The AI proposes, the human signs, and only the signed crosses

The problem

When a system is right 97% of the time, the problem is that nobody knows which the remaining 3% is.

In a fresh cheese dairy, the data point that does the most damage if it enters wrong is the variety — because the same white 250 g piece can be cow, goat, blend, salt-free or lactose-free. And without a human gate, the chain of consequences is direct:

• A variety recognition error enters the master — and from there it propagates to the labeling and appears on a retail chain's shelf, without anyone having seen it along the way.
• In a product with a declaration it is not an inventory failure — with an intolerance declaration (lactose-free) or a species one (goat), that error is a food safety incident, with everything that drags along.
• And the human gate cannot cost time — because if it costs time, on the line it gets skipped. It has to be faster than the alternative of not doing it.

Hence the answer is not choosing between automating and controlling, but separating the capture — which can be automatic — from the validation, which must be human but can take two seconds.

How it fits the IRIS system

Early human verification — a validation screen, not a form.

The design key is in the verb. A data entry screen asks for writing and on the line gets abandoned. A validation screen shows what the AI understood and asks for a gesture. The first competes with the operator's work; the second takes less than checking the paper order.

A sealed tablet at the packing station, suited to the environment. It shows a visual summary: the photo of what was captured, the four fields that matter in large type — variety, batch, expiry, format — and the contrast with what the work order says. If they match, one tap.

How early human verification operates at packing:

• A visual summary, not a form — four fields in large type and the photo of what was captured beside them. No keyboard in the normal flow and no nested menus.
• Contrast against the work order — the screen does not show loose data: it confronts what was captured with what the order says it should be, which is exactly the check the operator already does mentally.
• One tap if it matches, field correction if not — when something disagrees, the operator corrects the specific field on the screen itself, they do not rewrite the whole record.
• The system learns from the correction pattern — if a field is always corrected at the same point, that signals the capture is badly calibrated there, not that the operator is wrong.
• It works for everything captured, wherever it comes from — photo, voice, folder reading or email: everything converges on the same screen, with a record of who signed and when.

See the full IRIS architecture →

Before and after

Unsigned data vs. data validated in two taps

Aspect	Automatic capture with no human gate	With iLEAN + validation tablet
Data entering the master unreviewed	Happens, and nobody knows which	Zero
Validation time per batch	—	Seconds
Variety errors	Detected on the customer's shelf	Detected on the line
Product with a species or intolerance declaration	Risk of a food safety incident	Confirmed by a person before labeling
Traceability of the signature	Does not exist	Who signed and when, on every data point
Quality's stance on the project	Blocking	Backing: there is a recorded human signature

Impact estimate

Impact estimate for your plant — to be validated with your own numbers.

The block below is an estimate to be validated against your plant's actual data. We put it forward so the committee has an order of magnitude; we refine it during the assessment.

• Fresh cheese dairy packing several varieties in the same format, with species and intolerance declarations on the label.
• Pilot of the sealed tablet at the packing station together with the first capture case. It is not deployed on its own: it is the piece that makes the others signable. First value expected within a few weeks.
• Indicative payback between 3 and 7 months. Estimate to be validated.
• It is the cheapest example to deploy and the one eliminating the most risk — and the one making the project defensible before quality and the auditor, because there is a recorded human signature on every data point that enters.
• The risk it avoids does not average out: a variety error in a product with an intolerance declaration reaching the shelf is a food safety incident, not a mismatch.

And the fair question from the line manager

"If every batch must be validated, don't we end up just as loaded?" — no, because validating and typing do not cost the same. Typing is tens of seconds plus the load of remembering what goes in each field; validating is looking at four fields in large type next to the photo and confirming. The interaction is designed to take seconds, because if it cost time it would get skipped. And it is precisely that validation which lets the AI work on anchored tasks safely: even with an error below 1.5% ^[1], that remainder never reaches the master.

^[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.