Agent Found the Problem, Why Isn't It Fixed? Closing the Manufacturing Quality Loop

In a hypothetical case, early-shift inspection finds increased dimensional deviations in a batch of machined parts. An AI Agent quickly reviews inspection records, production orders, and equipment maintenance logs, lists candidate causes, and creates a pending inspection work order. If the demo stops there, it looks successful. But at shift handover, quality staff still ask whether the batch can proceed, maintenance says they just saw the work order, and the production supervisor is unclear who will arrange the inspection. The Agent's task shows "success" in the system, yet the physical problem remains untouched.

Four Types of "Completion" Cannot Share One Checkmark

The author distinguishes four record types, each answering a different question and having its own completion signal:

Anomaly Event – Which inspection or monitoring detected what? Completion: anomaly logged; acknowledgment or signal recovery recorded separately.

Diagnosis Task – What data was reviewed and what judgment reached? Completion: analysis delivered, or explicitly handed to human due to insufficient evidence.

Disposal Work Order – Who was assigned to do what, and what was actually done? Completion: prescribed inspection or repair work finished.

Quality Problem Record – How was the affected scope handled, and are measures effective? Completion: verification completed per agreed criteria, closed by authorized person.

A single quality problem record can link multiple diagnoses and work orders. They may reside in one software system, but their statuses must not be conflated: diagnosis completion cannot auto-close the problem, a maintainer's "done" entry cannot replace quality release, and signal recovery only means the current signal is normal – the earlier batch disposition still needs accounting.

Contain Impact First, Then Investigate Cause

When facing dimensional deviation, two things must happen simultaneously: define the suspect product scope and investigate the deviation source. Necessary temporary containment cannot wait for root cause confirmation. Quality leads can invoke existing procedures to identify, isolate, or add inspections for suspect batches to prevent further impact. The Agent can prepare materials but cannot unilaterally scrap or release the whole batch.

The investigation step lists candidate directions – tooling, fixturing, material, machining conditions, measurement – as starting points. The article references ASQ's 8D methodology, which separates temporary containment, root cause verification, corrective action, and effectiveness verification. The author borrows this distinction without requiring a full 8D report for every minor anomaly.

In the case, temporarily isolating suspect products does not fix the equipment; completing a repair does not prove all deviations came from that component. Post-repair sample conformance must still meet pre-agreed verification scope and observation requirements. Production restart, batch release, and quality problem closure can have different conditions and occur at different times. If long-term effectiveness observation is incomplete when production restarts, the verification task and owner must be retained – not hidden in a vague "follow-up" note.

Who Acts Next Matters More Than Generating Another Recommendation

A system showing "message pushed" does not guarantee the shift lead saw it; even if seen, they may wait for spare parts, an inspection window, or discover the task belongs to another role. The process must capture these differences.

Separating system records from physical actions clarifies handover points. The article illustrates the flow and notes three branches that must not be omitted:

Immediate Containment: On-site personnel act per existing procedures, log measures and owners synchronously, without waiting for Agent dispatch or diagnosis completion.

No One Picks Up or Conditions Unmet: Retain explicit waiting state and escalate per agreed rules.

Recurrence After Closure: Reopen per procedure or link to a follow-up problem, preserving original records.

Containment and investigation can run in parallel with separate progress tracking. A maintainer accepting an inspection work order only takes one work item; the overall quality issue still needs a lead person tracking: where product disposal stands, which role is waiting on whom, when verification results will return. For example, if a maintainer has accepted but awaits a production window, the state should specify what is waited for, who arranges it, and when to re-check. Beyond the escalation deadline defined by anomaly severity and site policy, it escalates to a designated owner.

Cross-shift handover demands explicit transfer. The Agent can generate a Pending Items Shift-Handover Summary from work orders, inspection records, and disposal logs: which batches are affected, what has been done, what is still waited on, who follows up next and when. Each item retains source and update time; unconfirmed or conflicting judgments must not be written as conclusions. The outgoing shift verifies the summary, the incoming shift acknowledges receipt – only then is handover complete. For the next shift lead, this actionable checklist is often more useful than a regenerated diagnosis report.

If inspection was completed offline, back-entry must preserve actual completion time, entry time, and evidence. Agent not seeing a write-back does not mean the site didn't act; receiving a "done" message does not mean verification passed.

Put Closure Conditions on a Single Card

What constitutes "done" for this anomaly can be written directly into the existing anomaly record. The following fields are suggested for site adaptation – not a mandatory form, not an extra sheet for crews:

Problem Object & Scope: Part, inspection characteristic, operation, time window; involved batches and pending investigation scope.

Linked Tasks: Anomaly ID, diagnosis tasks, inspection/repair work orders, batch disposal records.

Lead & Current Owner: End-to-end tracker, current assignee, waiting items, next-shift receiver.

Preconditions: Records complete; disposal authorized; site has inspection conditions.

Completion Evidence: Measure records, product disposal basis, inspection and necessary observation records.

Verifier & Closer: Who defines verification requirements; who is qualified to review; who has authority to close.

Timeout & Escalation: Escalation deadlines and recipients for unassigned tasks, prolonged missing conditions, or shift-handover leftovers.

Reopen Conditions: Verification failure, same-type recurrence, impact expansion, or original basis overturned.

The mapping between who arranges disposal, who sets closure requirements, and who maintains records must land separately on site, quality, and system owners. The Agent helps organize materials and remind progress; if the owner column only says "Agent", no one is accountable when issues arise.

If root cause remains unclear, keep "under investigation." A temporary measure can close, but root cause investigation and quality problem closure are separate judgments. Where current procedures or customer requirements mandate root cause identification and measure verification, continue; where closure is permitted otherwise, still document basis, approver, and residual responsibility per applicable procedure. "Accepting residual risk" is not a reason to skip these steps, nor can a model-generated conclusion fill the form.

What Work Does the Agent Actually Save?

Along this chain, the Agent can: consolidate scattered records and flag missing batch or operation information; present candidate causes with supporting evidence side by side; prepare work order drafts; compile progress and pending items from disposal records; deliver before-and-after inspection results to verifiers.

Doing these well eliminates time spent repeatedly searching documents, cross-checking records, and communicating. Site command and quality release remain with authorized roles, but the person taking over no longer starts from scratch each time.

Siemens' November 2024 press release described the Industrial Copilot at the Erlangen factory using machine data, historical records, manuals, and spare parts info to help personnel interpret error codes and find handling procedures. That reference covers analysis assistance only; who accepts the work order, how the batch is disposed, and who verifies and closes the problem still depend on the specific process – the press release does not answer those questions.

If the site already has a mature work order process, integrating the Agent into it suffices. Fixed-field checks, due-date reminders, and rule-based dispatch stay with conventional software. Agent assistance is evaluated for steps where data is scattered, expression is inconsistent, and investigation paths need context-aware adjustment.

Not all waiting in the disposal process should be blamed on AI. Missing spare parts, unavailable personnel, disputed quality responsibilities – these are exposed by the system but won't disappear by swapping models.

During Acceptance, Deliberately Pick the Hard-to-Close Cases

Demonstrating only "detect anomaly, generate work order, receive success receipt" easily bypasses real problems. Before trial, quality and site teams should jointly prepare anomaly test cases, then have the system owner verify flow and traceability.

Repeated alarms on same batch: can they link to existing problem to avoid proliferating work orders while not losing new impact?

No one picks up task: can it show wait time and escalate to designated role instead of auto-closing?

Critical inspection record missing: can it assign a supplementary evidence task while retaining "cause unconfirmed"?

Repair done but re-test fails: can it continue disposal without being forced closed by work order status?

Outgoing shift leaves with unresolved issue: can incoming shift receive complete materials and acknowledge takeover?

Same deviation recurs after closure: can it retrieve old measures and evidence to decide reopen vs. linked problem?

Acceptance must not only measure analysis speed. How long unassigned tasks waited, where disposal stalled, how many days re-test took, how many issues remain pending or reopened – all must be counted. Otherwise, the hardest-to-close problems are the ones most likely to vanish from statistics.

Summary

In this case, what was missing was not another anomaly analysis, but someone to continue arranging inspection, disposing the batch, verifying results, and carrying it across a shift without loss.

First clarify "what makes this thing done," then decide which steps the Agent owns. Otherwise, the system gets better at finding problems while the site merely gains another nagging assistant.

The next article "MES and ERP Are Connected, Why Can't the Agent Get Things Done?" will pursue: if the assignee receives materials where equipment, batch, and time don't align, what should subsequent judgment rely on?

References

[1] ASQ: Eight Disciplines 8D. Cited for its distinction of temporary containment, root cause verification, and measure verification; the responsibility card and flow design in this article are engineering inductions for the hypothetical case. https://asq.org/quality-resources/eight-disciplines-8d

[2] Siemens: Siemens Industrial Copilot expanded, adopted by thyssenkrupp, 2024-11-12. Only the operational assistance scenario from the press release is cited; planned capabilities or vendor performance claims are not treated as independent validation. https://press.siemens.com/global/en/pressrelease/siemens-industrial-copilot-expanded-adopted-thyssenkrupp