking and review mistakes to steer clear of.
• How to Build a Kitchen Training Manual With AI — A good month-two job for the kitchen once the basics work.
• What Are Diners Really Saying? Using AI to Analyse Your Reviews — A method for turning months of restaurant reviews into a few counted, checked themes, with a codebook, a tagging prompt and a worked nine-month example.
• How to Set Up an AI Phone Line for Takeaway Orders — Seven steps from counting your calls to a tested AI phone line that takes takeaway orders into your till, with the three routes compared.
• Where AI Actually Saves Time in a Small Restaurant — A task-by-task look at a small restaurant's admin week: where AI cuts real minutes, where it only moves them, and where it adds work.
• What AI Can and Cannot Do for an Independent Café — The desk jobs AI does well in a café, the ones it can't touch, and a 30-minute test on your own till data before you pay for anything.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI and Anthropic pricing pages for ChatGPT Business and Claude Team (checked September 2026); vendor privacy documentation on business-plan training defaults.