Main site

Accounting AI agent · Reconciliation

Reconciliation AI Assistant (Bank, Sub-Ledger & Balance Sheet)

Match statements and sub-ledgers to the general ledger, propose an explanation for every break and assemble the workpaper — with tolerances, evaluation, audit trails and preparer and reviewer sign-off before anything clears.

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

What this agent does

Automates the matching, not the sign-off

In

  1. 01

    Ingest bank statements, sub-ledger detail and general-ledger balances from feeds, camt.053/BAI2 files or API.

  2. 02

    Normalise dates, currency, references and counterparty identifiers across both sides.

Reason

  1. 03

    Match one to one, one to many and many to many inside the agreed amount and date tolerances.

  2. 04

    Separate timing differences already in flight from differences that are genuinely unexplained.

  3. 05

    Draw on standing reconciling items, prior clearing patterns and the account's own history.

Decide

  1. 06

    Classify every remaining break — timing, in transit, FX revaluation, fee, duplicate, error or unknown.

  2. 07

    Age each open item against the client's thresholds and route it to the named preparer.

Out

  1. 08

    Assemble the workpaper: matched sets, open breaks, proposed explanations and the evidence behind them.

  2. 09

    Propose adjusting journals for a preparer to raise — the agent posts nothing and clears nothing itself.

Product statement

The agent proposes matches and explains breaks; a named preparer clears them and a second person reviews the reconciliation.

Example workflow

One account, end to end

  1. 1Both sides receivedBank statement, sub-ledger extract and GL balance
  2. 2Sides normalisedDates, currency, references, counterparty identifiers and the opening position
  3. 3Matches proposedOne-to-one, one-to-many and many-to-many, in tolerance
  4. 4Breaks classifiedTiming, in transit, FX revaluation, fee, duplicate or unexplained — each with an age

No human action required

Stages 1 to 4 run without a person in the loop — the agent matches and classifies before anyone is asked to clear anything.

5DecisionSplits on the confidence threshold
High confidence

Matched in tolerance with a supported explanation — proposed in the workpaper.

Low confidence

Unmatched, aged or unexplained — held for the preparer.

Preparer review

The break is held with both sides, the proposed explanation, its age and the evidence.

Accept · Correct · Escalate

Cleared — handed back

  1. 6Workpaper assembledMatched sets, breaks and evidence — nothing posted
  2. 7Outcome evaluatedMatch rate, break rate, ageing profile and how many explanations the preparer kept

Overrides

Explanations the preparer rewrites are counted in the evaluation.

What should not run autonomously

Human approval stays in control

Outside the boundary — human approval required

8 items

  • Posting an adjusting journal to clear a break.
    • Closing or writing off a reconciling item.
    • Signing the reconciliation as preparer or reviewer.
    • Forcing a balance to agree with a plug.

  • Automation boundary

    Agent acts unaided

    • ✓Match both sides within the approved tolerances and date windows.
    • ✓Classify each break and age it against the client's thresholds.
    • ✓Draft an explanation and attach the evidence it rests on.
    • ✓Assemble the workpaper and route every break to its named preparer.

    Write actions run only inside the approval boundaries agreed during implementation. Clearing a break and posting its journal are not among them.

    • Carrying an unexplained item into the next period.
    • Suspense and clearing-account entries.
    • Intercompany differences between two entities' books.
  • Changes to matching rules, tolerances or ageing thresholds.

Example output

One break, annotated

Everything the agent proposes is attached to the two sides it came from.

Reconciliation output · single break

Illustrative example
Account
Main operating bank
Statement line
Credit, unmatched
Amount
$18,450.00
Proposed match
Two ledger receipts, one charge
Confidence
91%
Residual
Bank charge, journal proposed

As received

The statement line and the ledger side exactly as the sources delivered them — the agent normalises dates and references but invents no entry.

Evidence used

  • Statement line detail
  • Two open ledger receipts
  • Prior-period fee pattern

Why a residual

The receipts account for the line except the charge, so the difference is a fee the ledger has not booked — not a missing receipt.

What the score decides

Below the configured threshold no match is proposed and the line stays open. Either way a preparer clears the break and raises the journal.

Action

AcceptCorrectEscalate

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 lines, both sides

Statement, sub-ledger and GL

03Matching

Apply client-specific context

Use approved tolerances, standing reconciling items and the account's own clearing history.

01Approved path

Reduce routine tie-out work

Clean, in-tolerance lines are matched and set aside without anyone ticking the statement by hand.

02Human review

Focus the team on the breaks

Unmatched, aged and unexplained items reach the preparer with the work already done on them.

04

Build an evidence trail

Retain both sides, the match set, the proposed explanation, confidence, evaluator result and the preparer's decision — on both paths.

Integrations

Typical integrations

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

  • Banking & statements

    • Bank feeds
    • camt.053
    • BAI2
    • MT940
    • statement files
    • treasury systems

  • ERP & general ledger

    • NetSuite
    • SAP
    • Oracle
    • Dynamics 365
    • Sage
    • GL APIs

  • Sub-ledgers & sources

    • AR
    • AP
    • payroll
    • Fixed assets
    • card and payment processors

Agent

Reconciliation assistant

  • Reads both sides
  • Proposes matches
  • Explains breaks

  • Close & workpapers

    • BlackLine
    • FloQast
    • Trintech
    • Oracle ARCS
    • evidence stores

  • 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 a cleared break

Each control wraps the one inside it. A proposal clears every layer before it reaches the workpaper, and closing the break sits outside all six.

L6 · Outermost — last line of defence

Inward → L1 · closest to the model

L6Rollback / safe modeRestrict automation if evaluations or production signals degrade.Roll back
L5Version monitoringTrack model, prompt, matching-rule and tolerance changes.Track
L4TraceabilityRecord both sides, match set, explanation, evidence and overrides.Record
L3Preparer and reviewerDefine who clears a break and who signs the reconciliation.Gate
L2Match guardrailsEnforce tolerances, date windows and one-sided-entry checks.Restrict
L1Confidence thresholdsLow-confidence matches are not proposed; the line stays open.Hold open
Model coreMatch proposed — both sides, break class, residual and confidence
L1 – L2
Decide whether a match may be proposed
L3
Decides who clears and who signs it off
L4 – L5
Keep the workpaper and the rule history intact
L6
Pulls automation back when signals degrade

How Nestack evaluates it

Evaluate the full workflow — not only the match rate.

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

Surface — the workpaper the preparer opens

  1. E1Final-output evaluationWere the proposed matches and break classes correct?
  2. E2Step-level evaluationDid the agent apply the right tolerance and date window?
  3. E3Tool evaluationDid it read the correct account, period and statement version?
  4. E4Confidence calibrationDo low-confidence matches actually contain more errors?
  5. E5Slice evaluationHow does performance change across specific account cohorts?
  6. E6Business outcomeHow many breaks were explained correctly, and how many aged on?

Floor — the reconciliation a reviewer 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 · Retrieval

    2 modes

    • RC-01

      Partial statement or extract

      One side is read before the period's postings finish.

    • RC-02

      Aged items rolled forward

      Last period's unexplained items are carried in as settled.

    Stage gathers

    Statement lines, sub-ledger detail and GL balances

  2. 02 · Matching

    2 modes

    • RC-03

      Tolerance absorbs a break

      A real difference falls inside the allowed range.

    • RC-04

      Amount-only false match

      Unrelated items on two counterparties agree on value.

    Stage proposes

    Match sets, break class, residual and confidence

  3. 03 · Tool / write

    1 mode

    • RC-05

      Duplicate clearing

      The same item is matched twice across two runs.

    Stage writes

    Only where write access and approval policy allow it

  4. 04 · Output

    1 mode

    • RC-06

      Residual netted to tie

      An unexplained difference is absorbed so the account agrees.

    Stage returns

    The workpaper the preparer and reviewer see

  5. 05 · Change / Version

    1 mode

    • RC-07

      Silent tolerance drift

      A widened rule lifts the match rate and hides breaks.

    Stage tracks

    Matching rules, tolerances, model and prompt

  • Sev-1 · an unexplained difference clears
  • Sev-2 · wrong explanation reaches the workpaper
  • Sev-3 · input degrades, account routes to review
See the complete Accounting failure-mode catalogue

Affected slices

A high match rate can hide concentrated risk

Aggregate match quality can look acceptable while a small number of accounts carry most of the open breaks and nearly all of the unexplained value. Nestack reports performance by slice, not only in total.

Slice performance — reported separately, not only in aggregate

Illustrative example
SliceFailure rateLiftLift vs. thresholdStatus
Clearing and suspense accounts5.4%3.6×Review
Multi-currency revalued accounts4.0%2.7×Review
Intercompany pairs2.9%1.9×Watch
Stable low-volume accruals1.0%0.7×Normal

Bar: lift vs. stable-account baseline · scale 0–4.0× · tick marks the 2.0× review threshold

2 of 4 slices over threshold

Evidence-linked improvement

Closed periods are the regression set

Each cycle leaves a matching rule re-approved, a break class added or a tolerance changed — and a set of closed periods to replay it against.

Improvement cycle · five stages

Switchback — the path turns at Improve and returns at Learn

  1. 01

    Detect

    Break rate or ageing rises on an account.

  2. 02

    Diagnose

    Failure traced to the feed, a matching rule, a tolerance or a class.

  3. 03

    Improve

    The rule, tolerance or class is re-approved by its control owner and version-linked.

  4. 04

    Verify

    The change is re-run over signed periods.

  5. 05

    Learn

    Those periods become a standing replay set and the change enters the runbook.

  6. Learn → Detect

    The return edge. No tolerance widens until a replay shows which breaks it would have absorbed.

Typical build scope

Twelve workstreams across six weeks

The build scope read against the delivery timeline. Week structure follows the six-week plan — account scoping, data, matching and break handling, evaluation, integration, then a close run in parallel.

Workstream

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

  1. 01Account scoping, risk ranking and automation boundary.Week 1
  2. 02Source and statement-format assessment.Week 1, Week 2
  3. 03Tolerance, date-window and ageing-threshold mapping.Week 1, Week 2
  4. 04Ingestion and normalisation of both sides.Week 2
  5. 05Matching rules and one-to-many configuration.Week 2, Week 3
  6. 06Break classification and explanation drafting.Week 3
  7. 07Preparer and reviewer workflow and sign-off.Week 3, Week 4
  8. 08Workpaper assembly and evidence capture.Week 4
  9. 09Adjusting-journal proposals and routing.Week 4, Week 5
  10. 10Evaluation suite and closed-period replay.Week 4, Week 5
  11. 11Close-platform and general-ledger integration.Week 5
  12. 12Observability, 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 account groupProductionProduction integrationAdvancedMultiple entities / systems
Introduced at Pilot
Match and break proposals✓✓✓
Preparer and reviewer sign-off✓✓✓
Baseline evaluation✓✓✓
Introduced at Production
Client tolerances and match rules—✓✓
Workpaper assembly and evidence—✓✓
Adjusting-journal proposals—✓✓
Ageing and escalation thresholds—✓✓
Observability and evaluation—✓✓
Introduced at Advanced
Intercompany and multi-currency——✓
Multi-entity close orchestration——✓
Enterprise controls——✓
Build priceFrom $5,000From $8,000Custom quote
Final build price
Confirmed after discovery based on account count and risk ranking, statement formats and ledger integrations, transaction volume, sign-off 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. 01The account list, with risk ranking and named owners

    Account scoping, risk ranking and automation-boundary definition

    Week 1

  2. 02Current tolerances, date windows and ageing thresholds

    Tolerance, date-window and ageing-threshold mapping

    Week 1

  3. 03Statement files and sub-ledger extracts for a full period

    Ingestion and normalisation, and the matching baseline

    Week 2

  4. 04Access to bank feeds, ledger and sub-ledger APIs or exports

    Source-system and statement-format assessment, then integration setup

    Week 2

  5. 05Your break categories and standing reconciling items

    Break classification and explanation drafting

    Week 3

  6. 06Signed reconciliations from several recent closes

    Evaluation suite, closed-period replay and failure-mode testing

    Week 4

  7. 07Named preparers, reviewers and the escalation path

    Preparer and reviewer workflow, then the close run in parallel

    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 is the last evaluation week and the first week of a close run in parallel.

Phase

W1

W2

W3

W4

W5

W6

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

Week focus

W1Account scoping, risk ranking and the automation boundary

W2Statement and sub-ledger ingestion, and the matching baseline

W3Matching rules, break classes and preparer routing

W4Evaluation suite, closed-period replay and failure-mode testing

W5Ledger and close-platform integration, first live close in parallel

W6Reviewer comparison against the signed close, then Agent Care handover

Reading the band
Bars cover only the weeks their work is named in. Week 5 genuinely carries both.
At the end of W6
One close has run in parallel and been compared with the reconciliation your reviewer signed.
Duration
Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.

Next step · Accounting AI agent

Build a reconciliation assistant around your close.

Show us the accounts you reconcile, the statement and sub-ledger sources behind them, and who prepares and reviews each one. We'll start with the accounts that take longest to tie out, and show you where the agent stops.

Typical delivery
4–6 weeks
Deployment
Your finance stack
Human review
Every break and sign-off
After launch
Agent Care available

Nestack Agents · Reconciliation assistant

AGT-ACC-04 · Agent Care available after launch

Build a reconciliation assistant

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.