Algeria AgencyArtificial intelligence
AI in food production: from the printed lot to the question asked on a Thursday
The lot number is printed, and nobody knows where it went. Traceability is not settled at printing, it is settled at packing.
A cannery in the Mitidja, a Thursday afternoon. A distributor calls: a customer has reported a problem with a jar, and gives a lot number and a date. The question is simple — where did that lot go — and the answer takes two people, three binders and the rest of the day.
The lot was printed correctly all the same. What is missing is not the label, it is the trace of what became of it: the lot was packed, palletised, loaded and delivered without anybody recording the link between those steps. An artificial intelligence system can close that gap, and this article says how far exactly.
It does not repeat what your buyer actually reads on a shelf, nor the packaging order and the printing proof, which are covered elsewhere. It starts after them: the code is on the pack, and the question is what is done with it.
The question that arrives on a Thursday
The test of all this material is one question you will be asked one day, usually at a bad moment: which customers received a given lot, and how much of it is still with them. It comes from a distributor, from an end customer, sometimes from an inspection.
A business that takes more than an hour to answer does not have a computing problem: it has a recording-chain problem. The number exists, it is printed, it is even legible on the jar — but it was never tied to anything between the packing line and the lorry.
The time to answer is also what decides the scale of a withdrawal. When you do not know where a lot went, you withdraw everything it could be; when you do, you withdraw what is concerned. The difference is counted in pallets and in commercial relationships.
So this article starts from that question and works back: what should have been recorded, and when, to answer within the hour. The answer is three links, and none of them is difficult.
A lot is not a number, it is a boundary
A lot bounds a set of products made under conditions considered homogeneous. Its size is a decision, and it is the first one here: a lot per production day, per shift, per vat, or per raw material received.
The choice is a direct trade-off. A wide lot simplifies recording and makes every withdrawal massive; a narrow one multiplies the entries and makes withdrawal surgical. Nobody can settle it for you, because it depends on your volumes, your vats and what you can keep up without error.
The one mistake that cannot be recovered is a lot whose boundary corresponds to nothing real — a number that changes at midnight while production continues, or that stays the same when the raw material has changed. At that point traceability is an orderly fiction.
So before any tool, write the lot rule in one sentence and check that somebody on the line can repeat it without hesitating. If two people give two different rules, no system will make up for that. The place the rule usually breaks is the shift handover: if the evening team changes raw material without changing the number, two lots are mixed in the record and nothing shows it.
What a machine reads on a pack, and what it misses
Automatic reading of a lot code works, with precise limits worth knowing before buying. A code printed flat, dark on light, on a paper label, reads very well from an ordinary photograph.
What resists is predictable: inkjet printing on flexible film, which bleeds and distorts; curved surfaces, where part of the code is always out of focus; reflective metal packaging; and laser-etched codes with weak contrast. Those are precisely the commonest substrates in food production.
The consequence is not to give up but to put the reading in the right place. Photographing a pallet in a storage area under changing light gives a failure rate that makes the system unusable; photographing at packing, from a fixed station under constant light, gives the opposite.
Always ask for a demonstration on your own packaging, including the one you know is difficult. A reading done on a label printed for the occasion predicts nothing.
Record at packing, never afterwards
The principle is the one running through this file: information is captured where it still exists. The link between a lot and a pallet is obvious on the packing line, and becomes an investigation two days later.
Concretely, a pallet is given an identity as it is built, and that identity carries the lot or lots it holds. It is a few seconds of work — a photograph of the code and a confirming gesture — and it is the only one that cannot be caught up later.
The temptation is to reconstruct afterwards from quantities: that day we produced so much, that delivery was so much, therefore it held that lot. The reasoning works as long as there is one lot a day, and it collapses exactly on the day you need it most, which is when two lots coexisted in stock.
If you keep one sentence from this article, keep this one: traceability is not software, it is a record made at the second the product changes state.
Tying a lot to a delivery: the minimum chain
Three links are enough, and they are ordered. The first goes from lot to pallet, at packing. The second goes from pallet to delivery note, at loading. The third goes from delivery note to customer, and it already exists in your business.
The second link is the one most often skipped, because it is useless in normal operation: the driver leaves, the customer signs, the invoice follows. It serves only on a Thursday afternoon, and it is the only one missing when it is needed.
The form the record takes matters less than people think. A photograph of the delivery note with pallet identifiers visible, timestamped, is worth more than a sophisticated system nobody uses on the loading bay. What automation adds here is reading those photographs and filing them, not the gesture itself — and if the record lives in two tools at once, two systems are never quite in agreement and one of them has to be declared authoritative.
With those three links, Thursday’s question becomes a few minutes of searching. Without the second, it stays a day of binders, whatever was spent elsewhere.
What the system never declares
One sentence is forbidden and it should be named: the system never says a lot is compliant. It says a lot was produced on a given day, from recorded raw materials, delivered to given customers, and that the planned checks were or were not entered. Compliance is a decision that commits a person and a signature.
The second prohibition concerns analysis results: the system can record a certificate and attach it to a lot, it does not interpret it and does not conclude in your place. A certificate misread by a machine, on a microbiological parameter, is the kind of error that is not recovered from.
The third is subtler and comes up often in demonstrations: the system does not compute a "probability of non-compliance" from production parameters. That figure looks useful; it rests on nothing until you have years of annotated history, and nobody here has that.
Those three refusals leave the system what it does well: record, tie, retrieve. It is narrow, it is verifiable, and it answers Thursday’s question.
Remaining life is calculated, not predicted
Remaining shelf life at delivery is a real commercial argument and also a source of confusion. What is calculated is a subtraction: use-by date minus today, known exactly once the lot is identified on the pallet leaving.
What is not predicted is a product’s real degradation in a chain whose temperature you do not control. A system announcing an "effective" life taking transport into account is manufacturing data nothing measures, unless you have placed temperature loggers, which is another project.
The distinction has an immediate practical consequence: what you can automate without risk is alerting on stock whose date is approaching, at your site and at your distributors’ if you know their receipts. That is already considerable, because it is what produces unsold goods.
And if you want remaining life at delivery to be a commitment, it is guaranteed by a dispatch rule — nothing ships below so many days remaining — rather than by a clever calculation. The rule is keepable, checkable, and can be relied on.
The recall: what has to come out within the hour
A withdrawal or a recall is the moment all this machinery earns its keep, and it is prepared as an exercise rather than as a possibility. The list to produce has four elements: the quantities of the lot concerned, the customers delivered, the delivery dates, and what is left in your own stock.
That list has to be printable. A recall is handled by phone, often outside office hours, sometimes by somebody with no access to the system — and a procedure depending on a connection is one that will fail at the worst moment.
The useful exercise costs an hour: pick a lot at random from a production three weeks ago and trace it to the customers with nobody warned in advance. The time you get is your real recall time. Most businesses discover it is far longer than what their procedures state.
The list is also better sorted in calling order than by quantity: the distributor who received the most is not always the one selling fastest, and it is the second you have to reach first. That order is decided once, calmly, and written beside the procedure. What automation changes here is the time, not the responsibility: the decision to withdraw, informing customers and communicating remain acts of management, and no tool takes them on.
Supplier documents: reading is not checking
Traceability also runs upstream: raw materials, packaging, additives, with their own lots and certificates. Those documents arrive in disorder, usually by messaging, usually photographed, exactly as in a professional practice — and what a system extracts from a document received applies here word for word.
What automatic reading adds is the tying: this certificate belongs to this raw-material lot, which went into these production lots. Reconstructing that chain by hand is the most thankless job in a quality function, and it is the one a machine does well.
What it does not do is check. A certificate that arrives is not a certificate that is valid: the date, the body, the parameter measured and the exact product reference have to be looked at by somebody whose trade that is. The system flags what is missing, not what is acceptable.
The realistic gain is therefore filing time and instant retrieval, not automatic quality assurance. It sells less well and it is what holds in front of an auditor.
The paper register does not disappear
A business computerising its traceability often wants, in the same movement, to abolish the registers kept by hand. That is the moment to slow down, for two reasons that have nothing to do with computing.
The first is that an inspection asks to see what the business actually keeps, and a signed paper register has an immediate evidential value a screen does not always have. The second is that a production line keeps running during a breakdown, and traceability cannot stop because a terminal will not switch on.
The pragmatic position is a deliberate duplication during the transition, with a clear rule about which source counts. That duplication costs a few minutes per shift and avoids the situation where nobody knows which of the two sources is right.
Abolishing paper, if it happens, comes after a period in which the two sources have been compared and found to agree. It is a management decision, and it deserves to be dated and written down.
The check: a lot at random, traced by hand
Pick a lot produced three weeks ago and ask three questions, clock in hand. How many units were produced under that number. Which customers did they go to, and when. How much of it is left in your stock today.
The total time is your indicator, and it is the only one that counts. Under an hour, your recording chain holds and a tool will mostly buy you comfort. Beyond half a day, the missing link is almost always the second one in section 5 — pallet to delivery note.
Add a question that costs five minutes and reveals a lot: ask two people on the line what the lot rule is. If the two answers differ, start there, before any automatic photography of anything.
This exercise is also what you show a supplier. It turns a conversation about traceability into a conversation about a real lot, and it makes two proposals comparable.
What we do, and what we refuse
We install capture at packing — photograph of the code, pallet identity, one-gesture confirmation — the tie to the delivery note at loading, and the search that answers Thursday’s question. We provide the printable recall list, and we leave your data in a format you can export.
We refuse to declare a lot compliant, to interpret an analysis certificate, and to announce an "effective" shelf life taking transport into account. We also refuse to abolish your paper register until the two sources have been compared long enough for you to decide yourself.
We put forward no reading rate. It depends entirely on your packaging — inkjet on film, curved surface, reflective metal — and a figure given before seeing your products would describe nothing. We measure on your packaging, in front of you, before any commitment.
What you can do without us is the first step and the most decisive: write the lot rule in one sentence, and run the exercise in section 11. Many businesses discover there that the missing link is a thirty-second gesture on the loading bay, and the rest belongs to the label itself.
Frequently asked questions
Do we need management software before doing this?
No. Capture at packing and the tie to the delivery note sit upstream of any software, and they work even if your management runs on a spreadsheet. The reverse order — choose software then discover what is not recorded — costs more.
Our codes are inkjet-printed on film. Is that hopeless?
No, but the reading station becomes the subject: constant lighting, fixed distance, the shot taken at packing rather than in storage. It is an adjustment of a few tens of centimetres, and it makes the difference between an acceptable failure rate and a system nobody will use.
Can we trace our products to the consumer?
To the customer delivered, yes, and that is already what a recall requires. Beyond that it assumes your distributors record their own outflows, which cannot be decreed from your site. A complete chain to the distributor beats a theoretical one to the shelf.
How long does this take to put in place?
Capture and tying go live within weeks, most of the delay being the reading station and the team’s habit. What takes longer is the period of duplication with the paper register, which is a decision rather than a technical constraint.
What if we produce several lots in one day?
That is precisely the case where reconstruction from quantities collapses, and therefore the one where recording at packing becomes indispensable. If you are in it, the question is not whether the subject is worth the work but in which order to do it.
Can the system alert on short dates?
Yes, and it is the most profitable use after traceability itself, because it acts on unsold stock. The condition is knowing the stock — yours always, your distributors’ only if they send you their receipts.
Where we come in
The clock in section 11 gives you your real recall time, and it is almost always longer than the one written into your procedures.
- We put the shot where the code is still sharp, at packing, and not in the storage area.
- The pallet-to-delivery-note link is recorded on the loading bay in one gesture, because that is the one always missing.
- The recall list prints and reads with no connection, since a recall is rarely handled in front of a screen.
We will never write that a lot is compliant, nor what an analysis certificate is worth: both are judgements that commit a signature, and a machine has none.
Read next
Food and drink: the label is your first page
Your buyer reads you holding the product, on a shelf, in eight seconds. The website comes a long way after.The print run: ordering packaging without paying for it twice
Packaging is corrected before the press and never after. Here is the order of decisions, from the quote to the reversed calendar.Packaging: what reads at a metre
Packaging is judged at a metre, beside a competitor. Three things get through at that distance; everything else waits to be picked up.
Let us talk about your project
A free audit, no commitment: we look at your online presence and tell you what is holding it back.