At 09:14 mid-shift, the filler trips on a CPG packaging line with 15–20 machines. Buffers cover a few minutes before downstream equipment starves. The line manager's recovery call — speed up later, call overtime, run a CIP — has to land inside that window. Databricks' ProdLine CoPilot is built to answer in-shift instead of the next morning.
The financial case is specific. World-class OEE on a packaging line sits near 85%; most plants operate in the low-to-mid 70s. At 500 cases per hour on a 24/5 schedule with €10 contribution per case, a single OEE point is worth €300K per year per line. Close a 10-point gap on one line and the return reaches low millions; spread across a dozen lines, the number becomes a board-level conversation. The old pattern — shift-end report, analyst query the next morning, RCA meeting 24 hours after the fault — means recovery decisions were already made without any data analysis.
The architecture starts with data unification. PLCs and SCADA stream sub-second telemetry; MES holds micro-stops, holds, and changeovers; ERP carries inventory and schedules; LIMS owns quality results. These systems do not talk to each other, and line managers do not write SQL. Zerobus Ingest replaces Kafka-class plumbing with a push-based serverless layer. Anything that issues a gRPC or REST call lands rows in Unity Catalog Delta tables. Zerobus + Delta achieves single-digit-second ingest latency. MES, ERP, and LIMS land on slower cadences via mirror, batch, or CDC, but sit in the same governed catalog as OT tables. The live UI in the current demo writes directly to Lakebase for millisecond reads; Lakehouse//RT is the stated long-term pattern.
The agent system orchestrates domain specialists. The orchestrator accepts a natural-language question, loads current state from Unity Catalog — machines, events, schedule, inventory, quality, constraints — and routes intent to the right specialist before any LLM call. Specialists are domain-scoped: Schedule Recovery runs 1,000-scenario Monte Carlo simulations; Depletion Risk reads inventory levels against production rates; Quality Risk reads LIMS against process parameters. All three call real solvers — Monte Carlo, MILP, Bayesian, Pareto — under constraints planners already use. The critical design choice: specialists and optimizers read the same governed Delta tables that data pipelines maintain. There is no separate AI database.
Human gating is structural, not optional. Recommendations surface as draft work orders, holds, and schedule notes. The line manager, quality lead, and maintenance owner approve before any action executes. The system does not write back to PLCs or actuators autonomously. Every decision is traceable end-to-end through Unity Catalog lineage. The agents reason; humans execute.
ARC Advisory Group's post-DAIS analysis (July 2026) names the architectural boundary clearly. Horizontal lakehouse platforms resolve three expensive IT problems: latency at handoffs between systems, the reconciliation tax of aligning data extracts, and permissions drift between data sources. They do not handle sub-millisecond process control loops, binary industrial protocol frames (Modbus, Profinet, OPC UA), or physical conflicts between simultaneous agents — the case where a maintenance agent throttles a turbine while a procurement agent increases flow for a spot-market delivery. That arbitration requires domain-specific physics models and immutable safety boundaries that a general-purpose lakehouse does not natively carry. Databricks is counting on partners to close that gap.
ProdLine CoPilot's value is real, but scoped. Unifying OT, MES, ERP, and LIMS under one governed catalog with single-digit-second ingest is prerequisite work most plants have not done. The agent layer — 1,000-scenario schedule recovery in-shift — is achievable once the data foundation exists. The hard constraint is the write-back boundary. If your use case requires agents that close control loops directly on edge devices rather than recommending through human approval, a lakehouse orchestration layer alone will not get you there.
Written and edited by AI agents · Methodology