Pharmaceutical dispensing with AI — one badly signed weighing costs the whole batch.

The problem

A weighing is signed in four seconds. The deviation costs days.

Raw material dispensing is one of those steps that look simple until an auditor opens the dossier. You have a master formula living in the MES, a container of API or excipient with its code and its status (released, quarantine, retest), a scale that should be within tolerance, and an operator who in four seconds reads the weight, notes it down and signs. When everything lines up, nothing happens. When one of those four pieces breaks, you do not notice it in the booth — you notice it three weeks later, when QA spots the deviation in the batch record.

The typical failure modes are not dramatic, they are quiet: a quarantined API container picked up by mistake because the visual label was identical to the released batch; a scale that drifted 0.8% and nobody recalibrated since the last review; a master formula updated in the MES that the operator had printed on paper from the previous batch; a batch code the operator keyed in by hand because the reader would not scan the label. The classic system (operator + checklist + second signature) works 99% of the time. That 1% is the rework of a whole batch — and in biologics or ATMPs the batch can be worth six figures.

How it fits the IRIS system

iLEAN does not replace your MES — it seals the cracks between the MES, the scale and the person.

The dispensing problem is not a lack of systems, it is information living on islands and cross-checked by hand at the most critical moment. The MES has the formula. The scale has the weight. The container has its code. The operator has their eyes. And between the four of them there are a couple of cracks where the error slips through. iLEAN acts as the putty that fills those gaps, without asking you to replace the MES, the scale or the booth.

Edge sees the scale and the code. Connect brings the live formula from the MES into the booth. The agent cross-checks everything against the GMP rules and, if something does not add up, holds. The person signs — never the other way round.

The three iLEAN pieces applied to pharmaceutical dispensing control:

• Edge — a machine-vision terminal in the booth. It reads the scale (whether a modern Mettler or an old Sartorius with an analog display), the container code and the MES panel if there is one. It compares the weight it reads against the formula and triggers the hold (red light, weighing close blocked) in milliseconds if the deviation exceeds tolerance. It works with no network.
• Connect — captures the master formula from the MES (Werum, POMS, Camstar or whatever you run) and the container status from the LIMS/WMS. It also captures what arrives from outside — a supplier change with a new CoA, the retest of a batch, an urgent instruction from the compliance lead by email — and puts it in the booth at second zero, with nobody forwarding anything or calling a meeting.
• Agent — cross-checks formula + weight + batch + booth history + Annex 1 and Part I chapter 5 rules. If it detects a deviation, it does not send an email at 10 pm: it holds the weighing, proposes the action (recalibrate the scale, return the container to quarantine, escalate to QA) and alerts the person in charge. The operator validates; the booth does not restart on its own. And it assembles the evidence dossier for the auditor who will come by.

See the full IRIS architecture →

Before and after

Manual dispensing vs. cross-checked dispensing with iLEAN

Aspect	Manual dispensing + double signature	With iLEAN Edge + Connect + Agent
Master formula	In the MES; printed on paper in the booth	In the booth, live, synced with the MES at second zero
Scale reading	The operator reads it and keys it in	Edge reads it with vision and publishes it as structured data
Container status (released/quarantine)	Visual label + manual scan	Automatic cross-check against LIMS/WMS before the weighing is allowed
Tolerance deviation	Spotted by QA when reviewing the batch record	Held in the booth before closing, in milliseconds
Operation with no network	n/a	Edge keeps reading and holding on the booth's own power
Dossier for the auditor	Rebuild the batch record by hand	Per-batch file with a photo of the scale and the container, automatic

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.

• Mid-sized pharma plant, 2-4 dispensing booths, a mix of released and quarantined API raw materials, a mix of modern Mettler scales and old RS-232 Sartorius units, MES on site (Werum, POMS or equivalent).
• Edge pilot in one booth (vision terminal + integration with MES and LIMS + validation agent). First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on the documented frequency of deviations in the batch record and the average cost of rework / lost batch across your product portfolio.
• The hard lever is a single biological batch saved per year, or a couple of critical deviations avoided. The pilot pays for itself.
• Replication to the other booths done by your own people — the E.1 lesson works in pharma exactly as it does in automotive.

And QA's reasonable doubt

“What if the AI misreads the scale and signs an incorrect weighing?” — hallucination is a problem of free generation, not of anchored tasks. Reading a scale display and comparing it with the formula is an anchored task par excellence. In tasks of that kind, the best models brought error below 1.5% ^[1]. And even so, the reading does not decide: the booth holds whenever there is doubt and the person 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.