Zero unplanned downtime

The problem

On a line making 15,000 units an hour, a stopped minute never comes back.

Unplanned downtime on high-cadence lines is the sub-sector's biggest operational and economic risk. Every relevant incident costs unrecoverable lost production, plus emergency corrective maintenance, plus the risk of missing a major retail order.

• Unplanned downtime on high-cadence lines is the sub-sector's biggest operational and economic risk. Every minute stopped is production gone for good, not deferred.
• Each relevant incident adds emergency corrective maintenance on the mixer, the tunnel oven or the slicer, at the worst price and with whatever spare is at hand. Dough already mixed or proofing is lost along with the run.
• And with short-shelf-life bread, a lost run can mean missing a major retail order that cannot be made up the next day.
• The warning signs were there — a noise, an off-weight startup, a changeover that never settled — but each lived in a different place and nobody joined them until the line was already down.

How it fits the IRIS system

Four rings that verify each other — the eleven cases above, coordinated.

Ring 1 (Connect, cases 1 and 4): the startup log and issues dictated by the shift leader capture early deviation symptoms. Ring 2 (legacy integration, case 8): maintenance history is stitched to the ERP and the active Production Order. Ring 3 (Edge, cases 9 and 10): the line camera confirms the product is correct and that a format change started up well; if something doesn't match, JIDOKA AI alerts or blocks before the deviation spreads. Ring 4 (Agents, case 11): every intervention and format change gets documented, feeding root-cause analysis. SMED AI speeds up format changes so the coordination doesn't cost productivity.

No single ring prevents a stoppage. The strength is in the cross-check: a symptom dictated by the shift leader, the asset's maintenance history and what the camera sees on the belt, pointing at the same place before the line stops. When independent sources start to agree on a problem, that is the moment to act.

See the full IRIS architecture →

Before and after

The four rings against unplanned downtime

Ring	Cases it draws on	What it catches
1 · Early symptoms	Line startup (1) + shift leader headset (4)	Startup settings and noises before they become faults
2 · Maintenance memory	Integration via API, as with the coder (8)	Asset history tied to the ERP and the active order
3 · Product on the belt	Defect vision (9) + format change (10)	A deviation JIDOKA AI alerts on or blocks before it spreads
4 · Documented cause	Audit evidence pack (11)	Every intervention recorded for root-cause analysis
SMED AI	Format changeovers (10)	Faster changes, so coordination costs no output
Rings pointing different ways	All of the above	A person decides on the combined evidence

Impact estimate

Impact estimate — to be validated with maintenance and operations leadership.

The block below is an estimate to be validated against your plant's actual data. We put it forward so the committee has an order of magnitude; we refine it during the assessment.

• Estimated payback 6-12 months, counted against the first relevant stoppage avoided.
• Improved OEE on the covered lines, because recurring stoppages are caught in the process instead of corrected after the line stops.
• Lower emergency corrective maintenance cost, shifting spend toward planned and predictive work scheduled outside production hours.
• And less risk of missing a major retail order because a line went down on the wrong morning. Operations and maintenance leadership are the right people to put the final number on it, line by line.

Improved OEE on covered lines, reduced emergency corrective maintenance cost versus predictive. Estimated payback of 6 to 12 months against the first relevant stoppage avoided. Estimate to validate with the maintenance manager and operations leadership.

And the fair question from the production manager

“Do we have to deploy all twelve cases before this does anything?” — no, and framing it that way would be the way never to start. The rings are built in layers and most cases pay on their own. With line startup and the shift leader's headset you already have ring 1. Where rings cross-check, matching a symptom to an asset's history is an anchored task, where the best models drop below 1.5% error [1], and a person decides.

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