Multiproduct campaign changeover with AI — a badly executed changeover is not a delay, it is cross-contamination.

The problem

The campaign changeover works 99% of the time. That 1% is what puts the customer at risk.

Any production manager running a multiproduct reactor knows the changeover is not a boring routine — it is the most vulnerable moment of the month. Four things collide at once:

1. The cleaning between products: the operator cleans following the previous product's protocol, but the correct protocol is the one for the "previous product → next product" transition, and the transition matrix sometimes lives in an Excel file the operator never checks at the line.
2. The verification that the cleaning came out right: a sample goes to the lab, the result is awaited, and meanwhile the crew "gets ahead" on the next preparation. When the result runs late, there is a temptation to restart under "controlled risk."
3. The next recipe: production prepares it by typing or copying from the MES, with room for human error and dead time between the end of cleaning and the start of production.
4. The dossier for the auditor (FDA, IFS, EuGMP, ISO, or a B2B customer with its own quality team): it is assembled after the fact by cross-referencing shift reports, photos and the shift lead's memory. The day the inspection arrives, that reconstruction is what management suffers through.

The classic system (checklist + eye + signature + shift report) works 99% of the time. But the cost is not in the majority: it is in the rare changeover where residue from the previous product shows up in the next batch, or in the batch rejected because the traceability fails to convince the auditor. A single cross-contamination with a regulated customer pays for many years of a system done right.

How it fits the IRIS system

iLEAN does not replace the MES, the cleaning SOP or the shift lead — it seals the cracks between the three.

The changeover does not fail for lack of procedure: it fails because the procedure, the image of the reactor, the MES recipe and the transition matrix live in different places and get cross-referenced in the operator's head at the most critical moment of the month. iLEAN acts as the putty between those pieces: it sees the reactor with vision, reads the transition matrix and the MES recipe, cross-references them on the spot, and puts the decision in front of the manager with the evidence attached. The three rings guarantee that only what has been validated enters the MES — the person signs, always.

Edge sees the reactor clean before loading. An agent coordinates the cleaning with the correct protocol per the matrix. Tracer maintains the genealogy of the previous batch and the next one. The person signs the startup authorization — always.

The three iLEAN pieces applied to multiproduct campaign changeover:

• iLEAN Agent — an agent orchestrates the changeover: it pulls up the "previous product → next product" matrix, triggers the correct cleaning protocol, prepares the next recipe in the MES queue, validates raw material availability, checks that calibrations are current and, when everything is in place, leaves the authorization ready for the production manager. The agent does not authorize the startup: it prepares so the person can do it fast and do it well.
• iLEAN Edge — a terminal with computer vision (CNN) over the reactor, the lines, the centrifuge and the receiving tank: it compares against the accepted-clean reference image and detects remnants, out-of-range colorimetry, residue in gaskets. It runs in parallel with the rinse measurements from the DCS and the rapid lab (TOC, conductivity, pH). It works without a network; in a chemical plant, that matters.
• iLEAN Tracer — maintains the genealogy of the previous batch, the rinse and the next batch, and builds the evidence dossier while the changeover progresses: image of the clean reactor, rinse measurements, protocol applied per the matrix, human acceptance signature. When the audit arrives, the dossier is already built — it is not manufactured after the fact.

See the full IRIS architecture →

Before and after

Classic campaign changeover vs. changeover coordinated with iLEAN

Aspect	Classic changeover (SOP + checklist + shift report)	With iLEAN Agent + Edge + Tracer
Cleaning protocol	Previous product's SOP, operator's judgment	"Previous → next" matrix applied automatically
Cleaning verification	Visual + checklist + sample to the lab (waiting)	Edge vision + DCS measurements + rapid lab in parallel
Next recipe	Typed into the MES by production at the end	Pre-loaded into the MES by the agent while cleaning runs
Dead time between phases	Waiting for the lab + waiting for authorization	Reduction ≥ 30% over the baseline
Evidence dossier	Reconstructed after the fact from reports and memory	Built on the spot, with image and measurements
Audit preparation	Person-days reconstructing changeover by changeover	Dossier already built, just checked

Impact estimate

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

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

• Chemical or pharmaceutical plant with a multiproduct reactor, 8-15 SKUs a year, several changeovers a month, vertical MES (Werum PAS-X, SAP MII, Aspen, or each subsector's vertical).
• Pilot on one reactor: vision over the cleaning, an agent that cross-references the "previous product → next" matrix, recipe pre-loading in the MES, dossier built in parallel. First value expected within a few weeks.
• Expected reduction of total changeover time ≥ 30% (to be validated against your real SMED). On a reactor with 8-12 changeovers a month, that is new operable hours every month.
• Indicative payback between 4 and 9 months. The hard lever is not just time: it is avoiding a single cross-contamination with a regulated customer — one rejected pharma batch, or one from a B2B customer with its own audits, pays for the pilot many times over.

And the quality manager's reasonable doubt

“What if the AI accepts a cleaning that was not good, or misapplies the matrix and lets a changeover through with cross-contact risk?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely recontextualizes a piece of data from one system to another (reading the reactor image and comparing it against the reference, cross-referencing the transition matrix with the recipe), the best models brought error below 1.5% ^[1]. And even so, what is critical never closes on its own: iLEAN holds the changeover and the person signs. The three safety rings exist precisely for this — the MES remains the system of record, and only what a human has validated enters the MES.

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