Data Academy · Tutorial 5 of 10 · Manufacturing

Governed AI in manufacturing

Manufacturers are moving AI from dashboards into the work itself: agents that turn a condition alert into a planned work order, put a drifting lot on hold, or test whether a rush order can be promised. On a shop floor a wrong action stops a line or ships a bad part, so what decides whether AI scales is not the model but whether every action passes through the plant’s own rules, permissions and quality system.

01 What is changing

Where AI in manufacturing is heading

From prediction to action: the bottleneck is no longer spotting a failing bearing or a drifting process, it is everything that has to happen across CMMS, ERP, MES and QMS once the signal arrives.
From scattered pilots to a few scaled domains: maintenance, quality and order exception handling are where agentic work is moving into production, because the data and the decision rules already exist.
From plant-by-plant data to a shared model of the operation: agents fail when asset, part and lot data mean different things in each plant, so a common business model of the factory comes first.
From full autonomy to graded autonomy: plants are deciding explicitly which actions an agent may release on its own, which it may only propose, and which stay with a named engineer.

02 Use cases

Three use cases on the governed path

Each use case runs the same path: a business question, governed context, a deterministic rule, specialist agents, a policy check, an action and a record. How the path works →

Illustrative: names and figures are invented to show the flow.

Use case 1

Condition-based maintenance into the next planned stop

A critical asset showing early wear is usually caught by monitoring, then lost in the hand-offs between maintenance, stores and production planning. Closing that loop means fewer unplanned stops without adding stops of its own.

“Press line 3 is vibrating above alarm: can we fix it in this weekend’s changeover, or do we need to stop the line?”Asked by a maintenance planner

Context: CMMS: Press line 3 main drive, criticality A, main bearing last replaced 14 months ago
Condition monitoring: vibration velocity at 7.1 mm/s against a 7.0 mm/s alarm, rising steadily for 9 days
MES / APS: next planned stop is Saturday’s changeover in 4 days, a 6-hour window
ERP stores: 2 matching bearings on hand; a qualified fitter is rostered on Saturday
Rule: If a criticality A asset is above vibration alarm, a planned stop falls within 7 days, and the part and a qualified technician are available, schedule the replacement in that stop; otherwise escalate to the maintenance manager.
Decision: Replace in the next planned stop Severity: Medium
Agents: Condition agent Reads the vibration trend and fault frequencies and explains that the signature points to bearing outer-race wear rather than misalignment.
Planning agent Finds the Saturday window in the production plan and checks the job fits inside it with the fitter’s other tasks.
Parts agent Reserves one bearing and the seal kit in stores so the job is kitted before the stop starts.
Policy: Planners may schedule work orders inside an existing planned stop without further approval. Anything that extends a stop or changes the production schedule needs the production manager. The job carries a lockout/tagout permit, issued on site by an authorised person.
Action: Work order created in the CMMS and released automatically into Saturday’s stop, bearing kit reserved in ERP, production schedule unchanged.
Data: Asset register and maintenance history (CMMS)Condition-monitoring and historian dataProduction plan and planned stops (MES / APS)Spares stock and reservations (ERP)Technician skills and rosters

Use case 2

Process drift on a special characteristic

When a critical dimension starts to drift, the cost is decided by how fast the affected lots are found and stopped before they ship. Fast containment means fewer escapes, fewer customer complaints and less scrap.

“CNC cell 5 is drifting on the bore diameter: which lots are affected, and is anything on its way to the customer?”Asked by a quality engineer

Context: QMS control plan: housing H-220 bore diameter is a special characteristic, 42.000 mm plus or minus 0.015 mm
SPC: the last 8 subgroups sit above the centre line; Cpk has fallen from 1.6 to 1.1
MES genealogy: lots L-0912 and L-0913 produced in that window, 1,800 parts, 600 already in finished goods
WMS: none shipped yet; next dispatch to Northfold Axle Works is in 2 days
Rule: If a special characteristic shows a run of 8 or more points on one side of the centre line, or Cpk falls below 1.33, hold every lot made since the last good check and raise a nonconformance for disposition.
Decision: Hold lots and refer for disposition Severity: High
Agents: SPC agent Confirms the run rule and capability breach on the cell and shows when the drift started.
Traceability agent Walks the lot genealogy from the cell to finished goods and the pending dispatch, so nothing affected is missed.
Root-cause agent Links the drift to the tool-change counter and to similar past events on the cell, and proposes a likely cause for the engineer to confirm.
Policy: Agents may place a quality hold, because it is protective and reversible. Only a quality engineer may disposition the lots (rework, scrap or use as is). Use as is on a special characteristic needs the customer’s written deviation approval, as the customer-specific requirements under IATF 16949 expect.
Action: Quality hold placed on both lots in MES and WMS, the dispatch line held, and a nonconformance raised in the QMS and referred to the quality engineer for disposition.
Data: Control plans and special characteristics (QMS)SPC measurements by cell and toolLot genealogy and work orders (MES)Finished goods and dispatches (WMS)Customer requirements and deviation history

Use case 3

Rush order promise against capacity and margin

Expedited orders are often accepted on instinct and paid for in overtime, freight and other customers’ late deliveries. A governed check protects both service and margin before anything is promised.

“Can we promise Brindlemoor Fixings 2,000 brackets by Friday without hurting other orders or losing money on it?”Asked by a customer service manager

Context: ERP order entry: request for 2,000 brackets due in 5 days, order value £24,000 at standard price
APS: line 2 can make 1,200 in regular shifts this week; the other 800 need one Saturday overtime shift
Costing: contribution at standard is £5,800; overtime and expedited steel freight add £2,900, leaving £2,900
Commercial policy: expedited orders need a contribution floor of £3,600; no other order slips if the overtime runs
Rule: Confirm automatically if the order fits regular capacity and contribution stays at or above the floor; if it needs overtime or falls below the floor, hold it for approval.
Decision: Accept only with approval Severity: Medium
Agents: Capacity agent Tests the order against the finite schedule and shows that regular shifts cover 1,200 and that no existing order moves.
Margin agent Builds up the cost of overtime and expedited freight and shows the contribution falling below the floor.
Customer agent Reads the account history and terms and proposes an alternative: 1,200 by Friday and 800 the following week at standard cost.
Policy: Customer service may confirm orders within standard lead time and margin. Overtime needs the plant manager, within working-time limits. Contribution below the floor needs the commercial manager.
Action: Order entered in ERP as pending approval with two options (full quantity with overtime, or split delivery); no delivery promise sent to the customer until it is approved.
Data: Sales orders and customer terms (ERP)Finite capacity and schedules (APS / MES)Standard costs and freight ratesMaterial availabilityCommercial approval limits

03 The foundation

What the agents need to understand

Core entities in the ontology

PlantLineAssetWork orderPartBill of materialsLotCharacteristicCustomer orderSupplier

Systems they come from

ERP: orders, bills of materials, stock, costs and purchasing
MES: work orders, lot genealogy, production and holds on the floor
CMMS / EAM: assets, criticality, maintenance history and work orders
QMS: control plans, nonconformances, CAPA and deviations
APS: finite capacity, schedules and planned stops
Historian / SCADA: sensor, vibration and process data over time

04 Guardrails

The controls that let it scale

Read-only on the control layer

Agents read from the historian and SCADA but never write set points or start and stop equipment; that stays with operators.

Holds are automatic, releases are not

An agent may place a protective quality or stock hold, but only a named engineer may release or disposition it.

Safety permits stay human

Lockout/tagout and permit-to-work are issued on site by an authorised person and are never created or closed by an agent.

Customer and regulatory approvals respected

Use as is on special characteristics, and changes to validated processes, follow the customer’s deviation and change approval rules.

Every run recorded

The signal, rule, agent explanations, approver and resulting transaction are kept together so quality and audit can replay any decision.

05 Rollout

From the first use case to many

1
Pick one domain with clean signals
Start with maintenance on criticality A assets, where condition data, work orders and spares already exist and a wrong recommendation is cheap to correct.
2
Model the plant before the agents
Define assets, lines, parts, lots and orders once in the ontology, mapped to each plant’s ERP, MES and CMMS codes, so the same rule means the same thing everywhere.
3
Write the rules and autonomy levels down
Agree with maintenance, quality and production which actions release automatically, which wait for approval and which are referred, before anything goes live.
4
Run in shadow, then release
Let agents propose for several weeks while planners act as usual, compare the two, then allow automatic release inside the agreed limits.
5
Extend to quality and order exceptions
Reuse the same model and audit trail for quality holds and order promising, then carry it to the next plant.

06 What to measure

Outcomes, not activity

Unplanned downtime on critical assetsPlanned maintenance complianceTime from signal to contained lotCustomer escapes and complaintsOn-time in-full deliveryContribution on expedited orders

07 Pitfalls

What usually goes wrong

Agent on raw plant data. Asset and part codes differ from plant to plant, so an agent reasoning on raw tables gives answers that do not travel; map them to one model first.
Prediction without follow-through. A good failure model achieves nothing if the work order, parts and stop window are still arranged by email; govern the whole path to the action.
Autonomy granted by default. Letting an agent change schedules or release holds because it can is how a line stops or a bad part ships; set autonomy per action, not per agent.
Pilot that never meets the MES. Proofs of concept on exported data stall at integration; build the first use case on the live systems and their approval rules from day one.

08 Diagnostics

Questions to ask your team

1
Which of our maintenance, quality and order decisions already follow a written rule, and which depend on one person’s judgement?
2
Do asset, part and lot codes mean the same thing in every plant, and who owns that model?
3
Which actions would we allow an agent to release without approval, and who signs that list?
4
If a customer asked why a lot was released, could we show the data, rule and approver in one place?

09 Keep going

Frequently asked

Does the agent control the machines?

No. Agents read from the historian and production systems and create transactions such as work orders, holds and pending orders. Control of equipment stays with operators and the control layer.

Why use a fixed rule when a model could decide?

Rules are what maintenance, quality and customers have already agreed to, and they can be audited. The agents add the explanation, the context and the options; the rule decides the outcome.

Can this work across plants with different systems?

Yes, if each plant’s ERP, MES and CMMS codes are mapped to the same business model. That mapping is the main piece of work and the reason later plants go faster.