AI agents in manufacturing: applications and practical pilots

AI agents can help manufacturing teams turn disruptions into informed decisions. Learn where they add value, what evidence to check, and how to start a pilot that keeps your team in control.

A supplier tells you a component will arrive late. Do you change the production sequence, assess a substitute, or leave the plan unchanged?

Before you can decide, you need to know which work depends on that shipment. Purchasing has the supplier’s update. Planning has material requirements and scheduled orders. Engineering and quality hold the evidence for whether an alternate can be used.

Meanwhile, the customer-facing team needs to know whether a delivery commitment is at risk. AI agents in manufacturing can help assemble that evidence, identify gaps, and prepare options through approved connections to those records.

You’ll see where agents can help, what their recommendations need, and how to scope a practical pilot. The goal is to make the next decision easier without handing over control of the production plan.

TLDR

• Agents can connect material, production, maintenance, and quality information to prepare work for the people making manufacturing decisions.
• Useful outputs include supported dependencies, maintenance packets, alternatives, and customer-update drafts, with missing evidence made explicit.
• Start with one measurable workflow, reliable sources, and clear responsibilities before widening the agent’s access or actions.

What are AI agents in manufacturing?

AI agents in manufacturing are AI systems that interpret goals, use connected data, and coordinate tasks through permitted tools. They can gather production records, investigate material shortages, prepare scheduling recommendations, or carry out approved business-system actions.

Where can AI agents help manufacturing teams?

AI agents can take on the information gathering and coordination around production, maintenance, quality, and customer delivery. The useful application is specific: a defined task that produces something another person or system can act on.

The applications below describe industry possibilities, not a feature list for a particular platform. Each requires suitable interfaces, reliable records, and an operating design for the particular plant.

IBM’s manufacturing explainer covers material disruptions, maintenance coordination, scheduling, and engineering work. Its scenarios illustrate possibilities; they do not establish measured results for your operation.

Material shortages and supplier coordination

An agent can match a supplier notice to purchase references, component requirements, and production orders, then compare required quantities and dates against usable supply to highlight the work that may be affected.

The output should be a shortage brief: matched orders, dated supply assumptions, supported dependencies, and unresolved questions. That gives purchasing and planning a shared starting point instead of separate interpretations of the same notice.

Usable supply isn’t the same as stock on a screen. Reservations, quality holds, expected receipts, and the time needed to receive and release material can change what is available for production.

A proposed supplier alternative also needs evidence. Price and arrival date alone don’t establish that a different component meets the product’s requirements.

Production-planning support

Here the agent assembles current material constraints, work-order dependencies, and approved planning information, then compares documented alternatives and explains which assumptions need a planner’s decision.

Its output might list work unaffected by a shortage, orders needing review, and candidate changes for the planning process. That helps the planner focus on the decisions that matter.

This is different from generating a feasible production schedule. Feasibility can depend on machine capacity, tooling, labor, changeovers, routing, and other constraints the agent hasn’t validated.

In Deloitte’s manufacturing roadmap, Patricia Henderson and coauthors describe proposed schedules reviewed by production planners. Treat that as implementation guidance, not proof that a language model or a particular product can optimize your schedule.

Maintenance preparation and coordination

For a maintenance task, the agent gathers the relevant equipment history, approved procedures, spare-part information, and current request, then prepares a technician’s work packet and identifies scheduling dependencies.

The packet should distinguish a reported symptom from a confirmed diagnosis. It should also identify the procedure version and any missing equipment or parts information.

The value is better-prepared maintenance work: the technician sees the relevant context without searching each source separately. Deloitte’s maintenance example describes this coordination around a technician’s task, not an independently verified reduction in downtime.

Preparing maintenance information does not authorize equipment control. Physical work, isolation procedures, and return-to-service decisions remain subject to the plant’s established operating responsibilities.

Quality and engineering-change coordination

For a proposed production change, the agent gathers the configuration, revision, change record, and applicable quality evidence, then identifies which instructions or work orders need review and where records disagree.

For a substitute component, the packet should establish applicability to the intended configuration and effectivity. Effectivity means the dates, serial ranges, or other conditions under which a change applies.

Then check the specific material. An approved component design does not release every lot from inspection or quality hold.

A pre-approved alternate doesn’t necessarily need fresh engineering approval for every use. If its documented conditions are satisfied, follow the existing substitution procedure. Route exceptions or unclear applicability to the responsible owner rather than inventing another approval step.

This separates two useful outputs: evidence that an existing rule applies, and a clearly defined exception that needs a decision.

Customer-commitment updates

An agent can bring together an approved production decision, customer-order details, and the communication requirements for that order, then draft an update showing what changed and which commitment the team has authorized.

That gives the customer-facing team an evidence-backed message instead of a supplier estimate copied into a customer promise. Internal planning uncertainty can remain explicit while the team resolves it.

The draft should contain only information appropriate for its recipient. A quality investigation may inform a decision without belonging in the customer’s message.

AI knowledge management supports these cross-team relationships. The manufacturing benefit comes from connecting the relevant decision evidence, not simply returning more documents.

How can AI agents help teams manage a component delay?

An agent can turn a delay notice into a decision packet that identifies affected work, usable alternatives, and the next owner. Here’s a discrete-manufacturing example with illustrative records and quantities, not a reported factory result.

A plant assembles finished units from individual components. Work order WO-240 needs 120 units of component C-17 by Friday, with one component required per finished unit. A supplier moves a 60-unit receipt to Monday.

Identify which production work may be affected

The agent matches the notice to the purchase reference and WO-240’s component requirement. It finds 40 released, unallocated units and another 20 expected to be usable before Friday.

That gives the planner 60 units against a requirement of 120. The calculation excludes stock reserved elsewhere and material not released for this use.

The agent marks the 20-unit receipt as an assumption until receiving confirms the usable date. It doesn’t describe all work mentioning C-17 as affected: only supported, current dependencies enter the packet.

Prepare alternatives for review

The engineering record already permits alternate C-17B for this configuration during the relevant effectivity period. Its documented usage quantity is also one per finished unit.

The agent finds 60 units of C-17B, but 20 remain on quality hold. Only 40 qualify as usable in the current comparison.

Oracle’s substitution documentation illustrates why substitution start dates and usage quantities matter when a substitute is configured. Those are Oracle-specific mechanics, not a universal substitution rule or evidence of another product’s integrations.

The resulting packet is concrete:

In short: the agent gives the team an explainable shortage and reviewable alternatives, not a confident but unsupported production plan.

Clarify the customer commitment and finish the decision

The planner checks capacity and the order’s delivery conditions. In this example, splitting the order would not meet the customer’s batch requirement.

Purchasing and receiving confirm that Monday’s original-component supply can be usable before a Tuesday production slot. The planner validates that slot, and the responsible owner authorizes the revised plan.

The agent combines the approved plan with the customer order and drafts the update:

“Your full order is now scheduled for Tuesday production. The shipping date is still being confirmed.”

The draft distinguishes an approved production plan from an unconfirmed shipping commitment. It doesn’t depend on release of the held alternate lot.

The example ends with a supported next action: the customer-facing owner has a draft based on the accepted plan. The agent has connected the evidence and prepared the handoff; the planner has established feasibility.

What challenges should you address before implementation?

The main challenges are source reliability, integration constraints, operating authority, and adoption. Group them into the pilot design rather than adding a separate warning to every application.

Reliable records and usable connections

Define which source supports each decision. Enterprise resource planning, or ERP, manages business resources and transactions. A manufacturing execution system, or MES, supports production execution. Product lifecycle management, or PLM, can hold engineering definitions and changes.

Actual record placement depends on your installation. A supplier notice doesn’t inherently live in PLM, and an engineering record doesn’t establish inventory availability.

Check identifiers, timestamps, revisions, and ownership before connecting the workflow. Agent data readiness covers those foundations; manufacturing exception management needs them tied to the specific production decision.

Confirm which interfaces are available and supported. Industry applications do not establish native ERP, MES, or PLM connections in Computer, by DevRev. Product-specific capabilities require their own validation.

Access across the IT/OT boundary

Operational technology, or OT, monitors or controls physical processes. The IT/OT boundary therefore needs assessment of the actual information path, system load, and consequences of failure.

The NIST Guide to OT Security, Revision 3, published in 2023, addresses performance, reliability, and safety requirements. It is foundational OT guidance, not an agent-specific law or approval rule.

Read-only access still needs reviewed permissions, fields, query behavior, and handling rules. It can disclose restricted information even when it cannot change equipment.

For this pilot, use an owner-approved business or planning view and exclude machine-control commands. Apply appropriate agent sandboxing rather than relying on an evaluation score to contain execution.

Decision ownership and recovery

Define which recommendations can follow an existing approved procedure and which exceptions need review. Attach decisions to the relevant evidence and recheck when that evidence changes.

If a later phase allows a business-system update, specify completion evidence and recovery ownership first. A record update may succeed while a notification fails.

Inspect actual state before retrying. Stopping further activity doesn’t reverse an earlier change, and a human approval doesn’t create missing technical permissions.

Adoption and operating cost

A useful packet must fit the planner’s or technician’s work. Ask whether its evidence is easy to inspect and whether it creates additional review effort.

Include source maintenance, exception handling, human review, and integration support in the pilot’s cost. Those responsibilities belong in the build-versus-buy decision, not just the initial connection estimate.

Measure benefits before claiming them. Evidence completeness, unsupported recommendations, reviewer effort, and time to an accepted decision are possible measures, not promised improvements.

How do you start a practical manufacturing pilot?

Choose one application with a named recipient and a checkable output. For the component-delay workflow, that output is the decision packet, not an autonomously revised production schedule.

Agree its required evidence and compare it with the current process over a defined period. Include wrong revisions, stale schedules, held lots, ambiguous references, and missing receipts in testing.

Record accepted packets, rejected recommendations, unresolved evidence, and review effort. Explain exclusions so an easy test set does not stand in for the full workload.

Deloitte’s roadmap recommends focused trials that examine operating value, adoption, and compatibility with existing processes. Expand only after the people using the output can explain what helped and what still failed.

These answers distinguish useful assistance from assumptions that need a separate engineering or operating decision.

Start with the manufacturing decision that currently sends people searching across systems. Give its owner one packet with supported dependencies, alternatives, unresolved evidence, and a recommended next step.

AI agents in manufacturing earn a wider role when their output helps your team make the decision. The number of connected systems is secondary.