Accounting AI agent · Management reporting
Variance-Analysis & Commentary Copilot (Evidence-Linked)
Explain the movements in a set of results — decompose each variance, propose a driver with its evidence and draft the commentary — with materiality thresholds, traceable figures and a named human signing the narrative.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Human review
- Every narrative
- After launch
- Agent Care
What this agent does
Automates the first draft of the explanation
In
- 01
Ingest actuals, budget, forecast and prior-period balances on the same reporting structure.
- 02
Read the detail behind each balance — journals, sub-ledger postings, volumes, rates and one-off items.
Reason
- 03
Size every movement against the client's own percentage and absolute materiality thresholds.
- 04
Decompose material movements into price, volume, mix, rate and timing effects, and state what is left over.
- 05
Separate translation, acquisition, reclassification and restatement effects from like-for-like movement.
Decide
- 06
Detect offsetting movements that net down, thin evidence and drivers the underlying detail cannot support.
- 07
Route unexplained residuals and anything past threshold to the named analyst or reviewer.
Out
- 08
Draft commentary in which every stated driver is linked to the figure and the source it came from.
- 09
Prepare the variance schedule and the draft narrative for review — publishing stays a human action.
Product statement
The agent proposes drivers and cites the figures behind them; the explanation that is published, and the judgement inside it, belong to a named human.
Example workflow
One movement, end to end
- 1Period figures receivedActuals, budget, forecast and prior period on the same reporting structure
- 2Detail gatheredJournals, sub-ledger postings, volumes, rates, one-off items and last period's commentary
- 3Movement decomposedPrice, volume, mix, rate and timing effects, translation stripped out, and the residual named
- 4Driver proposed and citedCandidate driver, the figures behind it, threshold and offsetting checks, and confidence
No human action required
Stages 1 to 4 run without a person in the loop — a movement that decomposes cleanly and evidences itself reaches the gate unaided.
Decomposed, evidenced and no residual left — follows the approved path.
Residual unexplained, evidence thin or movements offsetting — enters human review.
Human review
The movement is held with its decomposition, the evidence behind each proposed driver and confidence.
Approve · Correct · Request review
Approved — handed back
- 6Commentary drafted and routedOnly where write access and approval policy allow it; nothing is published
- 7Outcome evaluatedDriver accuracy, commentary kept or rewritten, coverage of material movements and reviewer edits
Overrides
Every reviewer edit to a driver or a sentence is counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Signing off commentary that reaches the board pack.
- Naming a driver the underlying detail cannot evidence.
- Attributing a movement to a named team or decision.
- Calling a movement one-off or non-recurring.
Automation boundary
Agent acts unaided
- ✓Size every movement against the agreed percentage and absolute thresholds.
- ✓Decompose material movements into their measurable components.
- ✓Propose candidate drivers and link each one to the figures behind it.
- ✓Route thin evidence and unexplained residuals to review.
Write actions run only inside the approval boundaries agreed during implementation. Publishing the commentary is not one of them.
- Stating a forward implication or full-year effect.
- Changing a materiality threshold or the comparative basis.
- Adding, redefining or dropping an adjusted measure.
- Restating a comparative the pack has already quoted.
Example output
One movement, annotated
Every driver the agent proposes is attached to the figures it was read from.
Variance output · single reporting line
Illustrative example- Reporting line
- Distribution costs
- Basis
- Actual vs budget, month
- Variance
- $264,000 adverse
- Proposed driver
- Volume, then carrier rate
- Confidence
- 91%
- Residual
- Named, not explained
As received
Taken from the ledger and the plan at the same reporting structure — nothing on this side is inferred.
Evidence cited
- Shipment volume series
- Carrier rate schedule
- Journals over threshold
Why this driver
Read off the movement's own decomposition, each figure linked — evidence, not cause.
What the score decides
Below the configured threshold the movement routes to review. The wording that reaches the pack is a reviewer's either way.
Action
ApproveCorrectRequest reviewValue
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 movement
From the ledger and the plan
Apply client-specific context
Use the client's own reporting structure, thresholds and comparative basis.
Reduce routine schedule and drafting
Movements that decompose cleanly and evidence themselves are sized, explained and drafted without manual assembly.
Focus analysts on what is unexplained
Residuals, thin evidence and offsetting movements move to review instead of every line being written by hand.
Build an evidence trail
Retain the decomposition, the figures behind each driver, the drafted wording, confidence, evaluator result and every reviewer edit — on both paths.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
ERP & general ledger
- NetSuite
- SAP
- Oracle
- Dynamics 365
- Sage Intacct
Planning & budgeting
- Anaplan
- Workday Adaptive
- Pigment
- Planful
- Vena
Consolidation & close
- OneStream
- CCH Tagetik
- Oracle FCCS
- Close-management tools
Agent
Variance analysis & commentary
- Reads the movement
- Proposes drivers
- Cites the evidence
Reporting & BI
- Board and management packs
- Power BI
- Tableau
- Snowflake
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 your board pack
Each control wraps the one inside it. A movement clears every layer before its wording is drafted, and publishing sits outside all six.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Decide whether the explanation may stand
- L3
- Decides who signs the published narrative
- L4 – L5
- Keep the evidence and the basis traceable
- L6
- Pulls automation back when signals degrade
How Nestack evaluates it
Evaluate the full workflow — not only the final wording.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the draft commentary the reviewer reads
- E1Final-output evaluationWas the proposed driver the one the business recognises?
- E2Step-level evaluationWas the movement sized and decomposed on the right basis?
- E3Tool evaluationDid it read the correct period, entity and comparative?
- E4Confidence calibrationDo low-confidence drivers actually get rewritten more often?
- E5Slice evaluationHow does performance change across specific reporting cohorts?
- E6Business outcomeWhat share of material movements was covered, and covered correctly?
Floor — the narrative 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 →
01 · Retrieval
2 modes
- VA-01
Comparative moved since reported
Prior figures were restated after the pack last quoted them.
- VA-02
Like-for-like basis missing
Translated and acquired figures arrive with nothing to strip out.
Stage gathers
Actuals, plan, comparatives and the detail behind them
02 · Decomposition
2 modes
- VA-03
Correlation named as driver
A series that moved alongside is proposed as the reason.
- VA-04
Offsetting movements netted
Two real movements cancel and the line decomposes to nothing.
Stage proposes
Driver candidates, cited figures and confidence
03 · Tool / write
1 mode
- VA-05
Reviewer edit overwritten
A regenerated draft replaces wording a reviewer already changed.
Stage writes
Only where write access and approval policy allow it
04 · Output
1 mode
- VA-06
Movement restated, not explained
The sentence repeats the figure and names no driver at all.
Stage returns
The draft commentary the reviewer and the board see
05 · Change / Version
1 mode
- VA-07
Reclassification read as movement
An account remapping lands as if the business had moved.
Stage tracks
Model, prompt, threshold and account-mapping changes
- Sev-1 · unsupported wording could reach the pack
- Sev-2 · a real movement goes unexplained
- Sev-3 · basis incomplete, routed to review
Affected slices
Commentary that reads well can hide concentrated risk
Aggregate commentary quality can look acceptable while a small number of reporting cohorts carry most of the rewrites and nearly all of the drivers that did not survive review. Nestack reports performance by slice, not only in total.
Slice performance — reported separately, not only in aggregate
Illustrative example| Slice | Failure rate | Lift | Lift vs. threshold | Status |
|---|---|---|---|---|
| Lines with offsetting movements | 5.7% | 3.8× | Review | |
| Restated or reclassified comparatives | 4.2% | 2.8× | Review | |
| Foreign-currency and acquired entities | 2.8% | 1.9× | Watch | |
| Steady-state repeat lines | 1.1% | 0.7× | Normal |
Bar: lift vs. steady-state baseline · scale 0–4.0× · tick marks the 2.0× review threshold
2 of 4 slices over threshold
Evidence-linked improvement
Each reviewer edit is kept as a label
The labels are a by-product of review — every sentence a reviewer rewrites records the driver proposed and the one the business recognised.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Rewrite rate, unexplained residuals or missed material lines rise in a cohort.
- 02
Diagnose
Traced to the comparative basis, a driver rule or a missing source.
- 03
Improve
The threshold, basis or driver rule is re-approved by the reporting owner.
- 04
Verify
The change is re-drafted against the periods whose commentary was rewritten.
- 05
Learn
The reviewer's own wording becomes the reference answer for that line.
Learn → Detect
The return edge. Next period's draft is scored against every sentence the last one had rewritten.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, decomposition, drafting and review, evaluation, then a parallel reporting cycle and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Reporting-line and materiality-threshold discovery.Week 1
- 02ERP and planning-system API assessment.Week 1, Week 2
- 03Comparative basis and adjusted-measure definitions.Week 1, Week 2
- 04Actuals, plan and comparative ingestion.Week 2
- 05Movement sizing and decomposition logic.Week 2, Week 3
- 06Evidence linking and source resolution.Week 3
- 07Driver proposal and confidence scoring.Week 3, Week 4
- 08Commentary drafting and house wording rules.Week 4
- 09Reviewer edit, sign-off and audit trail.Week 4, Week 5
- 10Evaluation suite and regression periods.Week 4, Week 5
- 11Reporting-pack assembly and write-back.Week 5
- 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 tier | PilotOne reporting pack | ProductionProduction integration | AdvancedMultiple entities / systems |
|---|---|---|---|
| Introduced at Pilot | |||
| Movement sizing and thresholds | ✓ | ✓ | ✓ |
| Human sign-off on every narrative | ✓ | ✓ | ✓ |
| Baseline evaluation | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Price, volume and mix decomposition | — | ✓ | ✓ |
| Every driver linked to its figures | — | ✓ | ✓ |
| Drafted commentary in house wording | — | ✓ | ✓ |
| Reviewer edit and audit trail | — | ✓ | ✓ |
| Observability and evaluation | — | ✓ | ✓ |
| Introduced at Advanced | |||
| FX and acquisition separation | — | — | ✓ |
| Multi-entity consolidated packs | — | — | ✓ |
| Enterprise controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on reporting structure, ERP, planning and consolidation integrations, entity count, decomposition depth 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 reporting structure and the lines you explain
→Reporting-line and materiality-threshold discovery
Week 102Percentage and absolute thresholds, per account
→Movement sizing rules and the review gate
Week 103Your comparative basis and adjusted-measure definitions
→Basis rules, and FX and acquisition separation
Week 204Access to the ledger, planning and consolidation systems
→API assessment, then actuals, plan and comparative ingestion
Week 205The volumes, rates and operational data behind the lines
→Decomposition logic, evidence linking and source resolution
Week 306Recent packs, including one whose commentary was rewritten
→Evaluation suite, regression periods and failure-mode testing
Week 407Named preparers, reviewers and the reporting owner
→Reviewer edit and sign-off workflow, then a parallel cycle
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 drafts a live period alongside your analysts, so the two sets of wording can be read side by side.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Reporting lines, thresholds and the comparative basis agreed
W2Ledger, plan and consolidation ingestion, and the first decomposition
W3Driver proposal, evidence linking and confidence logic
W4Evaluation against past packs, and failure-mode testing
W5Parallel drafting on a live period — commentary written, nothing published
W6Reviewer sign-off in production, then Agent Care handover
- Reading the band
- Bars span only the weeks their work is named in. Evaluation and the parallel drafting run genuinely overlap in week 5.
- At the end of W6
- One reporting cycle has been drafted in parallel and read line by line against what your team wrote, then Agent Care takes over monitoring.
- Duration
- Six-week plan shown · typical delivery 4–6 weeks, sequenced so the parallel run lands on a real reporting cycle.
Next step · Accounting AI agent
Build a variance and commentary copilot around your reporting cycle.
Show us your reporting lines, your materiality thresholds and a recent pack with the commentary your team wrote. We'll score a draft against what your team actually published last period and show you the gap.
- Typical delivery
- 4–6 weeks
- Deployment
- Your reporting stack
- Human review
- Every narrative signed
- After launch
- Agent Care available
Nestack Agents · Variance analysis & commentary
AGT-ACC-12 · Agent Care available after launch