Written for Operations, quality, maintenance, technology and finance leaders in manufacturing, and the sponsors who have to fund a first case.
What's in this playbook
This 19-page playbook is a design method for operations, quality, maintenance, technology and finance leaders in manufacturing. The library has sometimes described it as “five deployment architectures” and a manufacturing return template. It does not contain five deployments we are reporting, and it will not invent a manufacturing-sector NPV, a downtime percentage, an OEE gain, or a payback period. The financial method is the method in the ROI guide in this library. This playbook will not stand in for it.
What it is, instead, is ten tests to apply before a manufacturing AI case is funded; five illustrative patterns — finance and administration as an honest first case, a quality-inspection assistant with a person on the reject, maintenance support read-only from your historian, an exception layer on automation you already run, and a document and SOP assistant — and a way to treat safety as a design input, not as a certification we will issue. We will not recommend, as a first case, an unsupervised shop-floor agent, or a program that promises to “optimize the factory.”
The regional manufacturer in the document is the same hypothetical firm as in the AI Readiness Assessment Framework and the 2026 diagnostic: Developing, overall 22, Process Automation a strength, with talent and governance as the constraints. The first case is the same: do not start with a shop-floor agent; raise the operator role and a proportionate governance path; reuse a pattern the finance automation already proved. It is labeled illustrative. It is not a client case study, and it is not a sample of the sector. A companion workbook holds the ten tests and the five patterns. A completed workbook is not a safety certification.
Safety, the boundary, and a first case that can be honest
The OT and IT boundary is named: what the system can reach, and what it cannot. We will not claim a certified connector to a named industrial vendor. Safety is a design input, not a late review: a person, and a gate, on anything that can stop a line, reject a part, or change a parameter. This is not a safety certification, and we will not certify that a system is safe to run on a line.
The source of truth is named — your SCADA, MES, or historian, as you run them. A write-back is a different design from a read. An operator, including on the floor, is named, and is not only the original builder. The mechanism matches the job: a determined workflow, an assistant, or an agent. An agent is not a promotion of a workflow. Read-only first; write-back only with a gate. A proposal is not an action until a named person approves it, in the system.
Hosting is decided before the case is funded. Shop-floor data may not be able to leave; the offline guide is the companion. Governance is proportionate to consequence: a quality-reject assistant and an invoice workflow are not the same consequence. Proportionate does not mean “none until later.” The first case reuses a proved pattern, or it says why not. A shop-floor agent is not a first case this playbook will recommend. Often P1, P4 or P5; P2 and P3 when the person, the gate and the source of truth are in the architecture.
Five illustrative patterns
Why each can be honest, when the tests pass. They are not five deployments we are reporting, and they are not a required sequence.
P1
Finance and administration, as an honest first case. The illustrative manufacturer already has finance automation in production. Reusing that pattern raises the operator role and a proportionate governance path on familiar ground, before anyone touches a line.
P2
A quality-inspection assistant, with a person on the reject. Quality is often where a plant wants to start. The person and the gate are architecture. This is not a certification that the system is safe to run.
P3
Maintenance support, read-only from your historian. A proposal, not an action, until a named person approves it. A write-back that can change a parameter is a different design.
P4
An exception layer on automation you already run. The workflow playbook is the companion. An agent is not a promotion of a workflow that already works.
P5
A document and SOP assistant. An internal, often read-only, assistant on a named corpus, with citations and permissions. The RAG guide is the companion.
A shop-floor unsupervised agent, or a program that promises to “optimize the factory,” is not a first case this playbook will recommend.
Ten design tests (excerpt)
A failed test on the boundary, on safety, on the operator, or on the first case is usually a reason to pause.
T1–T2
The OT and IT boundary is named. Safety is a design input, not a late review: a person, and a gate, on anything that can stop a line, reject a part, or change a parameter.
T3–T5
The source of truth is named, as you run it. An operator, including on the floor, who can explain and stop the system. The mechanism matches the job.
T7–T10
Read-only first; write-back only with a gate. Hosting is decided before the case is funded. Governance is proportionate to consequence. The first case reuses a proved pattern, or it says why not.
Get the guide and workbook
Enter your name, work email and company. The files appear on this page immediately. A backup download link is also emailed when email delivery is available.