AI in Food Retail: Jobs a Deli or Farm Shop Can Hand Over First

Start with jobs that are repetitive, already written down somewhere and checked by a person before they reach a customer: end-of-day markdown suggestions, the weekly order forecast, a weekly note about what has arrived and allergen label drafts. Leave freshness checks, use-by decisions, allergy conversations and producer relationships to staff.

The time saved is real but modest, and most of it comes from drafting, not deciding. In the illustrative deli used here, six jobs return about six hours a week, roughly one part-timer's shift, and none of them removes a human check. The order matters: begin where a mistake costs the least and the numbers already exist.

Follow me on Instagram@sagnikteaches

Six jobs, with the hours saved and the check that stays

Consider a deli and farm shop with five part-time staff, about 400 product lines, a cheese and cooked-meat counter, a weekly veg box and a busy Saturday. The hours are an owner's estimates for that shop, not measurements, so time your own jobs for two weeks before you believe them.

Connect on LinkedInSagnik Bhattacharya

At $20 an hour, 5.7 hours is about $114 a week, or roughly $5,900 a year. The bigger prize, if the forecast and markdown jobs work, is less stock written off, which the table does not count. The jobs do not carry equal risk, so the order I would hand them over in is markdown suggestions first, because they need only today's stock list; then the order forecast, which needs tidy sales data; then the newsletter; and last allergen labels, because the matrix takes the longest to build and the stakes are highest. The two small jobs at the bottom of the table, and the hamper replies covered further down, slot into any quiet week.

Subscribe on YouTube@sagnikteaches

Order forecasting: the weekly cheese order, worked through

A deli's weekly order is often a judgement call made from memory. The input for a better one already exists in the till: units sold per line per week. Pull eight to twelve weeks, merge lines that were renamed, and choose one unit per product, because cheese sold loose by weight and the same cheese sold in pre-packed rounds will not add up unless you convert one to the other. Square's custom item report, for example, lists a renamed item as a separate entry, so a rename in the spring splits one line in two.

For one line, a soft-rind cheese in 200 g rounds that arrives in cases of six with a 21-day shelf life, the last eight weeks look like this.

Here are eight weeks of unit sales for a soft-rind cheese (200 g rounds). Orders
come in cases of 6 and shelf life is 21 days. Week 7 included our Saturday food
fair, so treat it as an event week. Give me: the average for the normal weeks,
the range, a suggested order for a normal week and for an event week (whole
cases, rounded up), and the arithmetic. Tell me if any suggestion would leave
more than a week of spare stock.
Sales: 14, 16, 15, 19, 14, 17, 31, 15

The arithmetic the assistant should show: the seven normal weeks sum to 110, an average of 15.7 rounds, with a range of 14 to 19. A normal-week order of 18 rounds, three cases, covers every ordinary week but week 4. An event week of 31 needs six cases, 36 rounds, because five cases fall one short. A round costs $3.20 and sells for $7.50, so running out loses $4.30 of margin while a round that expires loses $3.20. The two costs are close, which argues for a small cushion over the average, not a large one.

Check three things in the reply. The averages: recompute them in your sheet, because models can slip on sums. The rounding: assistants sometimes round to the nearest case instead of up, so confirm each figure is a whole number of cases. And the event flag: if week 7 goes in as normal the average becomes 17.6, which with a cushion is enough to tip a normal-week order from three cases to four. Then apply what the data cannot know: the weather forecast, a school holiday, a wedding order already in the book.

Compare it with habit. An owner who orders four cases, 24 rounds, every week leaves about eight spare in a normal week. If even six of those expire, that is $19.20 a week on this one line, about $1,000 a year. If twenty lines behaved the same way the figure would run to five figures, but lines differ, so measure rather than multiply: record actual sales and leftovers each week and the sheet becomes its own back-test. The tutorial on forecasting stock and reordering for small shops covers the wider routine.

Forecast the lines where a surplus costs the most: short shelf life, high unit cost, bought in cases. In a deli that means cheese, cooked and cured meats, fresh dairy, ready meals and bread. Ambient lines such as tins and jars rarely justify the effort, because a minimum-and-maximum rule in the till does the job. Seasonal and weather-led lines, such as salad crops or barbecue food, have too little history for a forecast to mean much, so use last year's notes, the weather outlook and your own judgement. Begin with the ten perishable lines that cost you most to write off, which a glance at the bin will name.

A farm shop that buys from neighbouring growers works to a different rhythm, because it takes what is harvested. There the forecast is really a selling plan: how much of each item to promote, put in boxes and price this week. Give the assistant the growers' harvest list and last year's sales for the same weeks, and ask for a draft box plan and a list of what needs promoting. The call on what is good enough to sell, and on what to do with a glut, stays with the person who has looked at the crates.

Once the owner has settled the quantities, the assistant can turn them into each supplier's order in the format that supplier wants, whether that is an email, a message or rows for a portal.

Supplier: [cheese supplier], delivery Thursday
- Soft-rind round 200 g: 3 cases (18 rounds)
- Blue wedge 150 g: 2 cases (8 wedges)
- Mature hard cheese, 2 kg block: 1
Please confirm the delivery window and any substitutions before dispatch.

The assistant drafts and you press send. Keep your own copy of the approved quantities, because if a supplier disputes an order, your signed-off list is the record.

If your numbers live in Google Sheets, Gemini's =AI() function, on an eligible Workspace or Google AI plan, can draft text columns, but it generates only the first 350 selected cells in a run and refreshes only when you tell it to, so keep calculations in ordinary formulas. Excel's =COPILOT() function was retired on 14 September 2026; use the Copilot side pane or a chat assistant for the explanation instead.

Allergen labels: AI formats and flags, a named person decides

House-made items such as quiches, salads, cakes, dips and jams need a label with the ingredients and the allergens emphasised, and every recipe or supplier change can alter it. The format and the list of allergens you must declare depend on where you trade, so check the rules that apply to your shop with whoever inspects your premises. What follows is the workflow, not the rulebook.

What AI does well here is clerical. It reads each supplier's specification sheet and your recipe, copies every ingredient's own ingredient list and allergen statement into a master matrix, drafts the label text, compares a new draft with the last signed one and lists what changed. What it must not do is decide that a product is free from an allergen, choose "may contain" wording (that is a risk judgement about cross-contact in your own kitchen), or print anything without a signature.

Here is the kind of change a good comparison catches. A house-made granola uses oats, honey, sunflower oil, almonds and pumpkin seeds. The cook swaps in milk chocolate chips from a new supplier. The reply should read something like this (illustrative):

CHANGES SINCE LAST SIGNED LABEL (granola, 15 Oct):
- New ingredient: milk chocolate chips (supplier B).
- MILK is now present (from the chocolate chips). It was not on the last label.
- Supplier B's specification sheet states "may contain peanuts". Almonds and
  seeds already declared. Precautionary wording is a decision for [named
  person]; I have not changed it.
- Spec sheet for the oats is older than 12 months: please check it is current.
DRAFT LABEL: Granola. Ingredients: OATS, honey, sunflower oil, ALMONDS,
milk chocolate chips (sugar, cocoa butter, whole MILK powder, cocoa mass),
pumpkin seeds.
STATUS: DRAFT, not for printing until signed.

The named person reads the new spec sheet themselves, decides on precautionary wording, signs and dates the label, and the matrix is updated. Building the matrix is the real work: about 15 minutes a product the first time, so 30 products is seven or eight hours, and after that each change takes minutes. Counter staff answer customer questions from the signed matrix, never from memory or from an assistant. The wider arguments are in AI for delis: specials, platters and allergen labels and building an allergen matrix from recipes.

A sign-off routine that fits on one card: the label owner opens the new specification sheet, checks the draft against it line by line, decides any precautionary wording, initials and dates the master copy, and files the sheet. Print labels in small batches rather than months at a time, so an old recipe cannot live on in a drawer of pre-printed stickers, and keep the previous version for your records. Stamp every draft "DRAFT" until it is signed, so nobody prints from the wrong file on a busy morning.

Markdown prompts for the last two hours of trading

Around three o'clock someone walks the counter and the fridges and decides what to reduce. The decisions follow a pattern you can write down, which is why an assistant can draft the plan. The inputs are a stock list with quantities, dates, cost and normal price, plus your rules.

Shop closes at 5:30 pm today and reopens at 8 am tomorrow. Stock list below. Rules:
use-by today = must sell or be removed by close; best-before items may be
reduced; never go below cost plus 10%; never more than 50% off; first markdown
3 pm, second 4:30 pm; bundle where it helps. Give: the markdown for each item at
each step, shelf-edge wording under 8 words, and a 25-word message for our
regulars' group. List items staff must check by eye.
Quiche slice | use-by today | 3 left | cost $1.90 | price $4.20
Sourdough loaf | best-before tomorrow | 4 left | cost $1.60 | price $4.50
Soft cheese wedge | use-by tomorrow | 5 left | cost $3.10 | price $6.40

A typical reply (illustrative) reduces the quiche by 25% at 3 pm ($3.15) and by 50% at 4:30 pm ($2.10, just above the $2.09 floor of cost plus 10%); reduces the sourdough by 25% at 4:30 pm ($3.38); and puts the soft cheese at 20% off all afternoon ($5.12). The shelf-edge wording might be "Today's bargain: 50% off". Check the first draft for the usual slips. Assistants sometimes treat "use-by tomorrow" as urgent when only today's date forces the second step. They can slip below the floor on a cheap item, so check every price against the cost column. And they cannot see or smell anything, so the by-eye list is a reminder, not a safety check. Never reduce an item past its use-by date, and follow your own food safety plan.

Keep a simple log to see whether it works. An illustrative one reads: write-offs of $212, $185, $240 and $198 in the four weeks before, an average of $209, and $150, $138, $162 and $144 in the four weeks after, an average of $149. A drop of $60 a week would be worth about $3,100 a year, but only the shop's own log can say, and a holiday week or a new supplier can swing four weeks either way. The message to regulars goes only to people who have opted in to hear from the shop. Where the discount applies to short-dated stock across the whole shop, AI pricing for short-date markdowns goes into the arithmetic and the risks.

A weekly newsletter from the people who grow and make your stock

A short note on what has arrived this week, one producer in the spotlight and the opening hours is the easiest of the six jobs, because the owner sees every word before it goes out. The inputs are your delivery notes and two or three lines of notes from each producer. The assistant writes the first draft, and the rule that matters is "use only these facts".

Keep one sourcing sheet with a row per producer: what they supply, how far away they are, how the product is made, and the claims they have agreed you may make. The newsletter, the shelf cards and the label spec folder can all draw on it, so it is one sheet doing three jobs. It is also where you settle what "local" means for your shop. The word has no fixed definition, so write your own into the prompt, for example "call a product local only if its row gives a distance under 30 miles", and the assistant will apply it the same way every week.

Write a 200-word newsletter for our deli and farm shop customers from the notes
below. Use only facts in the notes. Do not use the words organic, free-range,
award-winning, local or fresh unless they appear in the notes. Friendly, plain
English, no exclamation marks. End with opening hours: [hours].
Notes:
- Eggs from [Producer A]: barn-kept hens, collected daily, delivered Tuesday.
- Chutney from [Producer B]: new batch of apple and onion, made last week.
- Cheese: soft-rind round back in stock after two weeks.

A first draft (illustrative) often comes back like this.

Fresh this week: free-range eggs from [Producer A], collected daily and on the
shelf from Tuesday. A new batch of apple and onion chutney from [Producer B],
made last week, sits beside the cheese counter, and our soft-rind round is back
after a fortnight away. Pop in and ask what goes with what.
Opening hours: [hours]

The draft broke two of the three rules: "fresh", which the prompt banned, and "free-range", where the notes said barn-kept hens. An upgraded claim is the typical failure. "Free-range" has a legal definition in many places, and a customer who cares will notice. Compare every adjective with the notes. Producers should see and approve the lines about them, and ask before you use their names or photos. Send through a proper email tool that handles unsubscribes, to people who have opted in; as of October 2026, Mailchimp's free plan allows 250 contacts and 500 sends a month, which is enough for a first list. The guide to writing a customer newsletter with AI in 30 minutes covers the layout.

Three smaller jobs worth handing over

Supplier invoice price check

Paste this week's invoice lines beside the last three invoices and ask for any line that moved by more than 5%. A typical flag: "Mature hard cheese, 2 kg: $9.80 per kg on the last three invoices, $10.60 on this one (+8.2%)." The owner queries the supplier or accepts the rise, and then changes the shelf price if the margin needs it. The assistant does not decide; it finds the line you would have missed on a Friday afternoon.

Shelf-talkers and tasting notes

A staff member's tasting note is the raw material: "nutty, a bit sweet, salty crystals, good with chutney." The assistant tidies it into a card: "Nutty and lightly sweet, with crunchy salt crystals. Try it with chutney." The slip to look for is invention. If the draft says "aged 24 months" and nobody gave that fact, strike it. Health or nutrition claims do not belong on a shelf card an assistant wrote.

Pre-order and hamper enquiries

For Christmas hampers and weekly boxes, an assistant can draft the acknowledgement and the answers to routine questions about collection times and deposits from a FAQ you supply. Anything about contents, substitutions or dietary needs goes to a person. A customer asking for a hamper without nuts needs a member of staff to confirm the contents from the signed matrix. The tutorial on farm shop boxes, labels and pre-orders goes deeper on the boxes themselves.

Jobs that stay with staff

• Judging freshness and quality. Looking, smelling and tasting. No export or photo replaces it.
• Use-by decisions. The date on the pack is a hard stop. A tool can remind staff that a date is near; it does not decide what is safe to sell.
• Allergy conversations. Staff answer from the signed matrix, and if the matrix does not cover it, the answer is "I can't be sure", not a guess.
• Ingredient changes. Anyone who swaps a supplier or an ingredient tells the label owner the same day. The assistant can compare documents only after someone supplies the new one.
• Producer relationships. What you say to the farmer about price, quantity or a bad batch is yours to say.
• Ripening and portioning. When a cheese is ready, and how a large piece is cut, rests on skills and the product in front of you.
• Prices across the range. An assistant can model margins. Deciding what a loaf costs against the shop down the road is a judgement about your customers.
• Temperature logs and cleaning records. Someone has to take the reading and sign it.

Six weeks, one new job at a time

1. Weeks 1 and 2. Markdown suggestions for the end of each day. Log what the assistant suggested against what staff did.
2. Weeks 2 and 3. Order forecast on eight weeks of history, run beside your normal order without replacing it. Compare it with what sold.
3. Week 4. The weekly newsletter, with producers approving their own lines.
4. Weeks 5 and 6. Start the allergen matrix with the five best-selling house-made lines, and add the rest after the sign-off routine has run for a fortnight.
5. Whenever there is a quiet week. Invoice price check, shelf-talkers and the hamper replies.

What it all costs is mostly your time. A chat assistant plan at about $20 a month list price, the spreadsheet you already use and an email tool's free tier cover all six jobs. Buy nothing else until a job has beaten the old way for four weeks, and treat any forecasting or label add-on as a decision to make with your own numbers in hand.

Four numbers show whether it is working. Write-offs in dollars each week, against the four weeks before you started. Stock-outs on your top twenty lines. Hours per job, timed. And the number of label corrections caught at sign-off, where a high count means the check is working, not failing. If a job does not beat the old way after four weeks, stop using it and keep the habit of logging, which is useful on its own.

Questions deli and farm shop owners ask

Can AI write allergen labels for a deli?

It can draft label text and a master allergen matrix from supplier specification sheets, and flag what changed since the last version. It must not decide that a product is free of an allergen or choose 'may contain' wording. A named person signs off every label against the specs, and staff answer customers from the signed matrix.

Which job should a farm shop hand to AI first?

Markdown suggestions or the weekly order forecast. Both use data you already hold, a person reviews the output before anything changes, and a mistake costs little. Leave allergen labels until the matrix exists, because building it is most of the work.

Do AI forecasts work for seasonal stock?

They help most for lines that sell steadily and least for seasonal or weather-driven ones with little history. Mark event weeks, include a few seasons if you have them, and treat the output as a starting figure that you adjust for the weather and the calendar.

Is it safe to put sales and supplier data into a chat assistant?

Item-level sales, stock lists and supplier specification sheets hold no customer data, so the risk is mainly commercial. Use a business plan or switch model training off, and keep customer names, emails and phone numbers out of chats unless your tool and privacy notice cover them.

Further reads

• AI Tools for a Grocery Store: Supplier Invoices and Shelf Prices — Supplier invoices and shelf prices in a larger grocery.
• AI in Hospitality Management: Catch Supplier Price Creep on Invoices — Catching quiet supplier price rises on invoices.
• How Bakeries Can Use AI to Predict Demand and Cut Unsold Stock — Forecasting daily demand for stock that goes off fast.
• Health Food Store Marketing Ideas: AI Posts Without Health Claims — Writing shop posts without making claims you cannot back.
• Email Newsletter QA Checklist Before You Hit Send — A final check before any newsletter goes out.
• AI for Butchers: Christmas Pre-Orders, Cut Lists and Pick-Up Slots — Pre-orders and cut lists for a counter-led food shop.
• AI Agents in Retail: What an Independent Shop Can Actually Run — Nine shop jobs split between an AI agent and staff, a line-by-line reorder run, and where Shopify Sidekick and Lightspeed AI fit in a one-till shop.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Square help pages on item sales reports and Google's help pages on Gemini in Sheets, as checked in October 2026; Microsoft Support notice on the retired Excel COPILOT function; Mailchimp's free-plan limits.