Galenic formulation with AI — from bench to plant without losing the conversation along the way.

The problem

The document reaches the plant. The conversation stays in R&D.

Transferring a galenic formula from the R&D lab to the plant works 95% of the time. The remaining 5% are the learning batches — the ones that cost two weeks, a quality committee and an uncomfortable conversation with the technical director.

The master document arrives, that part does not fail. What does not arrive is the living conversation around the document: the margin note in the lab notebook about how the blend behaved on a humid day in September, the email from the excipient supplier warning of a particle size change, the WhatsApp from the scale-up lead explaining why he raised tablet press speed by one point on the last pilot batch, the version of the spreadsheet with the final parameters that stayed in the personal Drive of the person who went on leave.

The three data sets you would need to cross-check for the transfer to go right the first time live on three islands:

1. The official document — master formula, manufacturing instructions, regulatory dossier. In the PLM or eDMS.
2. The living conversation — emails, WhatsApps, margin notes, intermediate spreadsheets, photos of a whiteboard. In personal accounts, on phones, in paper notebooks.
3. What happened in the plant on the first batch — documented deviations, operator comments, real SCADA parameters. In the batch record and in the shift lead's head.

Three islands, none of them talking to the others. The expensive learning batch lives in the joint between the three — and nobody was there coordinating in time.

How it fits the IRIS system

iLEAN does not replace the PLM — it listens to everything the PLM never receives.

Coordinating galenic formulation is not a problem of missing documents. The modern plant already has a PLM, an eDMS and signed batch records. The problem is that the living conversation — the one that decides whether the first batch comes out right — never enters any system; it lives in emails and WhatsApps and paper notebooks. iLEAN is the putty that seals that crack.

Connect listens to email, WhatsApp, calls, photos of notebooks, forgotten spreadsheets — everything that surrounds the official document. The agent cross-checks it and prepares the transfer file. The person signs — never the other way round.

The two iLEAN pieces that coordinate galenic formulation:

• Connect — inward listening (the formulator's notes, photos of the notebook, the scale-up lead's voice through their earpiece or tablet) and outward listening (the supplier's email about an excipient change, the WhatsApp from the tablet manufacturer with the result of the first pilot batch). And it digitizes: photograph a lab notebook and it extracts the data; read the panel of an old tablet press and it records the parameters. Bidirectional: it also drafts replies going out through the same channel, but the person signs.
• Agents on the Central platform — one per key user: an agent for the formulator (cross-checks notes against literature and previous design), one for the scale-up lead (cross-checks against the real pilot parameters), one for regulatory affairs (cross-checks against the eCTD and proposes what to update in module 3), one for the plant manager (cross-checks against the MES and batch records). The agents are boring and concrete: they take away the dumb work and leave the judgment work to the person.

See the full IRIS architecture →

Before and after

Classic transfer vs. transfer coordinated with iLEAN

Aspect	PLM + emails + notebook	With iLEAN Connect + Agents
Capture of the living conversation	In personal emails and WhatsApps	Connect captures everything at second zero
Margin notes in the notebook	Lost when people move on	Photo + extraction → permanent record
Supplier email	Lands in an inbox, not on the plant floor	Connect propagates it to whoever decides
Updating eCTD module 3	Manual RA work, weeks	The agent proposes what to update, RA signs
Learning batch in the plant	1-3 batches to tune parameters	Arrival with a complete, living file
Transfer audit	Rebuild from 5 sources	Automatic dossier with every source

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 organization. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Galenic R&D department with 3-5 active transfers to an internal plant or a CMO, RA with a backlog of variations, scale-up with incidents documented over the past few years.
• Connect + Agents pilot for one specific transfer (the next formula to scale up). First value expected within a few weeks: reduction of the time needed to prepare the transfer file by ≥ 30%.
• Indicative payback between 4 and 9 months, dominated by weeks of RA work saved and learning batches avoided.
• The hard lever: a single transfer done right the first time. The first batch comes out in spec, the dossier reaches the agency with fewer versions, RA gets to breathe.

And the CAIO's reasonable doubt

“What if the agent cites a wrong figure from the notebook in the dossier?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the model merely recontextualizes (reading a note from the notebook and proposing its impact on module 3), the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: the agent proposes, regulatory affairs signs with their own credential. The three safety rings exist precisely for this.

^[1] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks. ^[2] Fundación BBVA / Ivie — productivity from digitalization, 2000-2021.