Coordinating agents across plant, planning, and paperwork

The challenge

A representative multi-site manufacturer has three operational problems that are usually solved by three separate projects — and are in fact the same problem.

Asset health lives in historians, planning lives in the ERP, and supplier paperwork lives in mailboxes and a document store. Nobody sees them together, so a machine warning does not reach the production plan, and a late supplier confirmation does not reach the maintenance window. Each system is individually fine; the coordination between them is manual.

This is not a sixth practice area. It is what happens when the maintenance, automation, and document patterns we already build are composed into one engagement — which is also where most agent programs fail. Gartner expects more than 40% of agentic-AI projects to be cancelled by the end of 2027, and MIT NANDA found roughly 95% of enterprise GenAI pilots showed no measurable P&L return.

Setting: Representative — multi-site manufacturer
System concept: Custom Multi-Agent System
Illustrative duration: 20 weeks

The architecture

We build a custom multi-agent system that composes the patterns from three of our practices: condition monitoring on the asset side, exception-aware process automation on the planning side, and document extraction on the supplier side. A coordination agent gives them one shared view of state, so an asset warning can reach the production plan and a supplier confirmation can reach the maintenance window. Every cross-system action is a proposal with the triggering evidence attached, routed to the human who owns that decision today.

ROLE / 01

Monitoring Agent

Tracks asset condition, order status, and supplier confirmations across the systems that already hold them

ROLE / 02

Prediction Agent

Forecasts asset and schedule risk from historical and live data, with confidence attached to each signal

ROLE / 03

Optimization Agent

Proposes revised plans and maintenance windows, showing the trade-off behind each option

ROLE / 04

Coordination Agent

Maintains one shared view of state and routes each proposal to the human who owns that decision

A proposed delivery path

This sequence illustrates the engagement. The actual scope, timeline and quotation are agreed around your requirements.

1-4

Discovery & Scoping

Mapped the operational handoffs between plant, planning, and supplier paperwork, and identified the data sources and integration points

5-8

Architecture Design

Designed the multi-agent system, the inter-agent protocols, the shared state model, and the escalation rules

9-14

Core Development

Built the agent system and integrated it with the ERP, historians, and document store; models tuned and evaluated on the client's own historical corpus under a signed data agreement

15-18

Pilot & Evaluation

Ran a single-site pilot in shadow mode, scoring proposals against what the planning and reliability teams actually decided

19-20

Handover & Rollout

Extended to remaining sites with dashboards, an operator runbook, alert thresholds, and a documented rollback path

Benchmark context

Published figures used to frame this scenario. Baselines describe the referenced setting; targets are illustrative, not measured project outcomes.

Illustrative target20–50%

Less planning and coordination time

Source: Deloitte — predictive maintenance planning band

Illustrative target70%

Faster processing on the document-heavy flows

Source: Deloitte — IDP, 60-80% less processing time

Illustrative target40%

Less unplanned downtime on monitored assets

Source: McKinsey — 30-50% range for condition-based programs

Published baseline>40%

Of agentic-AI projects are forecast to be cancelled by 2027

Source: Gartner, 2025 — poll of 3,400+ respondents

Deliverables & controls

What the scope can include

A coordinated agent layer running across your ERP, historians, and document store
Shared state model, inter-agent protocol map, and documented escalation rules
Evaluation suite scoring agent proposals against your teams' recorded decisions
Operator runbook, alert thresholds, failure-mode plan, and a documented rollback path
All source code in your repository, with the deployment owned by your team

Where people stay in control

Cross-system actions are proposals. The person who owns that decision today still makes it.
Every proposal carries the triggering signal and the trade-off, so it can be overruled with the evidence visible.
The pilot runs in shadow mode and is scored against real decisions before anything is put in the path of production.
Agents read from the systems of record. Writes are scoped, logged, and reversible.

Sources behind the scenario

>40% of agentic-AI projects forecast to be cancelled by end of 2027
Gartner, 2025 (poll of 3,400+)
~95% of enterprise GenAI pilots show no measurable P&L return
MIT NANDA, 2025
Planning time -20-50% for condition-based maintenance programs
Deloitte
30-50% less unplanned downtime for condition-based programs
McKinsey
IDP: 60-80% less processing time
Deloitte