Manufacturing AI agent · Troubleshooting
Shop-Floor Troubleshooting AI Agent
Read the fault code and machine state, retrieve the approved procedure for that machine, and walk the standard diagnostic sequence — the person at the machine acts under the plant's isolation and permit rules.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Safety authority
- Never the agent
- After launch
- Agent Care
What this agent does
Retrieves the procedure, does not perform it
In
- 01
Take the fault code, the machine state and what the operator describes from the line, the HMI and the historian.
- 02
Identify the exact machine — variant, options fitted, controls version and the document revisions in force.
Reason
- 03
Retrieve the manual passage, the electrical drawing sheet and the diagnostic sequence approved for that machine.
- 04
Read the fault code against the controls version running on it, not against a code list from another build.
- 05
Pull past occurrences on this machine, and mark any fix carried over from a sister machine as a different build.
Decide
- 06
Mark every step that requires isolation, a permit or a competency the person at the machine does not hold.
- 07
Stop where the approved sequence runs out and escalate, rather than proposing a step nobody has approved.
Out
- 08
Show one step at a time in the language of the floor, with what to look at and what result ends the step.
- 09
Retain the fault, the state captured, the revisions cited, the steps walked and what actually fixed it.
Product statement
The agent retrieves and guides. The person at the machine acts under the plant's isolation, permit and competency rules, and anything past their authorisation goes to maintenance or engineering.
Example workflow
One fault code, end to end
- 1Machine stopsA fault code, an alarm, a jam nobody can clear or an operator's call from the line
- 2Machine identifiedSerial, variant, options fitted, controls version, and the manual and drawing revisions in force
- 3Guidance retrievedThe fault-code meaning for that controls version, the approved sequence, the drawing sheet and past occurrences here
- 4Safety envelope appliedEvery step tagged for isolation, permit and competency, and for whether an operator may do it at all
No human action required
Stages 1 to 4 run without a person in the loop — the machine is identified, the guidance is retrieved and every step is tagged before the operator is shown anything. A step needing isolation ends that stretch on the spot.
The approved step is shown, one at a time.
Stops there and hands over with the state.
Maintenance technician or controls engineer
Takes the fault, the machine state and the steps already walked, then works under the plant's own isolation and permit procedure.
Accept · Correct · Send back
Resolved — handed back
- 6Escalation raised with the stateFault, machine state, steps already walked and what has been ruled out — written to the maintenance system, not acted on
- 7Outcome evaluatedWhat actually fixed it, where the sequence ran out, and every unsafe step a technician struck out in review
Corrections
What the technician strikes out of the sequence is counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Declaring a machine safe to work on.
- Isolating, locking or tagging an energy source.
- Issuing, transferring or closing a permit to work.
- Authorising work beyond the operator's own competency.
Automation boundary
Agent acts unaided
- ✓Identify the machine, its variant, options fitted and controls version.
- ✓Retrieve the approved procedure, drawing sheet and fault-code meaning in force.
- ✓Show past occurrences on this machine and what actually fixed them.
- ✓Tag each step for isolation and competency, and escalate with the state captured.
The agent writes the fault log and the escalation, never the machine. The isolation and the restart are not its to write.
- Removing, defeating or bypassing a guard or interlock.
- Resetting a fault or restarting a stopped machine.
- Changing a parameter, setpoint or recipe on the machine.
- Revising an approved procedure or fault-code list.
Example output
One fault code, annotated
Everything the agent shows is tied to the machine it was read from and the revision in force at the time.
Guidance output · one fault code
Illustrative example- Fault raised by
- Machine HMI and operator
- Machine
- Filler L3-02, variant B
- Controls version
- v7.2, in force
- Guidance shown
- Approved sequence — steps 1 to 3
- Confidence
- 91%
- Safety authority
- Never the agent
As received
The code as the controller raised it, the machine it names and the version running on it — read, not inferred.
Evidence used
- Rev 6 manual, fault 412
- Drawing sheet 12
- Two prior occurrences
Why it stopped at step 3
Step 4 reaches inside a guarded enclosure. Isolation is not an operator task.
What the score decides
Confidence decides how much is shown before it hands over, never whether a step is safe to do.
Action
AcceptCorrectSend backValue
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 stopped machine
A fault code, an alarm or a call from the line
Work from this machine's record
Use the variant as built, the controls version running on it, the manual and drawing revisions in force, and what fixed this fault here before.
Give the operator the right page
The manual passage, the drawing sheet and the past occurrences arrive with the fault code, instead of a hunt through a folder while the line is down.
Hand over when the sequence runs out
Anything needing isolation, a permit or a trade stops at the operator and reaches maintenance with the machine state already captured, rather than being improvised.
Build an evidence trail
Retain the fault, the state captured, the revisions cited, the steps walked, the escalation and what actually fixed it — on both paths.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Control & HMI
- FactoryTalk
- Siemens WinCC
- Machine HMIs
- alarm servers
Documents & drawings
- OEM manuals
- electrical drawings
- Parts catalogues
- standard work
CMMS & fault history
- SAP PM
- IBM Maximo
- Infor EAM
- Closed work orders
- past occurrences
Agent
Shop-floor troubleshooting
- Identifies the machine
- Retrieves the procedure
- Escalates the rest
MES & historian
- Siemens Opcenter
- Aveva historian
- PI System
- downtime and stop records
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 person at the machine
Each control wraps the one inside it. A step clears every layer before an operator sees it, and the plant's isolation and permit procedures sit outside all six.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Keep the guidance about this machine
- L3 – L4
- Decide what may be shown at all
- L5
- Hands over in time, with the state
- L6
- Pulls automation back when signals degrade
How Nestack evaluates it
Evaluate the whole guided sequence — not only the fault it names.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the step the operator is shown
- E1Final-output evaluationDid the guidance name the fault the technician actually found?
- E2Retrieval evaluationWas it the right variant, the right revision and the right fault list?
- E3Step-level evaluationDid a technician judge each step accurate and in the right order?
- E4Escalation judgementDid it stop and hand over at the point the procedure ran out?
- E5Slice evaluationHow does accuracy change across machine age, shift and language?
- E6Business outcomeHow often was it resolved without escalation where that was correct?
Floor — whether the machine ran again, safely
Failure modes
Where each failure originates in the agent
Seven failure modes plotted against the five stages of the agent lifecycle. None of them puts a person's hands into a machine on its own — the plant's isolation, permit and competency rules, and a technician's review of every step, are the controls that stop them. If one gets through and an operator has already acted, the machine is stopped, the guidance is withdrawn from that fault code, and the record shows exactly what was shown, to whom and when.
Agent lifecycle
Direction of processing →
01 · Retrieval
2 modes
- TS-01
Variant or revision wrong
The procedure retrieved does not describe this machine.
- TS-02
Fix from a sister machine
What worked on the machine next to it, built differently.
Stage gathers
The machine, its revision and the fault-code list
02 · Reasoning
2 modes
- TS-03
Step assumes isolation
Written for a technician on a locked-out machine, shown live.
- TS-04
A next step where none exists
Plausible guidance where the honest answer is to call someone.
Stage assembles
The diagnostic sequence and the evidence per step
03 · Safety screen
1 mode
- TS-05
Drifts into maintenance work
The sequence carries on past what an operator may touch.
Stage blocks
Steps needing isolation, a permit or a competency
04 · Output
1 mode
- TS-06
Written for the wrong reader
Language or reading level the person at the machine cannot use.
Stage shows
One step at a time, in the language of the floor
05 · Change / Version
1 mode
- TS-07
Fault code changed meaning
A controls update reassigned the code; the list did not follow.
Stage tracks
Model, procedure, fault-list and controls changes
- Sev-1 · someone could be hurt at the machine
- Sev-2 · the wrong fault is chased
- Sev-3 · the operator gives up and calls anyway
Affected slices
The old machine at three in the morning
A current standard-build machine with a maintained manual reads cleanly. One rebuilt twice, on a night shift, with a manual nobody updated, does not — and that is the machine an operator needs help with. Reported by slice, not in total.
Slice performance — reported separately, not only in aggregate
Illustrative example| Slice | Failure rate | Lift | Lift vs. threshold | Status |
|---|---|---|---|---|
| Legacy machines, poor documents | 7.4% | 3.7× | Review | |
| Night and weekend shifts | 5.6% | 2.8× | Review | |
| Faults after a controls update | 3.6% | 1.8× | Watch | |
| Current standard-build machines | 1.6% | 0.8× | Normal |
Bar: incorrect-guidance rate vs. standard-build baseline · scale 0–4.0× · tick marks 2.0×
2 of 4 slices over threshold
Evidence-linked improvement
Every escalation says where the procedure ran out
What the technician found after the hand-over, and every step a reviewer struck out, come back as the cases the next release has to get right.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Struck-out steps, wrong-fault guidance or hand-overs that should have come sooner.
- 02
Diagnose
Traced to the machine identified, the revision retrieved, the fault list or the step tagging.
- 03
Improve
The retrieval key, the safety tagging or the procedure text changes under your document control, with a named approver.
- 04
Verify
Re-run over stored faults from that machine class, including every step a technician struck out.
- 05
Learn
The struck-out step is kept as a case, and the gap goes to whoever owns that procedure.
Learn → Detect
The return edge. An unsafe suggestion is an incident in its own right — the guidance is withdrawn from that fault code, and the operators who were shown it are told.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, machine and document access, guidance and safety tagging, evaluation, then supervised running and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Troubleshooting discovery and boundaries.Week 1
- 02Machine register, variants and revisions.Week 1, Week 2
- 03Control, HMI and historian access.Week 1, Week 2
- 04Fault-code list per controls version.Week 2
- 05Manual, drawing and procedure retrieval.Week 2, Week 3
- 06Past-occurrence and work-order history.Week 2, Week 3
- 07Safety tagging of every procedure step.Week 3, Week 4
- 08Escalation routing and state capture.Week 3, Week 4
- 09Operator language and reading-level pass.Week 4
- 10Evaluation suite and held-out machine classes.Week 4, Week 5
- 11Technician review and unsafe-step logging.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 machine class, one line | ProductionProduction document integration | AdvancedMulti-plant / multi-language |
|---|---|---|---|
| Introduced at Pilot | |||
| Fault-code guidance, variant and revision matched | ✓ | ✓ | ✓ |
| Every step tagged for isolation and competency | ✓ | ✓ | ✓ |
| Unsafe-step blocking and logging | ✓ | ✓ | ✓ |
| Escalation with the machine state captured | ✓ | ✓ | ✓ |
| Technician review of guidance steps | ✓ | ✓ | ✓ |
| Baseline evaluation | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Past occurrences and what fixed them | — | ✓ | ✓ |
| Every machine class on the line | — | ✓ | ✓ |
| Observability and evaluation | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Second-language and reading-level delivery | — | — | ✓ |
| Multi-plant and enterprise controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on the machine classes in scope, control, document and CMMS integrations, manual and drawing quality, the languages on the floor 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 machine register — variants, options fitted and controls versions
→Machine register, variants and revisions
Week 102Your isolation, permit and competency rules, and who they authorise
→Troubleshooting discovery and the safety boundary
Week 103OEM manuals, electrical drawings and the fault-code lists you hold
→Manual, drawing and procedure retrieval
Week 204The standard diagnostic sequences your technicians actually work to
→Procedure retrieval and step-by-step guidance
Week 305Who an operator calls, when, and how that differs by shift
→Escalation routing and state capture
Week 406A year of fault records with what actually fixed each one
→Evaluation suite, held-out machine classes and failure-mode testing
Week 407A named technician reviewer and operators on each shift
→Technician review workflow, 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 machine classes and the first faults your operators work from.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Troubleshooting discovery, machine register and the safety boundary
W2Control and document access, then the fault list per controls version
W3Procedure retrieval, step guidance and safety tagging of each step
W4Evaluation suite, escalation routing and the reading-level pass
W5Held-out machine classes, first supervised faults and corrections
W6Operators work live faults on one line, then Agent Care starts
- Reading the band
- Safety tagging is built in week 3, before a single step reaches an operator in week 5. An untagged step is the one that hurts someone.
- At the end of W6
- Operators have worked real faults alongside your existing call-out route, every step shown has been reviewed by a technician, 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 troubleshooting agent around your own machines.
Show us one machine class, the manuals and drawings you hold for it, and a year of faults with what actually fixed them. We'll guide one fault code from your own documents, mark every step that needs isolation or a trade, and show you where the approved procedure runs out.
- Typical delivery
- 4–6 weeks
- Deployment
- Your control and document stack
- Safety authority
- The plant's rules, always
- After launch
- Agent Care available
Nestack Agents · Shop-floor troubleshooting
AGT-MFG-06 · Agent Care available after launch