Manufacturing AI agent · Maintenance

Maintenance-Assistant AI Agent (Work Preparation)

Turn a condition signal or an operator report into a work request with the asset history, the current manual revision and a parts list attached — a planner schedules it, a supervisor authorises it.

Typical delivery: 4–6 weeks
Deployment: Your stack
Work authority: Never the agent
After launch: Agent Care

What this agent does

Prepares the work, does not authorise it

01
Take condition alerts, operator reports, inspection findings and PM due dates from the systems that raise them.
02
Pull that asset's own record — hierarchy position, configuration, past work orders and the manual revision in force.

Reason

03
Trend the signal against the asset's own baseline and mark condition that is drifting, not only condition that alarms.
04
Draft a job plan — task steps, trades, estimated hours and the isolation points the plant's own procedure names.
05
Draft a parts list from the bill of materials as fitted, marking what a rebuild changed since the last job.

Decide

06
Separate a probable instrument fault from a probable asset fault, and say which one the request rests on.
07
Match the fault against open and recently closed work orders on the asset before a second request is raised.

Out

08
Present a draft request with the evidence, the job plan, the parts list and a window production has not committed.
09
Retain the signal, the sources read, the revisions cited and every correction a planner or technician made.

Product statement

The agent prepares and proposes. A planner schedules the job, a supervisor authorises it, and a technician works under the plant's own isolation and permit procedures.

Example workflow

One work request, end to end

Agent

Human

1Signal receivedA condition alert, an operator report, an inspection finding or a PM falling due
2Asset identifiedHierarchy position, configuration as fitted, criticality, past work orders and the manual revision in force
3Fault characterisedThe signal trended against this asset's own baseline, and the instrument checked before the asset is blamed
4Work preparedDraft job plan, trades and hours, parts against the bill of materials, and the isolation points the procedure names

No human action required

Stages 1 to 4 run without a person in the loop — the signal, the asset record and the draft plan are all assembled before a planner is asked to read anything. Anything safety-related ends that stretch on the spot.

5DecisionSplits on confidence and on whether the fault is duplicated, safety-related or on a critical asset

Routine, single request

Enters the planned backlog with its evidence.

Safety-related, critical or duplicated

Goes to the maintenance supervisor first.

Planner and maintenance supervisor

Reads the evidence, the job plan and the parts list, then plans the job, schedules it against production and authorises the work.

Plan the job · Correct · Reject

Planned — handed back

6Request filed for planningWritten to the maintenance system as a request only — not scheduled, not authorised, not closed by the agent
7Outcome evaluatedWhat the technician actually found, the parts actually used, schedule conflicts and the duplicate rate

Corrections

What the planner rewrites in the job plan is counted in the evaluation.

What should not run autonomously

Human approval stays in control

Outside the boundary — human approval required

8 items

Authorising work on equipment.
- Issuing, transferring or closing a permit to work.
- Applying, removing or overriding a lock or tag.
- Declaring an asset safe to work on.
Automation boundary
Agent acts unaided
- ✓Turn a condition signal or an operator report into a draft work request.
- ✓Attach the asset history, past repairs and the manual revision in force.
- ✓Draft the job plan and parts list, citing the isolation points the procedure names.
- ✓Flag drifting condition, probable instrument faults and duplicate requests.
The agent writes the work request and the draft plan. The authorisation, the permit and the lock are not its to write.
- Closing a work order as complete.
- Committing a job into a production window.
- Changing an asset's criticality or PM interval.
Moving an asset to run-to-failure or off monitoring.

Example output

One condition signal, annotated

Everything the agent proposes is attached to the exact asset, work order and manual revision it was read from.

Work-preparation output · one condition signal

Illustrative example

Signal source: Condition system and operator
Asset: Pump P-214, criticality B
Manual revision: Rev 4, in force
Raised as: Draft request — bearing wear
Confidence: 88%
Work authority: Never the agent

As received

The signal, the asset it names and the revision in force at the time — read from the condition system, not inferred.

Evidence used

Trend vs. own baseline
Two prior bearing repairs
Rev-4 manual, task 6

Why it was raised

The trend moved on this pump's own baseline, and the transmitter checked out.

What the score decides

Confidence decides how hard the planner checks, not whether work is authorised.

Action

Plan the jobCorrectReject

Value

Where AI adds value

The same four claims, placed at the point in the workflow where each one applies.

Where the value lands

Value 01 – 04

Every signal and report

Condition system, operator, inspection or PM due

03Prepare & propose

Work from the asset's own record

Use the plant's asset hierarchy, its work-order history, the manual revision in force and the bill of materials as fitted.

01Approved path

Hand the planner a job already prepared

The history, the manual, the past repairs and the parts arrive with the request, so planning starts from evidence rather than a one-line fault report.

02Human review

Send the doubtful ones to a supervisor

Safety-related, critical-asset and duplicated requests reach a maintenance supervisor before they enter the backlog, instead of queuing behind routine work.

Build an evidence trail

Retain the signal, the sources read, the revisions cited, the parts proposed, the confidence and what the technician actually found — on both paths.

Integrations

Typical integrations

Five system groups connect to the same agent. Which of them are in scope is decided in discovery.

CMMS & EAM
- SAP PM
- IBM Maximo
- Infor EAM
- Oracle eAM
- Plant CMMS
Condition monitoring
- Vibration
- oil analysis
- ultrasound
- Thermography
- motor testing
Historian & control
- PI System
- Aveva historian
- Rockwell FactoryTalk
- Siemens Opcenter

Agent

Maintenance assistant

Reads the asset record
Drafts the job plan
Routes to the planner

Documents & parts
- OEM manuals
- drawing revisions
- Bills of material
- storeroom spares
Observability & evaluation
- OpenTelemetry
- Langfuse
- Supported monitoring/evaluation sources

Integration availability depends on the client's existing systems and API access.

Agent controls

Six layers between the model and the job

Each control wraps the one inside it. A request clears every layer before a planner sees it, and the authority to work on the equipment sits outside all six.

L6 · Outermost — last line of defence

Inward → L1 · closest to the model

L6Rollback / safe modeReturn work requests to your existing route if evaluation signals degrade.Roll back

L5TraceabilityRecord the signal, sources read, revisions cited, parts proposed and every correction.Record

L4Authorisation gateScheduling, authorisation, the permit and the lock stay with the people the procedure names.Gate

L3Duplicate matchOpen and recent work orders on the same asset are matched before a second request.Merge

L2Revision and fitmentNothing is cited from a superseded manual or a part the asset no longer carries.Withhold

L1Instrument checkA signal is tested against its own transmitter before it is blamed on the asset.Verify

Model coreDraft request produced — fault, evidence, job plan, parts list, proposed window and confidence

L1 – L2: Keep the request about the right thing
L3: Decides whether this is already known
L4 – L5: Keep authority with a person, trail intact
L6: Pulls automation back when signals degrade

How Nestack evaluates it

Evaluate the whole preparation — not only the request it raises.

Coverage runs the whole depth of the workflow, and every layer is cut by slice.

Surface — the request the planner opens

E1Final-output evaluationDid the request name the fault the technician actually found?
E2Step-level evaluationDid it read the right asset, configuration and manual revision?
E3Tool evaluationDoes the request tie back to the exact asset, work order and part?
E4Signal attributionWas the request raised on an asset fault rather than an instrument fault?
E5Slice evaluationHow does accuracy change across asset classes, ages and criticality?
E6Business outcomeHow many were duplicates, and how often did the parts list hold?

Floor — what the technician actually finds

Failure modes

Where each failure originates in the agent

Seven failure modes plotted against the five stages of the agent lifecycle. None of them authorises work — the planner's read, the supervisor's authorisation and the plant's isolation and permit procedures are the controls that stop them. If one gets through and a technician is already at the machine, the asset, the revision and the plan held in the request are what the stop-work and the record correction are built from.

Agent lifecycle

Direction of processing →

01 · Retrieval
2 modes
- MA-01
  Sensor fault read as asset fault
  A drifting transmitter is written up as a machine fault.
- MA-02
  History from a sister asset
  Repairs from a same-model asset with a different build.
Stage gathers
The signal, the asset record, history and manual
02 · Reasoning
2 modes
- MA-03
  Plan written to an asset that changed
  A superseded revision, or a part a rebuild removed.
- MA-04
  Window production had committed
  Preventive work proposed into a slot already sold.
Stage prepares
The fault named, the job plan and the parts list
03 · Request / write
1 mode
- MA-05
  Same fault, two work orders
  The alert and the operator report are raised separately.
Stage raises
The work request written to the maintenance system
04 · Output
1 mode
- MA-06
  Alert volume outruns the technicians
  So much is raised that work orders close unread.
Stage presents
The request, the evidence and the window proposed
05 · Change / Version
1 mode
- MA-07
  Redundancy the criticality misses
  A standby asset is ranked as if it stood alone.
Stage tracks
Model, threshold, PM-interval and asset changes

Sev-1 · a technician works to a wrong plan
Sev-2 · the asset record cannot be relied on
Sev-3 · the backlog fills, requests stop being read

See the complete Manufacturing failure-mode catalogue

Affected slices

A rebuilt asset is not the asset the record describes

Two pumps on a line are not the same problem. An asset with a full history and an unchanged build reads cleanly; one rebuilt, with a manual that never caught up, does not. Nestack reports performance by slice, not only in total.

Slice performance — reported separately, not only in aggregate

Illustrative example

Slice	Failure rate	Lift	Status
Rebuilt or modified assets	7.0%	3.5×	Review
Poorly documented old assets	5.2%	2.6×	Review
Operator-reported faults	3.8%	1.9×	Watch
Settled assets, full history	1.4%	0.7×	Normal

Bar: lift in requests that missed the actual fault · scale 0–4.0× · tick marks 2.0×

2 of 4 slices over threshold

Evidence-linked improvement

What the technician found goes back into the request

The technician's finding is the answer key. What was actually wrong, which parts were used and what the plan missed come back as cases to get right.

Improvement cycle · five stages

Switchback — the path turns at Improve and returns at Learn

01
Detect
Requests that missed the fault, duplicates or parts-list errors move on an asset class.
02
Diagnose
Traced to the signal read, the asset identified, the revision cited or the build assumed.
03
Improve
The retrieval rule, plan template or criticality input changes under your change control, with a named approver.
04
Verify
Re-run against held-out requests from that class, including the ones the technician overturned.
05
Learn
The overturned request is kept as a case and the reason enters the work-management review.
Learn → Detect
The return edge. A change to a PM interval or a criticality rating is an asset-strategy change — your reliability engineer approves it before it ships, not after.

Typical build scope

Twelve workstreams across six weeks

The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, asset and document access, plan drafting, evaluation, integration, then supervised running and handover.

Workstream

Week 1Week 2Week 3Week 4Week 5Week 6

01Work-management discovery and boundaries.Week 1
02Asset hierarchy and criticality mapping.Week 1, Week 2
03CMMS, historian and condition-system access.Week 1, Week 2
04Asset, work-order and part identity keys.Week 2
05Signal and operator-report ingestion.Week 2, Week 3
06Manual, drawing and bill-of-material retrieval.Week 2, Week 3
07Job-plan and parts-list drafting.Week 3, Week 4
08Instrument-fault and duplicate checks.Week 4
09Production-window and schedule-conflict checks.Week 4, Week 5
10Evaluation suite and held-out asset classes.Week 4, Week 5
11Work-request write-back and planner routing.Week 5, Week 6
12Observability, deployment and Agent Care handover.Week 6

12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel

Final scope and sequence confirmed in discovery

Engagement tiers

What each tier includes

Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.

Capability✓ in scope · — not at this tier	PilotOne asset class, one line	ProductionProduction CMMS integration	AdvancedMulti-plant / multi-system
Introduced at Pilot
Work requests drafted with the evidence	✓	✓	✓
Asset history and past repairs attached	✓	✓	✓
Manual-revision currency checks	✓	✓	✓
Authorisation by a named supervisor	✓	✓	✓
Baseline evaluation	✓	✓	✓
Introduced at Production
Job-plan and parts-list drafting	—	✓	✓
Instrument-fault and duplicate checks	—	✓	✓
Work-request write-back and planner routing	—	✓	✓
Observability and evaluation	—	✓	✓
Introduced at Advanced
Condition-trend and drift flagging	—	—	✓
Multi-plant and enterprise controls	—	—	✓
Build price	From $5,000	From $8,000	Custom quote

Final build price: Confirmed after discovery based on the asset classes in scope, CMMS, historian and condition-system integrations, document and bill-of-material quality, approval controls and deployment requirements.
Separate from build: Build pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.

What we need from you

What you bring, and what we build with it

Each input maps to a piece of build scope and a week in the delivery timeline.

You bring

We build with it

01Your asset hierarchy and the criticality ratings behind it
→
Asset hierarchy and criticality mapping
Week 1
02Access to the CMMS, the historian and the condition system
→
System access assessment, then asset, work-order and part identity keys
Week 2
03OEM manuals, drawing revisions and the bills of material as fitted
→
Manual, drawing and bill-of-material retrieval
Week 3
04Your standard job plans and the isolation procedures they cite
→
Job-plan and parts-list drafting
Week 3
05The production windows you protect and how they get committed
→
Production-window and schedule-conflict checks
Week 4
06Work orders from the last year, with what the technician actually found
→
Evaluation suite, held-out asset classes and failure-mode testing
Week 4
07A named planner and a maintenance supervisor
→
Planner routing and write-back, then supervised running
Weeks 5–6

Nothing else is required

Deployment, documentation and Agent Care handover are ours.

Delivery timeline

Four phases across six weeks

Phases are drawn over the weeks they actually occupy. Week 5 carries both the held-out asset classes and the first requests your planner works from.

Phase

DiscoveryW1
BuildW2 – W3
EvaluateW4 – W5
Pilot & LaunchW5 – W6

Week focus

W1Work-management discovery, asset hierarchy and criticality ratings

W2CMMS, historian and condition-system access, then identity keys

W3Manual and bill-of-material retrieval, job-plan and parts-list drafting

W4Evaluation suite, instrument-fault checks and production-window conflicts

W5Held-out asset classes, first supervised requests and targeted corrections

W6Your planner works from the request, then Agent Care starts

Reading the band: Instrument-fault and duplicate checks are built in week 4 — before any request reaches a planner's backlog in week 5.
At the end of W6: Requests have run alongside your existing route and been planned by the planner who schedules from them, then Agent Care takes over monitoring.
Duration: Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.

Next step · Manufacturing AI agent

Build a maintenance assistant around your work-management process.

Show us one asset class, its work-order history and the job plans and manuals your planners work from today. We'll prepare requests from your own condition signals and operator reports, then report what the technician actually found against what the agent proposed.

Typical delivery: 4–6 weeks
Deployment: Your maintenance stack
Work authority: A supervisor, always
After launch: Agent Care available

Nestack Agents · Maintenance assistant

AGT-MFG-03 · Agent Care available after launch

Maintenance-Assistant AI Agent (Work Preparation)

Work from the asset's own record

Hand the planner a job already prepared

Send the doubtful ones to a supervisor

CMMS & EAM

Condition monitoring

Historian & control

Maintenance assistant

Documents & parts

Observability & evaluation

Sensor fault read as asset fault

History from a sister asset

Plan written to an asset that changed

Window production had committed

Same fault, two work orders

Alert volume outruns the technicians

Redundancy the criticality misses

Detect

Diagnose

Improve

Verify

Learn

Build a maintenance assistant around your work-management process.