Main site

Quality assurance AI agent · Defect triage

Defect Triage AI Agent

Date the hour someone first knew, link the duplicates without burying the one that names an injury, and leave the reportability call to the qualified person who signs it.

Typical delivery
4–6 weeks
Deployment
Your stack
Named reviewer
Clock first
After launch
Agent Care

What this agent does

Prepares the judgement, never signs it

In

  1. 01

    A report arrives, and the hour an employee first held it is stamped beside it.

  2. 02

    A clock is already running, because 21 CFR 803.3 dates awareness to any employee.

Reason

  1. 03

    A duplicate is linked and not merged, so the one line naming harm survives the cluster.

  2. 04

    A cluster is a trend analysis, and 21 CFR 803.53 runs five work days from one.

  3. 05

    A severity band is set, and it is kept apart from the reportability question.

Decide

  1. 06

    A consumer hazard surfaces, and 16 CFR 1115.14 allows ten days to investigate.

  2. 07

    A not-reportable view is drafted, and 21 CFR 803.20(c)(2) leaves that call to a qualified person.

Out

  1. 08

    One defect meets two regimes, and CRA Article 14 adds a third from 11 September 2026.

  2. 09

    Execute write actions only inside the approval boundaries agreed during implementation.

Product statement

Intake, dating, linking and routing belong to the agent. The reportability decision, and the name that stands under it, belong to a qualified person the rule itself describes.

Example workflow

One report, arrival to decision

  1. 1Report receivedA complaint, a field report, a support thread, a returned unit or a failed test
  2. 2Awareness dated and recordedWho first held the information, the hour it reached them, the product and what it describes
  3. 3Duplicates linked, not mergedThe cluster, each member left whole, the outlier lines and confidence
  4. 4Controls appliedHarm-language checks, clock checks, regime checks and triage confidence

No human action required

Stages 1 to 4 run unaided, and no report is closed at any of them — the agent is preparing, and the reviewer lane opens at the reportability gate.

5DecisionSplits at the reportability gate
No harm language present

Goes to the qualified reviewer to decide.

Anything naming injury

Adds a regulatory-affairs read first.

Reviewer decision

The report is held with its arrival hour, the cluster it joined and the clocks it sits under.

Decide · Attach evidence · Send to regulatory review

Decided — by the qualified reviewer

  1. 6Complaint and event-file records updatedOnly where write access and event-file policy allow it
  2. 7Outcome evaluatedDecision accuracy, reopened duplicates, reviewer corrections and what the later audit turned up

Corrections

Each correction the reviewer makes counts in the evaluation.

What should not run autonomously

Human approval stays in control

Outside the boundary — human approval required

8 items

  • Concluding that an event is not reportable.
    • Signing the written report that goes to the Commission.
    • Closing a report whose text names an injury.
    • Deciding a quarterly summary form is enough.

  • Automation boundary

    Agent acts unaided

    • ✓Date the hour the first employee became aware.
    • ✓Link duplicate reports and preserve each reportability determination.
    • ✓Record the deliberation a reviewer relied on, dated and sourced.
    • ✓Hold the report for the reviewer who decides it.

    No report closes as not reportable except by a qualified person, inside agreed boundaries.

    • Judging whether a device may have contributed.
    • Telling a regulator the review is complete.
    • Setting which day a statutory clock is counted from.
  • Changes to the triage rules or the reportability gate.

Example output

One defect report, annotated

An engineering agent keeps the cause and a warranty agent keeps the unit; this one keeps whether a report must be told to a regulator. Below is one report exactly as the agent leaves it.

Triage record · single report

Illustrative example
Report
Customer report, one product line
Recorded as
Unexpected speed change, injury named
Awareness
Linked, not merged
Evidence of record
Support thread, 3 August 2026
Confidence
Held undecided
Held for
The qualified reviewer, by name

As received

Drawn from the support thread and the complaint file, and it concludes nothing those two do not carry.

What the record holds

  • Support thread lines
  • Complaint file entry
  • Linked duplicate set

Why no decision here

Calling a report not reportable is an act a qualified reviewer owns.

What the score decides

Below the configured threshold a report gets a regulatory read before the reviewer sees it.

Action

DecideAttach evidenceSend to regulatory review

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

Each report

From the hour it was first held

03Report

Where the report is used

The best published duplicate detectors put the true match in a top-ten shortlist about three times in five, so a linked cluster is a shortlist for a reader and not a closure.

01Approved path

The clock started already

A treadmill importer settled with the Commission this August over a hazard known from March 2018 and reported in October 2022; those reports were triaged, not missed.

02Human review

What was checked, and not found

No live standard defines defect-report severity: IEEE 1044-2009 has been Inactive-Reserved since 5 March 2020, and the classes in IEC 62304 and ISO 26262 attach to a software item or a hazard, never to a report. Corrective action moved to ISO 13485 clause 8.5 via § 820.10(c) on 2 February 2026.

04

Build an evidence trail

The report, the hour it arrived and the person who judged it stay together.

Integrations

Typical integrations

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

  • Defect and complaint tracking

    • Jira
    • Bugzilla
    • complaint logs
    • Field and test defect records

  • Support and field intake

    • Helpdesk
    • CRM
    • field reports
    • First-contact report lines

  • Regulatory event files

    • MDR event files
    • CPSC submissions
    • Reportability decisions and dates

Agent

Defect triage and reportability

  • Reads the reports
  • Dates and links them
  • Holds the decision

  • Trend and analysis records

    • Complaint trending
    • dashboards
    • Cluster reviews and escalation logs

  • Observability & evaluation

    • OpenTelemetry
    • Langfuse
    • Supported monitoring/evaluation sources

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

Agent controls

Six screens between the model and the file

Six screens down one chute, the last the tightest. What lands is drawn in the map below.

L6 · Outermost — last line of defence

Inward → L1 · closest to the model

L6Rollback / safe modeNarrow the agent to intake and dating when evaluation or production signals degrade.Roll back
L5Version monitoringTrack model, prompt and triage rules, and note the version each report was prepared under.Track
L4TraceabilityRecord each report, the hour it arrived, the cluster it joined and each read of the file.Record
L3Reviewer releaseHold the report for a qualified reviewer; the hold governs release, not whether the call is right.Gate
L2Triage guardrailsTest each draft against the harm language in the source, and return one that closes on silence.Restrict
L1Confidence thresholdsRoute a thin or injury-adjacent report to a regulatory read before the reviewer sees it.Require review
Model coreReport prepared — the arrival hour, the linked cluster, the clocks and confidence
L1 – L2
Test whether a report may close
L3
Leaves the decision to a qualified person
L4 – L5
Keep the report and the judgement behind it
L6
Routes to a qualified reader when signals degrade

How Nestack evaluates it

Evaluate the whole triage run — not only the decision that comes out.

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

Surface — the decision an event file carries

  1. E1Final-output evaluationDid the record hold the deliberation the decision rested on?
  2. E2Step-level evaluationDid the agent read the right reports, the right hours and the live triage rules?
  3. E3Tool evaluationDid it read and write the correct report and the correct event file?
  4. E4Confidence calibrationDo low-confidence triage drafts actually attract more reviewer corrections?
  5. E5Slice evaluationHow does performance change from one defect class to another?
  6. E6Business outcomeHow many reports needed a correction before the reviewer decided?

Floor — the file a regulator opens first

Failure modes

Where each failure originates in the agent

Seven failure modes, each placed where it first becomes visible.

Agent lifecycle

Direction of processing →

  1. 01 · Retrieval

    1 mode

    • VE-03

      The report that never reached the tracker

      A clock ran inside a thread nobody triaged.

    Stage gathers

    The reports, the hours, the units and the notes

  2. 02 · Reasoning

    2 modes

    • VE-04

      Merged away with the injury line

      The one report naming harm joins a cluster.

    • VE-06

      Two regimes, one due date

      The tracker counts down the wrong clock.

    Stage proposes

    The arrival hour, the cluster and the clocks

  3. 03 · Tool / write

    2 modes

    • VE-02

      Severity gates the reportability read

      A low-banded report skips the review.

    • VE-05

      Cluster seen, nobody told

      A trend surfaces on a screen, not to a name.

    Stage writes

    Only where write access and approval policy allow it

  4. 04 · Output

    1 mode

    • VE-01

      Closed, the deliberation unrecorded

      The file shows a call, not the reasoning under it.

    Stage returns

    The call a file carries and a regulator reads

  5. 05 · Change / Version

    1 mode

    • VE-07

      Quiet accuracy drift

      The wording moves on while the linking model does not.

    Stage tracks

    Model, prompt, triage rules and clock dates

  • Sev-1 · a reportable report closed
  • Sev-2 · a due date counted from the wrong hour
  • Sev-3 · signals thin, report held back
See the complete Quality failure-mode catalogue

Affected slices

Injury-adjacent reports absorb the corrections

A class-level triage figure can read clean while injury-adjacent complaints carry most of the reviewer corrections. Nestack reports the correction rate by defect class, not only in total.

Slice performance — reported separately, not only in aggregate

Illustrative example
SliceFailure rateLiftLift vs. thresholdStatus
Injury-adjacent complaints13.6%3.7×Review
Cross-regime field reports9.6%2.6×Review
Duplicate-cluster outliers6.0%1.6×Watch
Routine cosmetic defects2.7%0.7×Normal

Bar: correction-rate lift vs. routine-cosmetic baseline · scale 0–4.0× · tick marks the 2.0× review threshold

2 of 4 slices over threshold

Evidence-linked improvement

What a low-filed report costs

A cycle ends when the report held below the bar nobody set is a standing case. That suite is what the next intake window is measured against.

Improvement cycle · five stages

Switchback — the path turns at Improve and returns at Learn

  1. 01

    Detect

    Correction rate rises on injury-adjacent reports.

  2. 02

    Diagnose

    The corrective action opened on a low-filed report, which stopped no clock and discharged no reporting duty, is worked backwards until one cause is left standing.

  3. 03

    Improve

    The queue goes out numbered, and the reports beneath it ride with it.

  4. 04

    Verify

    One reportability case still failing is enough to hold the queue back.

  5. 05

    Learn

    It is kept for good, and the triage rules are amended in the same commit.

  6. Learn → Detect

    The return edge. The next intake meets a suite one case longer.

Typical build scope

Twelve workstreams across six weeks

The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, sources, triage workflow, evaluation, integration, then production validation and handover.

Workstream

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

  1. 01Reportability triage and automation-boundary scope.Week 1
  2. 02Complaint, tracker and event-file sources.Week 1, Week 2
  3. 03Reportability-clock and qualified-reviewer mappings.Week 1, Week 2
  4. 04Complaint and report intake.Week 2
  5. 05Report dating and cluster binding.Week 2, Week 3
  6. 06Confidence scoring and regulatory routing.Week 3
  7. 07Qualified-reviewer decision workflow.Week 3, Week 4
  8. 08Event-file system integration.Week 5
  9. 09Reportability and severity cases.Week 4, Week 5
  10. 10Guardrails and routing controls.Week 4
  11. 11Report-trail instrumentation.Week 5, Week 6
  12. 12Deployment, documentation 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 tierPilotOne defect class, one cycleProductionProduction triage workflowAdvancedMultiple regimes / product lines
Introduced at Pilot
Triage to your reporting rules✓✓✓
Qualified-reviewer decision✓✓✓
Complaint-intake baseline✓✓✓
Introduced at Production
Reporting by defect class—✓✓
Regulatory review workflow in your systems—✓✓
Approved event-file write-back—✓✓
Tracker-and-file integration—✓✓
Introduced at Advanced
Multi-regime triage rules——✓
Cross-regime report packs——✓
High report volume——✓
Multi-regime deadline controls——✓
Build priceFrom $5,000From $8,000Custom quote
Final build price
Confirmed after discovery based on integrations, workflow complexity, report volume, routing 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.

  1. 01Your live defect classes and the duty each one can trigger

    Triage-rule capture and clock versioning

    Week 1

  2. 02Representative reports, complaint sets and past decisions

    Source binding, linking logic and the intake baseline

    Week 2

  3. 03Your escalation path and the reviewers it names

    Triage-rule mapping, clock binding and the automation boundary

    Week 1

  4. 04Access to relevant APIs, feeds or exports

    Complaint, tracker and event-file source assessment, then integration setup

    Week 2

  5. 05Reports you would not want dated

    Reportability cases and the failure round

    Week 4

  6. 06What no triage may conclude

    Confidence scoring, review routing, guardrails and hold controls

    Week 3

  7. 07A qualified person who decides reportability

    Handover to the qualified reviewer, then pilot and production validation

    Weeks 5–6

Nothing else is required

Deployment, documentation and Agent Care handover are ours.

Delivery timeline

Four phases across six weeks

Week five carries two bands because those two phases coincide, not because the column reads better.

Phase

W1

W2

W3

W4

W5

W6

  1. DiscoveryW1
  2. BuildW2 – W3
  3. EvaluateW4 – W5
  4. Pilot & LaunchW5 – W6

Week focus

W1Triage-rule discovery, clock versioning and the automation boundary

W2Complaint and tracker integration and the intake baseline

W3Report dating, linking logic and hold controls

W4Evaluation suite, reportability cases and failure-mode testing

W5Event-file integration, pilot reports and targeted corrections

W6One intake cycle run under the quality lead, then Agent Care handover

Reading the band
Each bar covers the weeks its own work is named for, and week five holds a pair.
At the end of W6
When the intake record validates, Agent Care takes the agent on.
Duration
Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.

Next step · Quality assurance AI agent

Build a defect triage agent around the report your last cycle filed below the line.

Show us one report you closed as not reportable. Not how quickly it was triaged. The hour an employee first knew, what the file holds of the reasoning, and whose name stands under the call. A clock that ran while nobody counted comes back as a case.

Typical delivery
4–6 weeks
Deployment
Your defect reports
Named reviewer
Clock first
After launch
Agent Care available

Nestack Agents · Defect triage

AGT-QA-01 · Agent Care available after launch

Build a defect triage agent

Four questions so the call starts from your actual setup — then pick a time that suits you.

About you
Your reports

No obligation. We’ll map the workflow, name the automation boundary and tell you if an agent is the wrong answer.