Insurance AI agent · Premium audit
Premium-Audit AI Agent (Workers' Comp & GL Exposure)
Collect the payroll and sales records an audit needs, reconcile them against what was estimated, apply the governing bureau's classification rules, compute the developed exposure and hold the workpaper for the auditor who concludes.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Auditor review
- Pre-conclusion
- After launch
- Agent Care
What this agent does
Assembles the workpaper, not the audit's conclusion
In
- 01
Payroll records arrive, and the agent pulls the governing bureau's rules for that state and class.
- 02
A policy expires, and the agent maps the exposure types the audit covers — payroll and sales.
Reason
- 03
Each record is reconciled against what was estimated at binding, not accepted as filed.
- 04
The bureau's classification and remuneration rules for that state and year are applied to the exposure.
- 05
Every developed-exposure figure is bound to the record it was reconciled from.
Decide
- 06
Findings that would move the experience mod are flagged, not folded into this year's number alone.
- 07
The assembled workpaper is routed to the premium-audit manager and the auditor for conclusion.
Out
- 08
The record, the reconciliation, the applied rules and the auditor's conclusion stay on the audit.
- 09
Execute write actions only inside the approval boundaries agreed during implementation.
Product statement
The agent gathers the records, reconciles them and computes the exposure; a qualified premium auditor concludes the audit, and the insured can see the basis.
Example workflow
One audit, records to conclusion
- 1Audit records receivedPayroll ledger, sales journal, subcontractor certificates or policy declarations
- 2Exposure reconciledRecords matched to the estimated exposure, each figure with its source and its governing bureau
- 3Exposure computedDeveloped exposure, applied class codes, flagged findings and confidence
- 4Controls appliedBureau-rule checks, reconciliation checks, mod-impact flags and confidence threshold
No human action required
Stages 1 to 4 run unaided, and nothing is billed at any of them — the agent is computing, and the auditor's lane opens at the confidence gate.
Goes to the premium-audit manager and the auditor to conclude.
Adds a bureau-rule compliance read first.
Auditor review
The workpaper is held with its reconciliation, its flagged findings and the confidence.
Approve · Edit · Request more records
Approved — released to conclude
- 6Policy system updatedOnly where write access and approval policy allow it
- 7Outcome evaluatedReconciliation accuracy, flagged-finding outcomes, mod impact and post-conclusion corrections
Edits
Every edit made before conclusion is counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Concluding the audit or signing it as final.
- Billing additional premium to the insured.
- Applying a non-compliance or estimated-audit charge.
- Assigning the final class code.
Automation boundary
Agent acts unaided
- ✓Collect the payroll, sales and unit records the audit requires.
- ✓Reconcile each record against what was estimated at binding.
- ✓Apply the governing bureau's classification and remuneration rules.
- ✓Flag findings that would move the mod, and hold.
Any write happens inside the boundaries agreed at implementation, never ahead of conclusion.
- Closing an insured's dispute of the audit.
- Accepting a subcontractor's certificate as proof, unverified.
- Deciding which bureau governs a state-and-line risk, unreviewed.
- Changes to the bureau rule library or remuneration tables.
Example output
One class code in the workpaper, annotated
Everything the agent computes is attached to the record it was drawn from.
Exposure-computation output · single audit
Illustrative example- Audit
- General contractor, subcontractor exposure
- Computed line
- The subcontractor's payroll is included because the certificate on file was never verified with the issuing carrier
- Stated exposure
- Payroll register
- Record type
- Class 5606 remuneration schedule
- Confidence
- 89%
- Attribution
- Auditor of record and licence ID
As received
Taken from the payroll register and the certificate on file — nothing on this side is written by the agent.
Records used
- Payroll register
- Subcontractor COI
- Prior audit history
Why this finding
The certificate on file was never confirmed with the issuing carrier.
What the score decides
Below the configured threshold before it reaches the approver.
Action
ApproveEditRequest more recordsValue
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 audit
From the payroll and sales records
Compute from the records
Draw on the reconciled records and the bureau's configured classification and remuneration rules.
One class code, several years
Routine reconciliation and exposure computation arrive already done.
Point the auditor at what needs
Findings that would move the mod and unverified subcontractor exposure are flagged, so review starts where judgment is needed.
Build an evidence trail
The exposure, the record it was drawn from and the auditor who concluded stay on the audit.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Policy & rating systems
- Guidewire
- Duck Creek
- NCCI Scopes
- bureau manuals
Payroll & wage data
- ADP
- Paychex
- State wage-report files
Subcontractor & COI records
- myCOI
- Certfocus
- Certificate-tracking systems
Agent
Premium-audit & exposure computation
- Reads the records
- Computes the exposure
- Holds for conclusion
Workpaper & correspondence
- Audit management systems
- Dispute & appeal tracking
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 workpaper
The controls sit one inside the next. What they let through is named in the map below.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Test whether the workpaper may proceed
- L3
- Puts the conclusion in the auditor's hands
- L4 – L5
- Keep the exposure and the record behind it
- L6
- Reverts to record collection when signals degrade
How Nestack evaluates it
Evaluate the audit workflow — not only the final workpaper.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the workpaper the insured sees
- E1Final-output evaluationDid every developed-exposure figure reconcile to its source record and the applied class code?
- E2Step-level evaluationDid the agent use the right bureau, classification rules and remuneration tables for that state and year?
- E3Tool evaluationDid it read and write the correct policy and the correct field?
- E4Confidence calibrationDo low-confidence findings actually attract more auditor read-backs?
- E5Slice evaluationHow does performance change across specific exposure types?
- E6Business outcomeHow many workpapers needed an auditor correction or a post-conclusion revision?
Floor — the outcome the audit answers for
Failure modes
Where each failure originates in the agent
Seven failure modes, placed at the stage each one originates.
Agent lifecycle
Direction of processing →
01 · Retrieval
1 mode
- PX-03
Outdated payroll caps
Executive-officer payroll is capped against last year's minimum and maximum figures.
Stage gathers
Payroll records, prior audit history and bureau rules
02 · Reasoning
2 modes
- PX-04
Bureau conflated
An independent-bureau state is classified under NCCI's rules, or the reverse.
- PX-06
Mod impact dropped
A classification fix is applied to this term but never flagged for the mod recalculation.
Stage proposes
Reconciled exposure, class code, confidence
03 · Tool / write
2 modes
- PX-02
Charge applied unchecked
A noncompliance charge posts without confirming that bureau currently allows one.
- PX-05
Basis not retained
Only computed totals are kept, and the source payroll records are dropped.
Stage writes
Only where write access and approval policy allow it
04 · Output
1 mode
- PX-01
Certificate unverified
The subcontractor's certificate is accepted without confirming coverage with the issuing carrier.
Stage returns
The workpaper the auditor concludes
05 · Change / Version
1 mode
- PX-07
Silent rule drift
A bureau-rule or remuneration-table update widens what the agent will apply unflagged.
Stage tracks
Model, prompt, bureau rules and remuneration tables
- Sev-1 · premium bills outside the boundary
- Sev-2 · a wrong finding reaches the workpaper
- Sev-3 · records degrade, goes to review
Affected slices
One class code can hide where rework concentrates
A single class code's audit-rework rate can look acceptable in aggregate while a handful of exposure types carry most of the rework. Nestack reports the rework rate by slice, not only in total.
Slice performance — reported separately, not only in aggregate
Illustrative example| Slice | Failure rate | Lift | Lift vs. threshold | Status |
|---|---|---|---|---|
| Policies with subcontractor exposure | 7.7% | 4.0× | Review | |
| Multi-state payroll across bureaus | 5.4% | 2.8× | Review | |
| Executive officers and owner payroll | 3.6% | 1.9× | Watch | |
| Single-class, single-state payroll | 1.9% | 0.7× | Normal |
Bar: audit-rework rate lift vs. the single-class baseline · scale 0–4.0× · tick at 2.0×
2 of 4 slices over threshold
Evidence-linked improvement
Every cycle leaves the class code better tested
A cycle ends when the misclassification is a case the next release has to catch. That suite is what the next audit assembled is measured against.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Audit-rework rate rises in a class-code slice.
- 02
Diagnose
The code that carried the wrong class for three years is traced back through the reconciliation until the cause narrows to one record.
- 03
Improve
Version-stamp the correction and attach the audits that exposed it.
- 04
Verify
The affected cases run again, and a fail stops the release.
- 05
Learn
It becomes a permanent test, and the classification rules move with it.
Learn → Detect
The return edge. Detection next time runs against a suite one audit 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, records, computation workflow, evaluation, integration, then production validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Audit workflow discovery and scope and boundary definition.Week 1
- 02Policy and payroll-source assessment.Week 1, Week 2
- 03Bureau and classification-rule mapping and rule mapping.Week 1, Week 2
- 04Record ingestion and normalisation.Week 2
- 05Computation logic and record binding.Week 2, Week 3
- 06Confidence scoring and finding routing.Week 3
- 07Auditor review workflow.Week 3, Week 4
- 08Policy-system and bureau-data integration.Week 5
- 09Classification and payroll cases.Week 4, Week 5
- 10Guardrails and conclusion controls.Week 4
- 11Audit-trail instrumentation.Week 5, Week 6
- 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 tier | PilotOne state, one class | ProductionProduction audit systems | AdvancedMultiple states / bureaus |
|---|---|---|---|
| Introduced at Pilot | |||
| Computation to your bureau and rules | ✓ | ✓ | ✓ |
| Auditor review | ✓ | ✓ | ✓ |
| Classification-accuracy baseline | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Reporting by class | — | ✓ | ✓ |
| Auditor workflow in your systems | — | ✓ | ✓ |
| Approved write-back | — | ✓ | ✓ |
| Policy-system integration | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multi-bureau rule complexity | — | — | ✓ |
| Multi-stage audit approvals | — | — | ✓ |
| High audit volume | — | — | ✓ |
| Multi-bureau audit controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on integrations, workflow complexity, audit volume, review 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 class-code list and bureau assignments
→Bureau-rule ingestion and class-code mapping
Week 102Representative past audits, including disputes
→Exposure-computation baseline and record-binding baseline
Week 203Your remuneration and classification rule tables
→Bureau-rule and classification mapping
Week 104Access to relevant APIs, feeds or exports
→Policy-system and bureau-data assessment, then integration setup
Week 205Audits you would not want billed
→Bureau cases and the evaluation suite
Week 406What no audit may assume
→Confidence scoring, finding routing, guardrails and conclusion controls
Week 307Named premium auditors to review workpapers
→Auditor review workflow, 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
The bands follow the work as it actually falls, so week 5 holds evaluation and pilot at once.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Audit workflow discovery, bureau mapping and the automation boundary
W2Policy and payroll integration and the computation baseline
W3Computation workflow, confidence logic and auditor-review controls
W4Evaluation suite, bureau-rule checks and failure-mode testing
W5Policy-system integration, pilot audits and targeted corrections
W6One audit cycle run under the premium-audit lead, then Agent Care handover
- Reading the band
- A bar covers the weeks its work is named in, and nothing else. The week 5 overlap is real, not padding.
- At the end of W6
- Once the cycle validates, monitoring moves to Agent Care.
- Duration
- Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Insurance AI agent
Build a premium-audit agent around your bureau rules and audit workflow.
Show us your class-code list, your states and a recent audit file. We'll assemble a sample workpaper against your own bureau rules as the next step, so you can see exactly what stays with the auditor.
- Typical delivery
- 4–6 weeks
- Deployment
- Your policy & payroll systems
- Auditor review
- Pre-conclusion
- After launch
- Agent Care available
Nestack Agents · Premium-audit & exposure computation
AGT-INS-19 · Agent Care available after launch