| |
By Bill Murphy · Powered by Colony Spark |
THE OPERATIONS
BRIEF |
|
ISSUE #15 · MAY 29, 2026 |
|
| |
|
| |
|
Hi there,
AI is moving out of the reporting layer and into the execution layer — and the demos at Hannover Messe this year made that unmistakable. The agents can read the ERP now. The question is whether anyone has built the layer that checks what the agent can’t see before it touches production.
This issue is about that missing step: the materials, the capacity, and the customer commitments that need to be verified before an AI-suggested schedule change ever reaches the floor.
— Bill
|
| |
ERP AI Needs a Production Reality Check.
The agents moved into the execution layer. The control layer hasn’t caught up.
AI is moving out of the reporting layer and into the execution layer. That was the clear signal from Hannover Messe this year. SAP did not show up to demo dashboards. It showed up with a Production Master Data Agent, a Production Planning and Operations Agent, a Field Service Dispatcher Agent, and outbound task orchestration that touches materials, capacity, and customer commitments directly (ERP Today).
Microsoft framed Dynamics 365 as an agentic ERP foundation for manufacturing decisions across planning, sourcing, production, fulfillment, and service (Microsoft). The Accenture and Avanade agentic factory announcement included diagnostics tied to FMEA records and maintenance history, with early validation from Kruger pointing to MTTR reduction as a multimillion-dollar lever when scaled across production lines (Accenture).
This is no longer a visibility play.
|
Where the market actually stands — MaintainX 2026 report
· 58% of maintenance and operations teams are already using AI.
· 75% report measurable ROI in under six months.
· 79% still saw unplanned downtime stay the same or increase (MaintainX).
Those numbers do not contradict each other. They describe the same problem. Teams have adopted the tools. The execution maturity to back them up has not kept pace.
|
AI can now read the ERP. The question that matters is whether it understands what it cannot see: the machine that went down at 11 PM, the supplier who confirmed delivery but is running two days late, the work order that looks schedulable because inventory shows 40 units and the planning team knows 30 of them are already committed elsewhere.
The companies pulling ahead are not the ones giving AI the most access. They are the ones building the operational control layer before the agents get to execution. Materials checked before the schedule moves. Capacity confirmed before the work order drops. Customer commitments reviewed before the promise changes. Approval routing that reflects risk level, not org chart convention.
Samsung’s 2030 factory AI strategy makes the sequence explicit: digital twin simulation and safety governance come before deeper autonomous execution. That is not caution. That is the right architecture (Samsung).
|
The market is not waiting on AI to catch up. It is finding out, sometimes expensively, that the hard constraint is not the technology.
|
Sources: ERP Today, April 2026; Microsoft Dynamics 365 Blog, April 16, 2026; MaintainX Newsroom, May 5, 2026; Accenture Newsroom, April 20, 2026; Samsung Global Newsroom, March 1, 2026.
|
|
| |
The agent needs a number to beat.
Eric Ashby, COO of Kruger, on the real value lever in agentic manufacturing
Eric Ashby is Chief Operating Officer at Kruger, an early validator of the Accenture, Avanade, and Microsoft agentic factory. His framing puts a number on why downtime is the lever everyone in manufacturing is chasing right now.
|
“Unplanned downtime impacts safety, productivity and performance across our operations. The financial and operational value lever is significant. A 10-15% reduction in mean-time-to-repair quickly translates into multimillion dollar savings when scaled across production lines and sites. With this agentic factory, we can help our teams respond faster to issues, capture operational knowledge, and continuously improve how our factories run.”
|
The operators who are going to get the most from agentic manufacturing systems are the ones who already measure what the agent is supposed to improve. MTTR. Downtime frequency. Schedule adherence. If you do not have the baseline number, the system has nowhere to anchor its value.
|
The takeaway
A 10-15% MTTR improvement only shows up if you were measuring MTTR to begin with. Before you bring an agent into production scheduling, write down the number it is supposed to move — otherwise you will never be able to prove it did.
|
|
|
| |
The Production Planning Preflight.
The missing step between the AI recommendation and the action.
Four checks that run before any AI-suggested schedule change reaches the floor.
The promise of ERP AI is that it will make better production decisions faster. The risk nobody is talking about yet is what happens when it makes them without checking.
Not maliciously. Not recklessly. Just without asking whether the materials are actually there. Whether the machine is actually available. Whether the promised ship date is still intact after the schedule moves. Whether anyone with authority approved it.
That is not a technology problem. That is a missing step between the recommendation and the action. The Production Planning Preflight is that step.
|
Four checks, in sequence
1. Material availability. Pulls current inventory positions for every material and component the affected order needs. If there is a shortfall, it returns the quantity, the gap, and the open PO that is supposed to cover it.
2. Machine capacity & labor. Checks machine availability and labor coverage against the proposed shift window. If the resource is down, over-allocated, or missing skills coverage, it flags the constraint and the expected resolution date.
3. Customer commitments. Reviews ship date, order promise, and any priority flags already in the system before the promise changes.
4. Approval routing. Checks whether the change requires approval at the current risk level, based on order value, lead-time impact, or customer tier.
|
Each check returns one of three statuses — clear, blocked, or review needed. If everything clears, the schedule change moves. If anything is blocked, the change stops and the planner sees the exact constraint before anything touches production. If a check comes back review needed, the system routes it to the right supervisor with the full context already attached.
The agent does not override. It does not guess. It does not assume the plan is clean because the ERP said so last Tuesday.
|
What it eliminates
The schedule change that looked valid at 8 AM and turned into a material shortage by noon. The shop floor that got a new work order without the materials to run it. The customer call where no one knew the ship date had moved until the customer asked.
|
|
What the operator gets
A documented preflight result for every AI-suggested change. A clear record of what was checked and when. A production schedule that reflects reality before it reaches the floor.
|
There are two paths to build this.
|
Path 1: Internal build
If your ERP exposes an API or scheduled-job capability, the preflight logic can run as a lightweight script that queries inventory positions, work-center calendars, and open customer orders before any planning-agent output gets written back. NetSuite, Dynamics, and SAP all have the data. The constraint is usually not the system — it is the decision to require the check before the action.
|
|
Path 2: ERP partner
The firms already working inside your planning and scheduling configuration are the right people to build the constraint-validation layer. The ones adding this kind of governance to their post-go-live work are the ones their clients are still on the phone with two years later, instead of filing implementation complaints.
|
|
Build the check before you need it. Once a bad AI-suggested schedule change makes it to the floor, the cost of the lesson is already committed.
|
|
|
| |
|
THE OPERATIONS BRIEF
By Bill Murphy · Powered by Colony Spark
|
|
Building the control layer before your agents reach production — or watching a client skip it? Hit reply, I read every one. Bill
|
|
| |
|