Export yesterday's item sales and totals, add today's bookings, specials and kitchen notes, and ask an AI for a one-page briefing: yesterday in numbers, what sold out, today's expected covers, specials with allergens and one service focus. Toast IQ can draft a daily briefing inside Toast; any other POS needs an export and a chat assistant.
The draft is only as honest as its inputs. A POS knows what sold, not why, and it knows nothing about allergens, the chef's special or who called in sick. So the AI writes the notes and you supply the facts it cannot see. Treat anything it adds by itself, such as a reason for a slow night or an allergen line, as a mistake until you can match it to a record.
What a five-minute briefing needs, and where each line comes from
Before any tool, decide what the page must contain. I would build the briefing from seven blocks, each with an owner and a source. If a block has no source, drop it rather than let the AI fill the gap.
| Block | Source of the facts | Time to say it |
|---|---|---|
| Yesterday in numbers: covers, net sales, spend per cover against the same weekday | POS sales summary | 30 seconds |
| What ran out, and roughly when | Kitchen sold-out list | 30 seconds |
| Today's expected covers and shape of the evening | Bookings plus past same-weekday figures | 45 seconds |
| Specials and features, with allergens | Chef, and the allergen file | 60 seconds |
| Large parties, celebrations, access needs | Reservation notes | 45 seconds |
| One service focus, such as a dish to sell early | Manager | 30 seconds |
| People: sections, who is new, who earned thanks | Manager | 30 seconds |
That adds up to about five minutes spoken. Only three blocks come from the POS. The rest come from people, which is why the briefing cannot be fully automated and should not pretend to be.
Two routes: ask Toast IQ, or export the POS and paste
If you run Toast, start with its assistant. Toast's Toast IQ page lists a request to "create a short, easy-to-read daily briefing for my staff" among its examples, says it answers questions about sales, labour and menu performance, and states that Toast IQ is included with Toast POS at no additional cost. It can also mark items out of stock, with your confirmation. It works from the data inside Toast. Anything that lives elsewhere, such as bookings in another system, the allergen file or the chef's special, still has to be typed in, so ask it what it can see before you trust it with the forecast. My tutorial on what Toast IQ can and cannot tell you lists where the edges are.
For every other POS, or when you want control of the wording, use an export plus a chat assistant such as ChatGPT or Claude:
- Toast: Reports, Menus, Product mix. Toast's guide to the product mix report says it exports to Excel or CSV and can compare two periods, but it cannot be emailed on a schedule, so the export is a manual step of a couple of minutes.
- Square: Reports, Custom, Custom item report, which shows quantity sold, gross sales and net sales, with an Export button for CSV or Excel.
- Anything else: an end-of-day item report. You need item name and quantity at minimum, net sales if you can get it, and the day's covers and net sales from the summary.
Whichever route you take, the AI is a writer, and you remain the editor.
Tidy the export before you paste it
A raw item export is messier than the report on screen. Three kinds of clutter cause most wrong numbers. The same dish can appear twice because it sits on both a dinner menu and a takeaway menu. Modifier rows, such as "add extra bread", sit in the same list as dishes. And voids, comps or zero-quantity lines can pad the file. If you ask an AI for "the top five dishes" from a file like that, it will rank a dish's dine-in rows and takeaway rows separately and understate the dish. Here is a small illustration of the mess and the fix.
| Raw row | Qty | Net sales | Becomes |
|---|---|---|---|
| Burrata starter (Dinner) | 29 | $333.50 | Burrata starter: 34 plates, $391.00 |
| Burrata starter (Takeaway) | 5 | $57.50 | Joins the row above |
| Pappardelle | 27 | $580.50 | Pappardelle with slow-cooked beef: 29 plates, $623.50 |
| Pappardelle (Takeaway) | 2 | $43.00 | Joins the row above |
| Add extra bread | 11 | $33.00 | Extras list: 11, $33.00 |
Toast has a setting for part of this: the product mix report's "aggregate data by name" option combines items that share a name across menus. For everything else, let the AI propose the merge, but make it show its working so you can check the sums.
Below is an item sales export. Merge rows that are the same dish on different menus (for example "(Dinner)" and "(Takeaway)").
List modifiers and extras (rows starting "Add" or "No") separately and do not count them as dishes.
Ignore rows with quantity 0. Do not change any number. Show the merge as a table with the original rows beside the merged row,
then give me the top ten dishes by plates and by net sales.Check one merged row by hand: 29 plus 5 is 34 plates, and $333.50 plus $57.50 is $391.00. If the AI cannot show the original rows beside each merge, it is guessing. Delete employee columns at this stage too; a briefing does not need them.
One Saturday briefing built from Friday's figures
The example is a 60-seat restaurant serving grilled meats, pasta and a short fish list. All numbers here are illustrative. Friday's export and summary give you this, set against the average of the last four Fridays:
| Measure | Friday | Last four Fridays (average) | Difference |
|---|---|---|---|
| Covers | 138 | 148 | -6.8% |
| Net sales | $8,946 | $9,615 | -7.0% |
| Spend per cover | $64.83 | $64.97 | -0.2% |
The pattern is worth saying aloud to the team: a quieter room, not a weaker room, since spend per head held. The item export shows the top plates: burrata starter 34, ribeye 31, pappardelle with slow-cooked beef 29, sea bass 24, chicken under a brick 22. The kitchen notes say the sea bass went off at 8:10 pm and the pistachio cheesecake at 9:20 pm. For today, 96 covers were booked by 2 pm, there is a party of 12 at 8:00 on a set menu with a deposit paid, and a birthday table of six at 7:30. The chef's special is pork belly with apple puree, 14 portions, and two servers are new on the floor.
Write a one-page pre-shift briefing for restaurant staff from the facts below and nothing else.
If something isn't given, leave it out or write MISSING. Do not explain why sales were up or down.
Do not write allergen information; copy the line I give you. Sections: Yesterday, Today, Specials, Focus, Team.
Short lines, plain words. Show the covers estimate two ways (same-weekday average and booking pace) and give a range.
Yesterday (Friday): 138 covers, net sales $8,946. Last four Fridays: covers 149, 152, 131, 160; net sales 9,710, 9,980, 8,420, 10,350.
Item sales (plates): [paste the export: item and quantity]
Ran out: sea bass (last plate 8:10 pm), pistachio cheesecake (9:20 pm).
Today (Saturday): 96 covers booked at 2 pm. Last four Saturdays at 2 pm: 88, 91, 99, 83. Their final covers: 171, 166, 182, 158.
Party of 12 at 8:00 (set menu, deposit paid). Table of 6 at 7:30, birthday.
Special: pork belly with apple puree, 14 portions. Allergens: [copy from the allergen file].
Focus (manager): offer the pork belly special at every table before 8 pm.
Team: two servers new on the floor, each paired with a section lead.A clean answer looks something like this (illustrative):
PRE-SHIFT: SATURDAY
YESTERDAY (FRIDAY)
- 138 covers, $8,946 net, $64.83 a cover. Last four Fridays averaged 148 covers, $9,615 and $64.97 a cover. Quieter room, same spend per head.
- Most plates: burrata 34, ribeye 31, pappardelle 29, sea bass 24, chicken under a brick 22.
- Ran out: sea bass (8:10 pm), pistachio cheesecake (9:20 pm).
TODAY (SATURDAY)
- Booked at 2 pm: 96 covers. This time on the last four Saturdays: 83 to 99.
- Estimate: about 180 covers, range 175 to 187 (booking pace). Same-weekday average alone: 169.
- Party of 12 at 8:00, set menu, deposit paid. Birthday table of 6 at 7:30.
SPECIALS
- Pork belly with apple puree, 14 portions. Allergens: [copied from allergen file].
FOCUS
- Offer the pork belly special at every table before 8 pm (14 portions).
TEAM
- Two servers new on the floor tonight, each paired with a section lead.Checking the draft before it reaches the floor
Run five checks every time. They take two minutes and catch the errors most likely to embarrass you on the floor.
- Three numbers against the POS. Covers, net sales and one item count. If one is off, assume the rest are suspect and recheck them all.
- Dish names and prices as they appear on tonight's menu. AI tidies names, and a guest who orders "the burrata" may be ordering something slightly different from the line on the page.
- Allergens against the allergen file. Not against the AI, not against memory.
- Strike every reason you did not supply. "Quieter because of the rain" is a story, not a fact.
- Read it as a server would. Anything that blames a person or a shift gets rewritten.
That clean version is usually a second pass. A first pass from a looser prompt tends to carry errors like these. It may call the ribeye "our best seller", which is true by revenue ($1,116) but not by plates: the burrata starter sold 34 to the ribeye's 31. Ask for "most plates" and "most revenue" as separate lines. It may add "sales were down because Friday was wet", which you never gave it. Delete it. It may say the sea bass "sold out around 8:00" when your note said 8:10. Give times exactly as supplied. And it may give only the same-weekday average for Saturday, 169, ignoring that bookings are running ahead of normal; asking for two methods in the prompt fixes that. If a draft needs more than one round of fixes, tighten the prompt instead of editing the output again. A line like "use only the facts below" removes most of the invention.
Time and cost are modest. Export, paste and prompt take about three minutes, and the five checks about two, so five minutes a day once the prompt is saved. A free plan on one of the main assistants may handle an export of a hundred lines, though file-analysis allowances on free plans are limited and change, and paid individual plans run from roughly $5 to $20 a month at list prices. Check the current allowance on the plan you use. If you run Toast, the base Toast IQ assistant adds nothing to the bill; Toast also names Toast IQ Team and Toast IQ Grow packages without listing prices on that page, so ask before assuming a feature sits in the free part.
Covers forecast: three simple methods and a range
The forecast is the line staff care about most, because it drives how they set up and how quickly they are expected to turn tables. You do not need a model. Three short sums on figures you already have give a range, and the AI can show its working so you can check it.
| Method | Working | Covers |
|---|---|---|
| Same-weekday average | Final covers on the last four Saturdays: 171, 166, 182, 158 | 169 |
| Booking pace | 169 x (96 booked today / 90 average booked at 2 pm) | 180 |
| Typical finish from bookings | 96 x 1.88 (average of final covers divided by covers booked at 2 pm) | 180 |
| Range from the ratios | 96 x lowest ratio (1.824) to 96 x highest (1.943) | 175 to 187 |
On those four Saturdays the covers added after 2 pm (later bookings and walk-ins, less no-shows) were 83, 75, 83 and 75, so the room has been adding 75 to 83 covers after the afternoon. That is a more useful sentence for a floor manager than a single number: "booked 96, usually finishes between 175 and 187". Treat everything as an estimate and say so on the page. Adjust by hand for what the history cannot know, such as a street event or a closed competitor, and refresh at 5 pm if bookings move. If you have begun taking deposits or sending confirmation texts, your booking ratios will change, so start a fresh baseline; the no-shows tutorial covers how those rules shift the numbers. For a fuller forecasting method, see forecasting covers and cutting food waste.
A quiet Tuesday: the same page for a different kind of night
Saturday's briefing is about pace. Tuesday's is about using the space. In this example the last four Tuesdays finished with 58, 64, 61 and 66 covers (an average of 62), with 21, 24, 22 and 27 booked by 2 pm (an average of 23.5). Today only 15 are booked. Booking pace gives 62.25 x (15 / 23.5), about 40 covers, and the ratio method gives 37 to 42. That is well below the average of 62, and the covers added after 2 pm, 37, 40, 39 and 39 on those Tuesdays, are the number to watch. With so few bookings at 2 pm, treat a swing of a few late reservations as normal; the range is a floor, not a promise.
A slow evening needs a different block on the page. Give the AI a list of jobs you would like done in a lull and ask it to place them:
Add a QUIET-NIGHT block to the briefing. Estimated covers today: about 40 (usual Tuesday: 62).
Choose three tasks from this list and write each as a role, not a person: deep-clean walk-in shelves,
rotate dry goods, polish glassware, update the allergen folder for the new special, walk the new servers through the POS.
Do not suggest changing anyone's hours.The last line matters. Offered a slow night, a model can easily recommend sending a server home, and a staffing change is a manager's decision, taken before the meeting, with its own rules. If a draft says "consider sending one server home", delete it. Staffing against a forecast is its own job, and not one for a briefing prompt. The assistant's useful contribution here is the list of jobs and who does them, which is the part managers forget on a quiet night.
Sold-out dishes and specials: what the AI should and should not do
A sold-out list is the most actionable thing on the page, and also where AI overreaches. What it should do: report what ran out and when, and, if your export has timestamps, show plates per hour for those dishes so the chef can see how early the shortage began. What it should not do: decide tonight's prep numbers. The sea bass sold 24 and ran out at 8:10, which says demand was at least 24 and probably more; the chef knows delivery size, yield and whether a fish special is competing, so the chef sets the number. Put it as a question on the page: "Sea bass sold out at 8:10 on Friday. How many tonight?" That keeps the decision where it belongs. If you want the par levels behind that decision, see inventory par levels with AI.
For specials, treat the chef's note as the source and the allergen file as the authority. Paste the allergen line exactly, and write the prompt so the AI cannot alter it: "Allergens: copy this line unchanged." After the draft arrives, search the page for the dish name and confirm the allergen text sits next to it. If the special is new, the allergen file is updated before the meeting, not after.
What stays off the page
- Server-by-server sales rankings. A public league table can reward discounting and demotivate the people who covered the slow section. Talk to individuals privately, with numbers.
- Anything about named guests beyond first names and the need. "Table of 6 at 7:30, birthday" is enough. Medical or access information should be limited to what staff need to act, and kept out of any tool you have not checked.
- HR matters, performance comments and complaints. They belong in a private conversation.
- Guesses. If it is not in your sources, it is not on the page.
The page also has to be easy to read at a glance: short lines, no paragraphs. If the AI writes a paragraph, ask for bullets.
Closing the loop on yesterday's focus
A briefing that never looks back becomes wallpaper. Add one line to tomorrow's input: yesterday's focus and what happened to it. Keep it factual, written by the manager in two short lines, and let the AI carry it into the next page. For example: "Focus: mention the pistachio cheesecake before 8:30. Result: 22 sold, ran out at 9:20." On its own that is a data point. After a few weeks of the same kind of focus line you can see which nudges move plates and which are ignored, and you stop repeating the ones that do nothing. Do not ask the AI to decide whether the focus "worked" from a single night, because it cannot separate your nudge from the weather or a large party. Ask it only to list the last five focus lines and their results side by side, and judge them yourself.
Monday's longer look back, and keeping the routine alive
Once the daily page works, add a weekly version for Monday that looks back over seven days against the previous four weeks. The prompt changes in three ways: you paste weekly item sales instead of one day; you ask for the five dishes with the biggest change in share of sales, not raw plates; and you add any price changes or menu changes from your log so the AI does not read a price rise as a loss of appetite. If you run price tests, testing menu price changes by dish shows how to keep that log.
A useful Monday table is share of mains, not plates, so a busy week does not flatter every dish. An illustrative version, last week against the four weeks before it:
| Main | Share, previous four weeks | Share, last week | Change |
|---|---|---|---|
| Grilled sea bass | 7.4% | 8.9% | +1.5 points |
| Pork belly special | not on | 6.1% | new |
| Chicken under a brick | 9.1% | 7.9% | -1.2 points |
| Pappardelle with slow-cooked beef | 14.0% | 13.1% | -0.9 points |
| Ribeye, 10 oz | 12.6% | 12.4% | -0.2 points |
Read the chicken line carefully. Its share fell 1.2 points in the same week a special arrived at 6.1%, which suggests the special took some of its guests. That is a "why" the AI cannot see and you can, so write it in a notes column and let the AI repeat it, rather than letting it guess.
The routine itself matters more than the prompt. I would keep the prompt in a Project in ChatGPT or Claude, which both store instructions and reference files, and share it so any manager runs it the same way: ChatGPT allows project sharing on every plan, while Claude shares projects only on its Team and Enterprise plans. Set a time: export at 2 pm, paste, two minutes of checking, print or post by 3. Name a backup for days the usual manager is off. Review the page monthly with the team and delete any block nobody uses. A briefing that grows each week soon runs past ten minutes, and the habit dies the first time service is busy. For turning the same kind of daily list into standing opening and closing routines, the restaurant SOP template tutorial picks up where this one stops.
Pre-shift briefing questions from restaurant managers
Can Toast IQ write the pre-shift briefing for me?
Toast's own page for Toast IQ shows a request for a short daily briefing for staff, and says Toast IQ is included with Toast POS at no additional cost. It works from data held in Toast, so bookings, allergens and specials kept elsewhere still need typing in, and you should check its figures against your reports.
How long should a pre-shift meeting last?
I would aim for five to ten minutes, short enough that people stay on their feet and listen. If the briefing runs long, trim the numbers first and keep what changes what staff do tonight: sold-out dishes, large parties, specials and the one service focus. Longer team conversations belong in a weekly meeting.
Is it safe to paste POS data into ChatGPT or Claude?
Item sales hold no guest details, so the risk is low. Reservation notes are different: leave out surnames, phone numbers and health details, and use first names or booking IDs. Business plans from the main vendors do not train on your content by default, and consumer plans have a model-training switch in privacy settings.
What if my POS cannot export item sales for one day?
Most systems can print or export an end-of-day item report. If yours only prints, key the top ten items by hand, which takes about two minutes and is more reliable than asking an AI to read a photo of a receipt roll. Weekly exports still help for the Monday look back.
Further reads
- How to Build a Restaurant Staff Rota With AI Demand Forecasts — Staff against a covers forecast with a demand-based rota.
- AI for Restaurant Order Management: One Queue for Every Channel — Bring takeaway and delivery orders into one queue for the shift.
- What Are Diners Really Saying? Using AI to Analyse Your Reviews — Add one line of guest feedback to the weekly briefing.
- How to Build a Kitchen Training Manual With AI — Turn recurring briefing reminders into training pages.
- ChatGPT Prompts for Restaurants: 10 for Menus, Rotas and Replies — Ten ready prompts for menus, rotas and replies.
- Can AI Do a Restaurant's Bookkeeping? What a Person Still Checks — What a person still checks when AI handles the books.
- How to Use ChatGPT for Restaurants: Five Jobs to Hand It Each Week — A Monday-to-Sunday routine: ordering, prep and specials, rota, weekend promotion and the weekly review, with allergen answers kept human-checked.
- Restaurant Labour Cost Percentage: Test Rota Changes With AI First — A what-if rota test from last month's takings and payroll: baseline the percentage, cost the change and find the break-even sales loss, with AI doing the sums.
- Hotel Shift Handover Notes With ChatGPT: From Desk Logs to a Daily Report — A one-line-per-event desk log, anonymising rules, two ChatGPT prompts and a worked Tuesday that shows what the AI gets wrong and how to catch it.
- Restaurant Waitlist Software: Where AI Helps at the Door — AI can quote waits by party size, text when a table is ready and answer waitlist calls. Here is what it still needs a host at the door to do.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Toast IQ product page; Toast support (Product Mix report overview); Square support pages (item sales reports). The restaurant, figures and briefing in the worked example are illustrative.