Nothing enters the system without a human signing it

The problem

The two blockers that kill a digitization project — and both are reasonable.

When an automatic capture project collapses in a plant, it is almost never the technology. It collapses on two objections raised in the first meeting, both of them perfectly legitimate:

• "AI is going to put bad data into the ERP" — the fear of the quality manager and of IT. A contaminated master costs more to clean than the automation saved, and whoever signs for data quality does not want a system that writes on its own.
• "The operator will not type" — the production manager's fear, and they are right: any solution that adds screens and fields at the booth is abandoned within three weeks, however good it is.
• The usual way out is to do nothing — either you automate without validating and pay in data quality, or you require full manual validation and nobody does it. Either way the project dies, and with it everything that depended on it.

The contradiction is only apparent: the answer is not choosing between automating and controlling, but separating capture (which can be automatic) from validation (which must be human, but can take two seconds).

How it fits the IRIS system

Connect in early-human-verification mode — a UI for validating, not for typing.

The design key is in the verb. A data entry screen asks people to write, and gets abandoned. A validation screen shows what the AI understood and asks for a gesture: correct, or fix me. The first competes with the operator's work; the second takes less time than reaching for a pen.

Every capture arrives on the tablet as a two-second visual summary, with its source in view — the photo of the sheet, the supervisor's dictation, the customer email, the panel frame — alongside what the AI extracted. The responsible person validates or corrects. Only then does the data cross into the ERP or the QMS.

How early human verification works in a paintshop:

• Two taps, not a form — the default interaction is to confirm. Correcting is the exception, and when it is needed you correct the specific field, not rewrite the record.
• The source always visible — the evidence it came from is shown next to what was extracted. Whoever validates does not have to trust: they compare. That is what turns the signature into a real signature.
• Doubtful items arrive flagged — if a field had low confidence, it appears highlighted and is the first thing reviewed. The system does not hide its uncertainty to look more confident.
• Who signed is recorded — every validation carries a person and a moment. That feeds the audit evidence pack directly: not just the data, but who stood behind it.
• At the line side, not in an office — the tablet sits where the process happens, because validation that requires walking to an office piles up and ends up being done in bulk, which is the same as not doing it.

See the full IRIS architecture →

Before and after

Automating without validating vs. early human verification

Aspect	"Autonomous" automatic capture	With iLEAN Connect + validation
ERP master data quality	Risk of silent contamination	Only what a person has signed gets in
Operator time	Forms that get abandoned	Two taps on a visual summary
Traceability of the decision	You know the data, not who backed it	Person and moment recorded
Low-confidence data	Gets in anyway, indistinguishable	Arrives highlighted and reviewed first
Acceptance in the committee	Blocked by the quality manager	The quality manager keeps control
Evidence for audit	Data exists, ownership does not	Data with signature and source behind it

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.

• Automotive paintshop that wants to capture startup sheets, oven panel, lab and external alerts automatically, but needs to guarantee master data quality.
• Pilot of the line-side tablet alongside 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.
• This case is enabling and strategic, not a direct saving: without it, the rest of the matrix does not get past the committee. Value to be validated together with the cases it enables.
• Its measurable effect shows up in the other cases: they are the ones that deliver payback, and they only get approved if this piece comes first.
• Data governance benefit: every record keeps its source, its confidence level and the person who validated it, which later translates directly into audit evidence.

And the fair question from the production director

"If every capture has to be validated, do we not end up as loaded as we were typing?" — no, because validating and typing do not cost the same. Typing a record takes tens of seconds plus the whole cognitive load of remembering what goes in each field; validating means reading a summary and confirming. In practice the gesture per capture is counted in seconds, and the volume concentrates on the responsible person, not on every operator. And it is precisely that validation that lets the AI work safely on anchored tasks: 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.