Cosmetics manufacturing with AI — the emulsion that split, the batch that passed GMP by hand and the audit that started over.

The problem

The batch record closed by hand is always the batch record the audit sends back.

A cosmetics plant lives at the intersection of three realities that regulation requires you to document and that are almost never in the same place at the same moment:

1. The emulsion — order of addition, jacket temperature, homogenization energy and time, cooling rate. It lives in the homogenizer SCADA or on the operating sheet.
2. Traceability of the active — the batch and the variability of the active or pigment declared in the supplier's certificate, which arrives by email or PDF and almost never makes it into the GMP room. The R&D manager's formula lives in a spreadsheet or in a PLM.
3. The GMP batch record — the file that proves every step was carried out according to the approved procedure. It is closed at the end of the batch, almost always by hand, with signatures and dates to be reconstructed.

The technical director knows it. The veteran formulator corrects by eye when the emulsion starts to split, but that know-how does not stay behind. The filling operator has the screen and the paper. And when the audit arrives — ISO 22716, COSMOS, Ecocert, FDA if you export to the US, the national health authority if there is a complaint — the dossier is the first thing the auditor looks at. The classic system works 95% of the time. That other 5% is the batch that gets reformulated, the dossier that gets rebuilt or the finding that takes months to close.

How it fits the IRIS system

iLEAN does not replace your GMP system — it seals the cracks between formula, homogenizer, filling and batch record.

The problem in cosmetics is not a lack of procedure: it is information living on islands (formula in the PLM, active certificate in an inbox, the real homogenizer profile in SCADA, the Edge reading on the line, the signature on paper) that only the final batch record tries to bring together, almost always by hand. iLEAN acts as the putty that fills the gaps without touching your ERP, PLM, MES or your approved GMP procedures.

Edge sees the emulsion and the filling. Connect captures the active certificate and the formula. The agent cross-checks the three realities and prepares the batch record at close. The person signs — the batch is never released on its own.

The three iLEAN pieces applied to cosmetics manufacturing:

• Edge — a terminal with vision (CNN) and analog inputs in the GMP room and on the filling line. It detects emulsion instability, early separation, color drift against a reference plate, continuous viscosity, abnormal foaming. It reads real jacket and coil temperature, the scales and the pressure gauges. It verifies the pack and the crimping at close. It works with no network: in a cleanroom, the last thing you want is to depend on WiFi.
• Connect — captures the certificate for every active and pigment, whether it arrives by EDI or as a PDF in the technical director's inbox, the batch formula in the MES, PLM or a spreadsheet, the SDS for every ingredient, and historical variability. Whatever comes from outside (a retailer's change in natural cosmetics, new INCI regulation, a new batch of active with a different declared activity) enters at second zero.
• Agent — cross-checks the batch certificate, the formula, the real homogenizer profile and the Edge reading. It recognizes the pattern (“with this batch of active and this cooling rate, this cream tends to show early separation within 48h unless the temperature drops 2°C”) and proposes the in-line adjustment. And it prepares the batch record at batch close: ingredient-by-ingredient traceability, temperature curve, a photo of the appearance, the scale reading, mass reconciliation. The person reviews and signs.

See the full IRIS architecture →

Before and after

Cosmetics manufacturing run on paper vs. cross-checked with iLEAN

Aspect	SCADA + procedure + batch record by hand	With iLEAN Edge + Connect + Agent
Variability of the active	Certificate by email, stays with purchasing	Cross-checked against the formula and the homogenizer profile
Detecting a split emulsion	Inspection at the end of the batch or after storage	Anticipated during homogenization, before cooling
Color and viscosity	QC at the end, rework if out of spec	In-line reading, adjustment proposed before packing
GMP batch record	Manual reconstruction, signatures after the fact	Complete dossier at batch close, with photo and traceability
Operation with no network	n/a	Edge keeps reading and holding the data
Audit (ISO 22716, COSMOS, FDA)	Days of preparation for every batch audited	Signable file in minutes, automatically

Impact estimate

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

The block below is an estimate to be validated with the real data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Cosmetics plant working private label, contract manufacturing or natural cosmetics, multi-SKU, multi-customer, with the batch record on paper + spreadsheets and recurring ISO 22716 / COSMOS / FDA audits.
• Edge pilot in one GMP room + one filling line, Connect for the active certificate and the formula, an agent that cross-checks and closes the batch record. First value expected in a few weeks.
• Indicative payback between 4 and 9 months, depending on the percentage of batches reformulated/rejected, the average cost of rework and how long it currently takes to close the batch record + audit dossier.
• The hard lever is threefold: rework cut by ≥30%, an audit dossier prepared by default, and a lower cost in formulator hours spent rebuilding paperwork.

And the GMP quality manager's reasonable doubt

“What if the AI gets it wrong and lets an out-of-spec batch through?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI compares what it sees and measures against the recipe, the certificate and the procedure, the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN holds the batch and the person signs. The three safety rings exist precisely for this — and they fit the GMP logic of human validation you already know.

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