Main site

Accounting AI agent · Expenses

Expense-Audit & Policy-Compliance AI Agent

Audit every claim against your expense policy, evidence each flag with the receipt, card line and rule behind it, and route it to a named reviewer — the agent never rejects a claim itself.

Typical delivery
4–6 weeks
Deployment
Your stack
Human review
Every flag
After launch
Agent Care

What this agent does

Audits every claim, evidences every flag

In

  1. 01

    Ingest expense claims, receipt images, corporate-card feeds, mileage logs and per-diem entries.

  2. 02

    Normalise merchant, date, currency and category, and read the claimant's grade and entitlement.

Reason

  1. 03

    Test each line against the policy version in force on the date the money was spent.

  2. 04

    Read receipt, claim and card line together to find duplicates, split claims and missing evidence.

  3. 05

    Check mileage, per-diem overlaps and VAT/GST reclaim eligibility against the rules for that country.

Decide

  1. 06

    Raise a flag carrying the rule, the evidence and a reason written for the person being flagged.

  2. 07

    Rank flags so reviewers see the defensible ones first, and pass the rest through unflagged.

Out

  1. 08

    Retain the rule version, evidence, reviewer decision and the reason the reviewer gave for it.

  2. 09

    Write the flag and its evidence only — clearing, rejecting and reimbursing stay with people.

Product statement

The agent flags and evidences inside the boundaries agreed during implementation; rejecting a claim, and anything that follows from it, stays with a named human.

Example workflow

One expense claim, end to end

  1. 1Claim receivedExpense platform, card feed, mileage log or receipt capture
  2. 2Evidence gatheredReceipt, card line, prior claims, and the claimant's grade and entitlement
  3. 3Policy evaluatedLimits, categories, class of travel, per-diem and mileage rates
  4. 4Duplicate and pattern checksThe same spend across cycles or across card and cash, and split amounts

No human action required

Stages 1 to 4 run without a person in the loop — nothing reaches the claimant or their manager before the gate.

5DecisionSplits on the confidence threshold
High confidence

Passes with no flag raised.

Low confidence

Held as a flag for review.

Human review

The claim is held with the rule it tripped, the evidence and a reason the claimant can read.

Clear · Uphold · Ask the claimant

Decided — handed back

  1. 6Expense platform updatedThe reviewer's decision and its reason, where write access allows it
  2. 7Outcome evaluatedFlag precision, flags overturned in review, and rates by team

Overturns

Flags cleared in review are counted, by team and by grade.

What should not run autonomously

Human approval stays in control

Outside the boundary — human approval required

8 items

  • Rejecting a claim or withholding reimbursement.
    • Calling a claim fraudulent.
    • Referring a person to HR or an investigation.
    • Granting or recording a policy exception.

  • Automation boundary

    Agent acts unaided

    • ✓Read the claim, the receipt and the matching card line together.
    • ✓Test each line against the policy version in force.
    • ✓Detect duplicates, split claims and per-diem overlaps.
    • ✓Raise a flag carrying its rule, its evidence and a written reason.

    Write actions run only inside the approval boundaries agreed during implementation. Rejecting a claim is not one of them.

    • Overriding a grade, per-diem or mileage entitlement.
    • Changing the expense policy or its thresholds.
    • First claims from a new joiner.
  • Any flag the claimant contests.

Example output

One expense claim, annotated

Every flag is attached to the line it came from and the rule it was tested against.

Audit output · single claim line

Illustrative example
Merchant
Hotel, overseas trip
Category
Accommodation
Amount
$412.00
Flag raised
Nightly cap exceeded
Confidence
92%
Routed to
The claimant's approving manager

As received

Taken from the expense platform and the card line it settled on — nothing on this side is inferred.

Evidence used

  • Itemised receipt
  • Card settlement line
  • City cap for that date

Why this flag

The city cap in the policy version in force that night — not a judgement about anyone.

What the score decides

Above the threshold the flag reaches a reviewer with its evidence. A human clears or upholds it.

Action

ClearUpholdAsk the claimant

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

All expense claims

From every submission channel

03Policy checks

Apply client-specific context

Use the policy version in force, grade entitlements and per-diem tables.

01Approved path

Check the population, not a sample

Every claim is tested against every rule, not the percentage a sampling audit draws.

02Human review

Send people only defensible flags

Reviewers get ranked, evidenced flags with a reason attached, rather than a queue of claims to read in full.

04

Build an evidence trail

Retain the rule version, evidence, confidence, evaluator result and the reviewer's decision and reason — on both paths.

Integrations

Typical integrations

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

  • Expense & ERP systems

    • SAP Concur
    • Expensify
    • Navan
    • NetSuite
    • ERP/GL APIs

  • Corporate-card feeds

    • Card transaction feeds
    • settlement files
    • Issuer APIs
    • statement imports

  • Receipt capture

    • Mobile capture
    • email intake
    • E-receipt feeds
    • scanned documents

Agent

Expense audit & policy compliance

  • Reads the evidence
  • Tests against policy
  • Evidences each flag

  • HR & approval workflow

    • Employee master
    • grade and entitlement
    • Approval systems
    • email

  • 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 flagged

Each control wraps the one inside it. A flag clears every layer before anyone is asked to look at it, and the decision sits outside all six.

L6 · Outermost — last line of defence

Inward → L1 · closest to the model

L6Rollback / safe modeRestrict automation if evaluations or fairness signals degrade.Roll back
L5Fairness monitoringCompare flag and overturn rates across teams, grades and claim types.Compare
L4TraceabilityRecord the rule, evidence, reviewer decision and the reason given.Record
L3Human decisionDefine who reviews a flag and who may act on what it says.Gate
L2Policy guardrailsTest only against the policy version in force on the date of spend.Restrict
L1Confidence thresholdsLow-confidence checks are logged, not raised against a person.Hold back
Model coreCheck result proposed — rule tested, evidence gathered, reason drafted and confidence
L1 – L2
Decide whether a flag is raised at all
L3
Decides who may act on what it says
L4 – L5
Keep the reason and the flag rates visible
L6
Pulls automation back when signals degrade

How Nestack evaluates it

Evaluate the full workflow — not only the final flag.

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

Surface — the flag the reviewer and the claimant see

  1. E1Final-output evaluationWas the flag correct, and was the right rule cited?
  2. E2Step-level evaluationDid the agent read the receipt, card line and entitlement correctly?
  3. E3Tool evaluationDid it read or write to the correct claim and claimant?
  4. E4Confidence calibrationDo low-confidence checks actually produce more false flags?
  5. E5Fairness evaluationDo flag and overturn rates differ across teams and grades?
  6. E6Business outcomeHow many flags survived review, and what reclaim was recovered?

Floor — the decision a named human signs

Failure modes

Where each failure originates in the agent

Seven failure modes plotted against the five stages of the agent lifecycle.

Agent lifecycle

Direction of processing →

  1. 01 · Capture / retrieval

    2 modes

    • EX-01

      Cropped receipt image

      The itemisation or the tax line is cut off.

    • EX-02

      Card and cash duplicate

      The same meal claimed once on the card feed and once in cash.

    Stage gathers

    Claim lines, receipts, card transactions and entitlements

  2. 02 · Policy evaluation

    2 modes

    • EX-03

      Split claim under a limit

      One dinner submitted as two lines under the cap.

    • EX-04

      Stale policy version

      Tested against today's limits, not that night's.

    Stage tests

    Rules, limits, per-diem and mileage rates, confidence

  3. 03 · Tool / write

    1 mode

    • EX-05

      Flag on the wrong line

      The result lands on another claim in the report.

    Stage writes

    Only where write access and approval policy allow it

  4. 04 · Output

    1 mode

    • EX-06

      Flag without its reason

      The rule fires but no evidence reaches review.

    Stage returns

    The flag, its evidence and the reason given

  5. 05 · Change / Version

    1 mode

    • EX-07

      Flag rate drifts by team

      A rule change flags one team more than the rest.

    Stage tracks

    Policy versions, model, prompt and rule changes

  • Sev-1 · a person is flagged unfairly
  • Sev-2 · a real breach passes unflagged
  • Sev-3 · input degrades, claim routes to review
See the complete Accounting failure-mode catalogue

Affected slices

A low flag rate can still hide an unfair one

A low overall flag rate can hide a few claim cohorts that carry most of the flags a reviewer later overturns — and one cohort flagged far more often than the rest. Nestack reports performance by slice, not only in total.

Slice performance — reported separately, not only in aggregate

Illustrative example
SliceFailure rateLiftLift vs. thresholdStatus
Overseas multi-currency trips5.4%3.4×Review
Card-and-cash claim pairs4.2%2.6×Review
New joiners' first claims3.0%1.9×Watch
Routine domestic claims1.1%0.7×Normal

Bar: overturn lift vs. domestic-claim baseline · scale 0–4.0× · tick marks the 2.0× review threshold

2 of 4 slices over threshold

Evidence-linked improvement

A wrong flag retires the rule behind it

Each cycle retires a rule that flagged the wrong people and keeps the claim that proved it, and the flag-rate comparison is re-run before it ships.

Improvement cycle · five stages

Switchback — the path turns at Improve and returns at Learn

  1. 01

    Detect

    Overturn rates rise, or one team's flag rate pulls away from the rest.

  2. 02

    Diagnose

    Traced to a rule, a policy version, a receipt read or a card-feed join.

  3. 03

    Improve

    The rule or its threshold is re-approved with finance and version-linked.

  4. 04

    Verify

    Re-run on held-out claims, then re-checked for flag rates by team.

  5. 05

    Learn

    The overturned claim becomes a regression case; the reason is rewritten.

  6. Learn → Detect

    The return edge. Every rule that flagged the wrong person leaves a test behind.

Typical build scope

Twelve workstreams across six weeks

The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, policy encoding, detection logic, evaluation and fairness testing, integration, then a shadow cycle and handover.

Workstream

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

  1. 01Workflow discovery and automation-boundary definition.Week 1
  2. 02Expense-platform and card-feed assessment.Week 1, Week 2
  3. 03Expense-policy encoding and rule library.Week 1, Week 2
  4. 04Claim, receipt and card-feed ingestion.Week 2
  5. 05Duplicate, split-claim and per-diem detection.Week 2, Week 3
  6. 06Mileage and VAT/GST reclaim checks.Week 3
  7. 07Confidence scoring and flag ranking.Week 3, Week 4
  8. 08Flag explanations written for claimants.Week 3, Week 4
  9. 09Reviewer workflow and contested flags.Week 4
  10. 10Evaluation suite, regression and fairness testing.Week 4, Week 5
  11. 11Expense-platform and GL integration.Week 5
  12. 12Shadow cycle, deployment and Agent Care handover.Week 5, 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 policy, one entityProductionProduction integrationAdvancedMultiple entities / countries
Introduced at Pilot
Policy checks and flags✓✓✓
Human decision on every flag✓✓✓
Baseline evaluation✓✓✓
Introduced at Production
Duplicate and split-claim checks—✓✓
Card-feed reconciliation—✓✓
Claimant-facing flag reasons—✓✓
Fairness reporting by slice—✓✓
Observability and evaluation—✓✓
Introduced at Advanced
Per-diem, mileage and VAT checks——✓
Multi-country policy and currency——✓
Enterprise controls——✓
Build priceFrom $5,000From $8,000Custom quote
Final build price
Confirmed after discovery based on policy complexity, integrations, claim volume, the number of countries 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 expense policy and its past versions

    Expense-policy encoding, version history and the rule library

    Week 1

  2. 02A representative sample of submitted claims and receipts

    Claim, receipt and card-feed ingestion, and the duplicate-detection baseline

    Week 2

  3. 03Corporate-card feeds and access to the expense platform

    Expense-platform and card-feed assessment, then integration setup

    Week 2

  4. 04Employee grade, entitlement and cost-centre data

    Grade-entitlement, per-diem and mileage rate checks

    Week 3

  5. 05Who decides a flag, and who hears a contested one

    Automation-boundary definition, reviewer workflow and contested flags

    Week 4

  6. 06Claims you have previously upheld and previously waived

    Evaluation suite, regression cases and fairness testing

    Week 4

  7. 07Named reviewers and a team willing to run a shadow cycle

    Flag explanations for claimants, then the shadow cycle 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

Phases are drawn over the weeks they actually occupy. Week 5 runs the agent in shadow against a live expense cycle while evaluation is still open.

Phase

W1

W2

W3

W4

W5

W6

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

Week focus

W1Workflow discovery, policy-version mapping and the automation boundary

W2Claim, receipt and card-feed ingestion, and the detection baseline

W3Detection logic, confidence scoring and flag explanations

W4Evaluation suite, fairness testing and the reviewer workflow

W5Expense-platform integration and a shadow run on live claims

W6Shadow flags decided by named reviewers, verified, then handover

Reading the band
Bars cover only the weeks their work is named in. Evaluation stays open into week 5 because the shadow run is what tests it.
At the end of W6
One expense cycle has been flagged in shadow and every flag decided by a named reviewer, then Agent Care takes over monitoring.
Duration
Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.

Next step · Accounting AI agent

Build an expense-audit agent around your policy and your people.

Show us your expense policy, your card feeds and who decides a contested flag. We'll encode one version of the policy, test it against claims you have already settled and compare the flags.

Typical delivery
4–6 weeks
Deployment
Your finance stack
Human review
Every flag decided
After launch
Agent Care available

Nestack Agents · Expense audit & policy compliance

AGT-ACC-08 · Agent Care available after launch

Build an expense-audit agent

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

About you
Your books

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