Multibatch ISO 22716 GMP records with AI — digitize the dossier without throwing away the paper or changing the ERP.

The problem

The records are done right — it's rebuilding the batch that kills you.

A cosmetics plant that earned its ISO 22716 certification many years ago does not have a record-quality problem — the operator writes it down, the shift lead closes the batch, the lab files the control. The problem shows up at three moments:

1. When an external audit walks in (ISO 22716, health authority, a global retailer like Sephora or Watsons) and asks you to rebuild the file of several batches within a few days: someone has to fetch the physical logbook, scan it, cross it with the shift lead's spreadsheet and with the paper cleaning log. Weeks of a quality team busy filing instead of producing.
2. When a complaint arrives and you need to see what happened in that specific batch: the trail exists, but it is scattered across three folders, two spreadsheets and a logbook taken home by the shift lead who just retired.
3. When the veteran's knowledge walks out with him. The fine-grained decisions he made on each batch (why he raised the temperature half a degree in reactor 3, why he extended the mixing time) live in his head and in a notebook, not in the system. The day he retires, half the plant goes with him.

The consultant's classic way out (moving everything to a vertical cosmetics MES) costs six or seven figures and forces the plant to change how it records. During the months of transition, record quality gets worse, floor resistance runs high, and the ROI fades away. It does not have to be that way.

How it fits the IRIS system

iLEAN Connect is the putty that digitizes GMP without throwing away the paper.

The GMP record problem in small and mid-sized cosmetics plants is not what gets recorded — that part is already fine — it is where it lives and how it gets rebuilt. Three islands: the plant logbook, the shift lead's spreadsheet, the old machine panel. iLEAN acts as the putty that fills the gaps between those islands, without asking you to change the machine, the ERP or the way the plant works.

Connect reads the logbook from a photo, the spreadsheet wherever it lives, the old panel even if it is isolated. The agents cross-reference, index by batch and build the dossier. The quality manager signs the batch closure — the paper stays in the filing cabinet if you want it, and the digital trail builds itself.

The specific piece for this pain:

• iLEAN Connect — the piece that fills the dead zones between ERP, paper and panel. Graduated capture adapted to each source: (1) manual — the operator or the shift lead photographs the logbook or the analog panel when closing the record, and Connect transcribes and indexes it by batch; (2) intermediate — if the machine has a local computer, however old and isolated, Connect connects and extracts without changing the machine; (3) integrated — if there is a modern interface, direct integration. Plus outward capture: the supplier's CoA that arrived by email, the incident note the customer sent over WhatsApp — everything lands in the batch file at second zero.
• iLEAN Agents — a swarm of agents in Central that cross-reference what Connect captures with the production order in the ERP, the lab control and raw-material traceability. They build the batch file during the operation: by the time the auditor arrives, the dossier already exists. And they keep the supporting documentation (procedures, GMP work instructions, staff training) versioned — which ISO 22716 also requires.
• The three rings guarantee that the human signature is mandatory at batch closure and that records travel cryptographically signed. What is critical is never decided alone.

See the full IRIS architecture →

Before and after

Paper + spreadsheets + vertical MES vs. iLEAN Connect on top of what you already have

Aspect	Classic model (paper + spreadsheets) or vertical MES	With iLEAN Connect + Agents
Tool change for the operator	Paper stays, or a new MES that breaks the flow	None — Connect captures where the operator already records
Rebuilding the batch file	Weeks, with the quality team tied up	Minutes — built during the operation
Capturing the shift lead's spreadsheet	Lives in a forgotten shared folder	Connect reads it wherever it is and indexes it to the batch
Analog panel on an old machine	Nobody records it digitally	Operator photo or a read from the isolated PLC
The veteran shift lead's knowledge	Leaves with him when he retires	Captured as voice notes / indexed notes
Project cost and timeline	Vertical MES: six-seven figures, 12-24 months	Connect pilot: first value in weeks

Impact estimate

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

The block below is an estimate to be validated with the specific data of your plant. We put it forward so management and the quality manager have an order of magnitude; we refine it during the diagnostic.

• Mid-sized cosmetics plant with several reactors and parallel packing lines, a mix of modern and old machines, records split across paper + spreadsheets + ERP.
• Connect pilot on 1-2 reactors and the packing line most critical for your main retailer: capture of the logbook and the old panel, connection to the ERP, batch file construction. First value expected within a few weeks — from the first captured batch.
• Indicative payback between 4 and 9 months, depending on the annual time your quality team spends rebuilding files for audits/complaints and that team's labor cost.
• Defensible floor for reducing per-batch file rebuild time: ≥ 30%. The hard lever is not just labor cost — it is passing the audit without surprises and keeping the global retailer as a customer.

And the quality manager's reasonable doubt

“What if the AI transcribes the logbook wrong and the record ends up with an error?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely recontextualizes a data point from one format to another (reading the photo of the logbook and pouring it into a structured record), the best models brought the error below 1.5% ^[1]. And even so, what is critical is never decided alone: Connect proposes the transcribed record next to the original photo, and the quality manager signs the batch closure. The three safety rings exist precisely for this.

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