Paraben-free challenge testing with AI — when the natural preservative leaves less margin, the LIMS cannot be a spreadsheet.

The problem

Without parabens the efficacy margin narrows — and the data that proves it usually sleeps in a spreadsheet.

The classic paraben + phenoxyethanol system had been delivering a wide window for decades: it tolerated a careless pH, an imperfect emulsion, a packaging change from marketing. A natural/paraben-free range is not that forgiving. ISO 11930 is still the same — bacteria and fungi counts at day 7, 14 and 28, criteria A and B — but the real window of the formulation narrows and the number of iterations multiplies.

And this is where the operational problem shows up, which is not regulatory but a matter of capture:

1. The count — the microbiologist writes it down in their notebook or on a sheet in the isolated LIMS.
2. The formula in force at that time — R&D has its version, which does not always match the one that produced the batch.
3. The specific packaging of the pilot — because the same natural preservative behaves differently in an airless bottle and in an open jar, and that data point usually lives only in an email from the packaging manager.

When the EU Regulation 1223/2009 auditor arrives or the customer asks for the safety dossier, somebody has to reconstruct those three data points by hand. They almost always get it right — and that "almost" is where a perfectly good formulation falls apart.

How it fits the IRIS system

iLEAN does not replace the LIMS — it seals the cracks between the notebook, the LIMS and the recipe.

The problem with paraben-free challenge testing is not a lack of information: it is information that lives on islands and that, at the critical moment (the audit, the urgent reformulation because a supplier changed an extract), does not reach the decision-maker in time. iLEAN acts as the putty that fills those gaps, without asking you to change any of your current systems.

Connect captures the count wherever it lives. The Agents cross-reference it with the recipe in force and the ISO 11930 standard. The batch dossier comes out in one click — and the person signs.

The iLEAN pieces applied to the challenge test LIMS:

• Connect — captures the day 7, 14 and 28 count whether it comes from the modern LIMS, from the isolated legacy LIMS or from the microbiologist's physical notebook (photo reading, voice dictation, direct integration depending on the case). And it also captures what comes from outside (the supplier's email with a new extract specification, the importer's WhatsApp with the US retailer's rule) at second zero, without asking anyone to forward anything.
• Agents — cross-reference the count, the formula in force when the test was run, the pilot packaging, the batch in production and ISO 11930. If a formulation changes (a change of extract supplier, a change of packaging) and the data suggests the efficacy margin is approaching the limit, the agent flags it before it goes to the shop floor. The person validates — the agent never decides alone.
• Three safety rings — the critical batch data (the formula that went into production) lives in ring 1, isolated. What comes from outside (a count from an external LIMS, the supplier's email) is validated in ring 2 before it touches ring 1. For CAIO/IT: it fits the AI Act and regulatory traceability without giving up the power of AI on the outside.

See the full IRIS architecture →

Before and after

Manual LIMS vs. LIMS cross-referenced with iLEAN

Aspect	LIMS + notebook + spreadsheet	With iLEAN Connect + Agents
Capturing the count	Physical notebook → manual transcription into the LIMS	Capture by photo, dictation or direct integration
Cross-reference with the formula in force	Emails between R&D and microbiology	Unified at second zero, with version control
Change of packaging / extract supplier	R&D is told if someone remembers	The agent cross-references the new data with historical efficacy and flags the margin
Iterations for a paraben-free formula	Multiplies the paperwork	Multiplies the clean data, not the microbiologist's workload
Dossier for the ISO 11930 / 1223/2009 auditor	Reconstructed by hand, weeks	Per-batch dossier, automatic, with notebook + LIMS + packaging
The veteran microbiologist's knowledge	Walks out with them the day they retire	Pattern captured as a permanent capability

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 the committee has an order of magnitude; we refine it during the diagnostic.

• Cosmetics plant with a growing natural/paraben-free range, legacy LIMS or modern LIMS with partial capture, physical notebook still in use by the veteran microbiologist.
• Connect pilot on the LIMS + notebook + integration with the recipe in force in the PLM/ERP. First value expected within a few weeks: the first ISO 11930 dossier that comes out in one click instead of taking a whole morning.
• Indicative payback between 4 and 9 months, depending on the number of challenge tests per year, the cost of late reformulations and the hidden cost of audit hours prepared by hand.
• Hard lever: a single formulation rejected in an audit avoided, or a single production run not put on hold because the data was there in time.
• Expected reduction in dossier preparation time for the auditor of ≥ 30% as soon as the agent picks up the first complete challenge test cycle.

And the microbiology director's reasonable doubt

“What if the AI makes up a count or misreads the notebook?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely recontextualizes a data point from one system into another (reading a figure from the notebook or the LIMS and moving it into the dossier), the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: the microbiologist validates and the quality manager signs. The three safety rings exist precisely for this.

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