Mock Recall Template: Run a Traceability Exercise With AI Help

Pick one finished batch, then trace it backwards to every ingredient lot and forwards to every customer who received it, and time the whole thing against a target such as two hours. Reconcile the quantities, record each gap, and fix the records. AI helps by reading and matching the paperwork, not by deciding what to do.

The exercise only works if it hurts a little. If every record turns up at once and the numbers add up, you chose an easy batch. Choose one that used several ingredients, shipped to several channels, and ran on a busy day, because that is the one a real recall will land on.

Follow me on Instagram@sagnikteaches

What a mock recall is checking

A mock recall is a rehearsal of the moment you have to answer two questions fast: where did the ingredients in this product come from, and where did the product go? Regulators and retail customers ask these questions in a real event, and at least one food traceability rule will require covered firms, once it is enforced, to hand over an electronic, sortable spreadsheet of the relevant records within 24 hours of a request during an outbreak or recall. You do not need to work to that rule to benefit from practising the same discipline.

Connect on LinkedInSagnik Bhattacharya

It tests four things, each of which can fail on its own.

Subscribe on YouTube@sagnikteaches

• Traceability back: from the finished lot to each ingredient lot and supplier.
• Traceability forward: from the finished lot to each customer, order and quantity.
• Quantity balance: whether made equals sold plus held plus destroyed, plus a small, known loss.
• Speed and people: whether the right person can find the records, and how long it takes.

Quantity balance catches the quiet failures. Records that look tidy can hide two missing pallets, which in a real recall are the two pallets nobody knows to pull.

Choose the batch and set the targets before you start

Pick the batch on the day, using a rule you can defend. Good rules: the batch with the most ingredients made in the last ninety days, or one that contains an allergen-bearing ingredient. Avoid a batch that was a single supplier, a single customer and a single day. Write the choice and the reason at the top of the record.

Then set three targets, written down before the clock starts.

These are starting points. A larger producer, or one working to a retailer's code, may be given tighter numbers, so check what your customers' supplier terms say before settling on yours.

The mock recall template, filled in

Here is the one-page record, completed for an illustrative granola maker with four staff. Copy the headings; replace the entries with your own.

MOCK RECALL RECORD
Date and start time: [date], 09:30
Run by: [name, role]    Observer: [name, role]
Trigger (realistic): a customer reports a plastic fragment in a bag of oat-and-almond granola, lot code shown on the pack as G-2611-B.

BATCH SELECTED: G-2611-B (finished goods)
Reason chosen: five ingredients, three sales channels, made on a day with two other batches.
Made: 1,200 x 400 g bags

TARGETS: trace complete in 120 min; every ingredient lot and customer named; quantities within 2 per cent.

BACKWARDS TRACE (ingredient lots used)
Oats: supplier lot [ ] received [date] qty used [ ] kg
Almonds: supplier lot [ ] received [date] qty used [ ] kg
Honey: supplier lot [ ] ...
Rapeseed oil: supplier lot [ ] ...
Packaging film: roll no. [ ] ...

FORWARDS TRACE (where the 1,200 bags went)
Account / channel | Order or invoice no. | Date | Bags | Delivered or collected
...

QUANTITY BALANCE
Made 1,200 = sold [ ] + in stock [ ] + samples [ ] + destroyed [ ] + unexplained [ ]
Unexplained as a percentage of made: [ ]

FINISH: time [ ]   Elapsed: [ ] min   Pass / fail per target: [ ]
FINDINGS: [numbered list]
ACTIONS: [what, who, due date]
SIGN-OFF: [name, date]

Keep it on one sheet. People complete short forms. The long recall procedure lives elsewhere, and the exercise tests whether anyone can follow it.

Running the trace: step by step

1. Start the clock when the trigger is read out. Record the time on the sheet.
2. Find the production record for lot G-2611-B and list every ingredient lot used. This is the backwards trace.
3. Check each ingredient lot against the goods-in record: supplier, date received, delivery note and the supplier's own lot code.
4. List every dispatch that included the lot. This is the forwards trace and usually the slower half, because dispatch records sit in a different system.
5. Add the quantities and run the balance.
6. Stop the clock when you have a list you would be prepared to send to a customer or inspector.
7. Debrief the same day, while the awkward moments are fresh.

The observer matters. The person running the trace is concentrating; the observer writes down what slowed it, such as the file that had to be fetched from another room or the invoice with no lot code on it. Those notes are the real output.

Where AI helps: reading, matching and gap-spotting

A small producer's records tend to be a mix: a goods-in spreadsheet, photos of delivery notes, invoices exported from accounting software, a production log in a notebook. An AI assistant can turn that mix into one table far faster than you can retype it, and it is good at spotting what does not match. It is poor at knowing what is true. That is why it works on records you give it and never on your memory of what happened.

Replace customer names with account codes before you upload anything, and check that your assistant plan and your customer contracts allow business records to be shared with it. Business plans such as ChatGPT Business or Claude Team do not train on your content by default. Free consumer plans may; check the privacy setting on the plan you use.

Prompt 1: build the backwards trace from goods-in records

Below are (a) the production record for lot G-2611-B and (b) our goods-in log for the last six months.
Task: for every ingredient on the production record, find the matching goods-in row using supplier lot code and date. Return a table with: ingredient, supplier lot, date received, quantity received, quantity used in this batch, and the goods-in row you matched.
Rules: match only on exact lot codes. If a lot code is missing or two rows could match, do not choose one. Put "NO MATCH" or "AMBIGUOUS" and say why. Do not fill any gap from general knowledge.

A realistic output has four rows matched and one marked "AMBIGUOUS", for example two honey deliveries on the same day with no lot code on the delivery note. That ambiguity is the finding. The instruction not to choose is what stops the assistant from handing you a neat but invented answer.

Prompt 2: build the forwards trace and the balance

Here is our dispatch log (columns: date, account code, invoice no., product, lot code, bags) and the production record for lot G-2611-B (1,200 bags).
1. List every row for lot G-2611-B and total the bags by account code.
2. Add the totals and compare them with 1,200. Show the sum.
3. Tell me how many bags are unaccounted for and what percentage of 1,200 that is.
4. List rows where the lot code field is blank but the product and date fall inside the production window, because those may belong to this lot.

The fourth instruction is the useful one. Blank lot codes on dispatch lines are the commonest hole in a small producer's records. A model can flag the rows where the date makes the lot likely, which you then confirm by hand.

A worked run with numbers

Here is how the granola exercise might come out, with all figures illustrative.

The trace started at 09:30. The backwards trace was done by 10:05: four of five ingredients matched cleanly, and the honey was ambiguous because two deliveries arrived on the same day without a lot code on the note. The forwards trace was slower. Dispatch records were in the accounting software, and half the farm-shop orders were taken by phone and recorded on a paper pad. The list was finished at 11:50, which is 140 minutes against a target of 120.

Made 1,200, accounted for 1,174, which leaves 26 bags unexplained: 26 divided by 1,200 is 2.17 per cent, just over the 2 per cent target. So the exercise produces three findings: time over target by 20 minutes, one ambiguous honey lot, and 26 bags unexplained. Each is a concrete job. That is the point of the exercise; there is no "pass" that would have been more useful than this.

Turning findings into actions

List each finding with a cause you can verify, an owner and a date. An assistant can draft the wording, but the cause must come from the observer's notes and the records, not from what sounds reasonable. For the granola run, the list might read as follows.

1. Honey lot ambiguous. Cause: two deliveries on the same day, no lot code on the note. Action: goods-in staff record the lot code from the jar label at receipt; ask the supplier to print it on the note. Owner: production lead. Due: two weeks.
2. 26 bags unexplained. Cause: not yet known. Action: check the pad orders and the sample log for the production week; if still open, recount stock. Owner: owner-manager. Due: one week.
3. Phone orders on paper. Cause: the pad is not entered in the system. Action: enter the order the same day with the lot code, or photograph the pad daily into the shared folder. Owner: dispatch. Due: one week.

Where a finding sits with a supplier, such as the missing lot code, the guide to drafting a supplier corrective action request gives wording that asks for a fix without sounding like an accusation. Rerun a smaller version of the exercise a month later, on a different batch, to check the fixes held.

Making the next trace faster: the codes and the records

Most of the time in a mock recall goes on hunting, not thinking. Three changes cut the hunting.

• One lot code per finished batch, printed on every pack and carried on every dispatch line. If you are still deciding on the format, see designing batch codes for a food producer.
• Goods-in record with the supplier's lot code, not only your own. When a supplier reports a problem, they will quote their code.
• Dispatch lines with the lot code, not only the product. If your accounting software has no field for it, put it in the line description in a fixed format such as "G-2611-B".

A shared spreadsheet with three tabs (goods-in, production, dispatch) and one lookup column per tab is enough for a small producer. An assistant can write the lookup formulas, and you test them by tracing one batch by hand. The guide to AI quality control for small manufacturers covers where a spreadsheet stops being enough.

Variations for different kinds of producer

A bakery with short-life products

A bake that lasts three days has often been eaten by the time the trace finishes, so the useful measure is how quickly you can stop further sales and tell customers, not the paper trail alone. Add a step to your template: "Message sent to every account holding stock, and time sent."

A sauce or drinks producer with contract packing

If a co-packer fills your product, their batch record is part of your trace. Request it as part of the exercise and time how long it takes them to answer. A day's delay at their end counts against your target, because it would in a real event.

A producer selling mostly through marketplaces and online orders

Order exports often lack lot codes. Match orders to the production window by date, then confirm from the pick record. The risk is shipping an older lot on a newer order because stock was picked in the wrong sequence, which a date match cannot see.

Drafting the customer message while the clock runs

A trace that finishes with no one told is only half an exercise. Add a timed step to the template: draft the message that would go to each account holding the lot. An assistant is good at first drafts here, as long as you hand it facts and forbid it from adding any.

Draft a short, plain message to a trade customer who holds stock of our granola, lot G-2611-B.
Facts you may use: product name, lot code, pack size (400 g), that we are withdrawing this lot as a precaution while we investigate, that they should stop selling it and set it aside, and that we will collect it and credit the invoice.
Facts you may not use: any cause, any health claim, any number of complaints, any statement that the product is safe or unsafe.
Leave [your name] and [phone number] as placeholders. Keep it under 120 words.

The restriction list matters more than the facts list. Without it, assistants often add a reassuring sentence such as "no illnesses have been reported", which you cannot know. A person with authority should read the final wording before it is ever used, and your adviser should approve the template itself in advance.

A realistic output is three short paragraphs: what to stop selling, how to identify the lot on the pack, and what happens next. Time how long it takes to turn that into a message to all five accounts, using the account list you built in the trace. If sending takes longer than the trace, the bottleneck is your contact list, not your records.

What a quiet failure looks like in the records

The most dangerous findings do not look like failures. Consider a hypothetical sauce maker whose dispatch sheet showed a tidy 480 jars sold from a 500-jar batch. The balance was 4 per cent out, which the owner assumed was breakage. When the mock recall asked for the breakage log, none existed. A look at the pick sheets showed a customer collection of 20 jars that was taken from the shelf and never invoiced. In a real event those 20 jars would have been the ones with no customer attached.

The way to catch it is the balance, not the trace. A trace follows the paper that exists. A balance asks what the paper should add up to. Ask the assistant to show the sum and to say which of the four buckets (sold, held, destroyed, sampled) is thin on evidence. If "destroyed" is a number with no record behind it, that is the finding.

Before and after, for a single dispatch line, shows how small the fix is.

Before: "10 x granola, farm shop, 14 Nov." There is no lot, so the line cannot be matched to a batch.

After: "10 x granola 400 g, lot G-2611-B, farm shop A, invoice 2204." The line now answers both questions in one glance.

When the exercise exposes something serious

Occasionally a mock recall finds a real problem, such as a lot that is still on sale and cannot be traced to its ingredients. Do not file it away. Treat it as an incident the day you find it, bring in whoever owns food safety, and decide what is needed for the live product. Then record what you did, including the decision not to act where that was the call. If you are unsure whether something triggers an obligation to notify an authority or customer, ask your food safety adviser or the scheme body you certify with, not an assistant.

The same applies to allergens. If the trace shows a lot that may have gone out under the wrong allergen label, check it against the matrix and the labels by hand; the guide to AI food label makers versus manual checks explains why that check should not rest on a generated label alone.

Naming a deputy for every recall role

A recall plan fails less often on records than on people: the owner is away, the person who knows the pad system is on holiday, nobody has the supplier's phone number. Use the exercise to test cover. Run it once with the usual lead, and the next time with the deputy leading and the lead only observing.

In a four-person business one person holds two roles. That is fine, provided the sheet says so and the second run changes who holds them. If the second run is slower by more than a third, the process lives in one person's head, and the fix is a written step list rather than a faster person.

Keeping the evidence

Save the record, the observer's notes, the prompts and outputs, and the corrected spreadsheet in one folder named for the date. Auditors and retail customers ask for evidence that the exercise took place and that findings were closed. A dated sheet with an owner against each action is usually enough. Do not tidy the failures out of the record. A mock recall with no findings looks like one that was never run.

A short list to put on the wall

• Choose the batch on the day, and make it a busy one.
• Write the targets down before starting the clock.
• Strip customer names before uploading anything to an assistant.
• Tell the assistant never to guess at a missing code.
• Check three of its matches by hand, every time.
• Give every finding an owner and a date.
• Run a smaller repeat a month later.

An assistant will save you an afternoon of matching codes. What it cannot do is make the records exist, so the first thing the exercise improves is how you record things on an ordinary Tuesday. That is where the real speed comes from.

Mock recall questions from small producers

How often should a food producer run a mock recall?

At least once a year, and again after a change that touches traceability: a new supplier, new batch codes, new software or a new sales channel. Widely used food safety certification schemes expect a traceability test at least once a year. Check the requirement of any scheme or customer you work to, as either may also set a time limit for completing the trace.

Should I tell staff in advance?

Tell the people running it, but not the people supplying the records. A mock recall that staff have prepared for tests their preparation. Pick the batch on the day and give the team a realistic trigger, as you would in a real event.

Can I upload batch records to an AI assistant?

Only if your plan and your agreements allow it, and with customer names replaced by account codes. Business plans such as ChatGPT Business or Claude Team do not train on business content by default. Check the privacy settings and your customer contracts first.

What counts as a pass?

Set the pass marks before you start. Common ones are a complete trace inside your time target, every ingredient lot and every customer for the chosen batch identified, and quantities that reconcile to within a small, stated margin. Anything short of that is a finding to fix, not a failure to hide.

Does AI replace the recall plan?

No. A recall plan assigns people, decisions and contacts, and someone accountable must own it. AI helps with the dull parts of the exercise: reading records, matching codes and flagging gaps. A person decides what happens to the product.

Further reads

• AI for Food Manufacturing: Six Jobs From Food Safety to Sales — Six jobs in a food business where AI helps.
• AI for HACCP: What a Small Food Producer Can Draft and Must Verify — See what to draft and verify in a food safety plan.
• Allergen Matrix Template: Build From Recipes With AI, Then Verify — Build the allergen matrix a recall decision depends on.
• Food Production Schedule: AI Plans Around Allergen Changeovers — Plan production around allergen changeovers.
• AI Supplier Management: Track Prices, Lead Times, and Risk — Track supplier lead times and risk.
• AI for a Soap Maker: Batch Records, Ingredient Labels and Listings — Keep batch records of a similar small-batch kind.
• Best AI for Welding: Can It Keep Weld Procedure Records in Order? — How a welding shop can use AI to match material certificates to jobs, flag missing documents and build a customer audit pack, with a qualified person signing.
• AI for a Cottage Food Business: Labels, Pricing and Order Forms — Build a recipe-to-label record, price from actual batch yield and collect orders without guessing food rules.
• Foreign Body Complaint: Investigate and Reply as a Food Producer — A step-by-step foreign body complaint routine for a food producer: sample, batch trace, line records, reply letters and escalation triggers, with AI as helper.
• Retailer Supplier Questionnaire: Answer It From Evidence, Using AI — Build an answer bank from your own food-safety documents, let AI match questions to it and draft replies, then link evidence to every answer.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: a food regulator's published page on its traceability final rule (electronic sortable spreadsheet within 24 hours of a request; enforcement not yet started); OpenAI's business data page and Anthropic's plans page (no training on business content by default), all checked November 2026. The exercise itself is illustrative.