← Case studies
Industry / 01 Generative AI Multi-agent Manufacturing

Manufacturing Copilot

A generative-AI assistant that lets a factory floor ask its own data a question, and get back the stoppage, the likely cause, and the fix, in plain English. One question fans out across MES, sensor logs, ERP and Salesforce, and fans back into a single answer.

My role
AI strategy & delivery lead
Built at
IBM Consulting
Client
Nestlé
Scope
3-plant pilot to 25 U.S. factories
Approach
Multi-agent LLM over live plant systems
Status
In production pilot

Confidentiality note. Proprietary figures are withheld. Screens, data and metrics shown here are sanitized recreations built for this portfolio, directionally accurate but not the client's actual numbers.

Walkthrough · one question, four systems, one answer
Sanitized demo · illustrative data
Manufacturing Copilot
connected MESHistorianERPSalesforce / CMMS
Line 3 at Plant 07 was our worst performer this week, what drove the downtime, and where should we fix it first?
Interpreting the question, then planning the work across four live systems.
The plan
  1. Scope to Line 3 · Plant 07 · last 7 days and pull every stoppage
  2. Rank causes by lost minutes, then quantify the cost impact
  3. Trace the top cause to its root step and the right fix owner
Orchestrator
Retrieval agent
MES + historian → stoppage log
Manuals agent
Vector DB → manual step + diagram
Signal agent
Sensor logs → anomaly trace
Analytics agent
ERP → cost & output impact
Action agent
Salesforce / CMMS → fix owner
Underneath, that one question fans out across four live systems and is joined on the fly.
one question plain English MES + Historian time-series · 2.4M points / day Sensor / OT logs 18k tags · 1-second cadence ERP 60+ tables · cost & output Salesforce / CMMS work orders · tickets one answer chart · cause · fix joined live across all four systems · nothing leaves the plant network
Line 3 lost 14h 20m this week, 38% above its 8-week average. The driver is the filler changeover, not the mechanical fault the shift logs assumed. The capper jam and labeler misfeed are secondary, and the misfeeds trace to a single roll lot.
Lost minutes by cause · Line 3 · 7d
Filler changeover
312
Capper jam
205
Labeler misfeed
148
Conveyor E-stop
96
Detector reject
61
Filler vibration · changeover window
Anomaly at 23:04, aligned with the purge step in the changeover, not a bearing fault.
Recommended: re-sequence the 23:00 changeover (≈ 3.5 h/wk recovered) and quarantine labeler roll lot L-22.
Certainly. 10 stoppages totaling 2h 14m across three shifts at Plant 07 last week, pulled live from the MES and joined with each operator's note.
FactoryLineShiftMachineMinCauseOperator note
Plant 0732Filler 1312Changeover"purge step ran long again"
Signal agent: 4 of 5 overruns this week begin at the same valve-purge step. Recommend re-sequencing the 23:00 changeover. Est. recovery ~3.5 h/wk.
Plant 0733Capper 2205Infeed jam"caps bridging at infeed"
Plant 0731Labeler 1148Misfeed"roll L-22 again"
Plant 0732Conveyor 396E-stop"belt slip, product backed up"
Plant 0722Detector 161False reject"metal detect over-sensitive"
Plant 0721Filler 248Seal temp"weak seals on the line"
Open the work order for the changeover fix, attach the evidence, and brief the night shift.
Work order #14207 · createdCMMS
Re-sequence the 23:00 filler changeover, valve-purge step. The four matched overruns and the vibration trace are attached. Priority: high · est. recovery ~3.5 h/wk.
Owner Night maintenance Due Next changeover Evidence 5 files attached
Synced to ERP Logged in Salesforce Assigned · nights
Done. The work order is open, the procedure change is drafted, and a shift brief went to the night lead.
Next: want me to alert you if the 23:00 purge step runs long again, or roll the same check out to Lines 1 and 2?
01 / 05
Step 01A single plain-English question is planned and routed across four enterprise systems by a team of specialized agents.

A designed walkthrough of the deployed assistant. The operator types one question; underneath, an orchestrator fans the work out to specialized agents across MES, sensor logs, ERP and Salesforce, then fans it back into a chart, a sensor read, a live table, and a work order, all in plain English.

01 · The problem

A fifth of production time disappears into stops nobody can explain.

On a connected line the data that explains a stoppage already exists. It is just scattered across the historian, the MES, fault codes, and a maintenance tech's free-text shift note. Unplanned downtime runs around 20% of total production time, and the people closest to it can't get answers, because the answers sit behind ERP screens and BI dashboards they were never trained to drive.

Off-the-shelf chatbots don't close the gap either. They reason over static text; they can't reach a live plant database, don't speak the plant's vocabulary, and can't take an action. The floor needed something that could read the live system and reason about it in the plant's own language.

02 · What I built

An assistant the floor can talk to in plain English.

I scoped and led the build of the Manufacturing Copilot: a conversational assistant with real-time access to live plant systems. An operator asks a question the way they'd ask a colleague; it works out what they mean, plans the analysis, queries the production systems itself, and hands back readable insight, a chart, a ranked cause list, a recommended fix, instead of a wall of SQL.

  • Reads the live systems, the historian, MES and shift logs in real time, not a static export.
  • Plans before it answers, breaking a question into a structured plan, then generating exactly the query it needs against the production schema.
  • Knows the plant's vocabulary, trained on how operators ask versus how the data is organized, so it scales past a hundred chained questions and supports real root-cause digging.

"It can generate and act on solutions by reading the latest operational data, a more direct way to cut downtime than any dashboard."

From the technical write-up
03 · The agents

One question, a team of agents, four systems.

The operator sees one plain-English answer. Underneath, an orchestrator breaks the question into a plan and hands the pieces to specialized agents that fan out across the plant's systems and fan back in:

  • Retrieval agent, reads production context from the MES and historian.
  • Manuals agent, runs semantic search over the machine manuals, diagrams included, to pull the exact procedure step behind a fault.
  • Signal agent, scans sensor and actuator logs for the anomaly behind a stop.
  • Analytics agent, joins output, cost and inventory from the ERP, and builds the chart.
  • Action agent, opens and tracks work orders in the CMMS and Salesforce, and writes the summary.

Because the agents are grounded in how the plant's data is actually structured, they can narrow a fault the way an engineer would, from "Line 3 keeps stopping" to "the 23:00 changeover's purge step." New data, like scheduling and forecasting, drops into the same frame, moving the system from explaining downtime toward heading it off.

The manuals are not just text. We ingest the OEM PDFs, photos and exploded diagrams, build multimodal embeddings over both the words and the images, and index them in a vector database. So when a fault traces back to a torque spec or a purge sequence buried in a diagram, retrieval-augmented generation pulls the right step, image and all, instead of guessing, keeping every answer grounded in the actual procedure.

04 · Solution design

One question, fanned across the whole plant.

It looks like a chat. Underneath, every question runs through an agentic platform, an orchestrator that plans the work and dispatches a team of specialized agents across four live enterprise systems, each speaking that system's own schema and protocol, then reconciles their answers into one grounded response.

OPERATOR ORCHESTRATOR AGENTS LIVE PLANT SYSTEMS ANSWER "Why did Line 3 stop this week?" plain English Orchestrator interpret intent plan · route · reconcile self-planning code-gen LLM Retrieval agent production context Signal agent anomaly detection Analytics agent join · cost · charts Action agent writes work orders MES + Historian time-series · batch Sensor / actuator OT logs · PLC ERP output · cost · inventory Salesforce / CMMS work orders · tickets chart + cause + fix one answer underneath it all → access control · guardrails · evaluation · human-in-the-loop · audit trail
The agentic platform. The operator types one sentence. An orchestrator plans the analysis and dispatches four specialized agents that fan out across MES, OT sensor logs, ERP and CMMS, each grounded in that system's schema, then fan their findings back into a single plain-English answer.
05 · The 0 → 1 plan

From a notebook to ten factories, in six months.

Month 1
Lay the tracks
  • SQL Q&A, text RAG
  • Data pipeline
  • Guardrails, instrumentation
★ Stakeholders aligned
Month 2
Build foundations
  • Multimodal RAG
  • Scenario planning
  • Early trust controls
★ Pilot users onboarded
Month 3
Proof in the plant
  • E2E reliability
  • Human-in-the-loop loops
  • Access control, CI/CD
★ MVP shipped
Months 4–5
Production & scale
  • Monitoring, scaling
  • Operational playbook
  • Security & RBAC
★ Operational at 3 factories
Month 6
Next gear
  • Predictive maintenance
  • Real-time sensor integration
  • Cross-factory optimization
★ Scaled to 10 factories
SpeedValueGovernanceEfficiencyScale
06 · Measuring value

Three layers of metrics, one question: is it worth it?

01
Operational & business impact
solution metrics
Downtime / factoryMTTREscalation rateDowntime % changeResolved issues / wkDaily active users
Fewer unplanned stopsLess raw-material wasteDirect cost savings
02
Trust, quality & decision confidence
model metrics
Context precision / recallHuman (SME) evalDeferral rateAnswer correctnessUser feedback score
Enough trust to act on answersFewer mistakes vs. manualConfidence to scale
03
System health & scalability
platform metrics
Query latency / agentError rate by componentSystem logsTime-per-stepCache hit rateTokens / query
Faster debuggingPredictable performanceControlled cost, clear ROI
07 · Leading it

I owned the strategy and led the team that shipped it.

This was a joint effort with the client's Innovation & Technology group. I ran the discovery that scoped where generative AI could actually help on the floor, designed the multi-agent architecture, then led a cross-functional team, ML engineers, data engineers, plant operations leads and enterprise IT, from prototype to a pilot running against live plant systems. A big part of the job was the unglamorous half: data access, guardrails, and a hand-off the plant could own.

08 · Impact

From three lines to a national rollout.

3 plants
live pilot, covering roughly 15% of the client's U.S. revenue
25
national factories targeted for rollout
2%
achieved reduction in unplanned downtime
8-figure
projected annual savings at scale (figure withheld)

Beyond the headline savings, the system compounds into labor efficiency, raw-material conservation and quality, an early, concrete step toward end-to-end automation across a connected plant.

09 · Talks & recognition

Where this work was shared.

NYC
Workshop hosted at the IBM Innovation Studio for P&G, Colgate & Unilever leadership
Workshop · Host
Invited
From Dashboards to Dialogue: Conversational Analytics for Manufacturing
Consumer Goods Forum, invited industry session
Invited talk
Conference
Multi-Agent LLMs for Operational Root-Cause Analysis
Retail Industry Leaders Association (RILA), technology track
Conference
View the project on GitHub ↗
Next · Industry / 02
Fleet Assignment Under Disruption