<!-- Canonical: https://myoctopus.ai/trust/system-cards/variance-commentary -->

1.  [Home](/)
2.  /[Trust](/trust)
3.  /[Model System Cards](/trust/system-cards)
4.  /Variance Commentary AFI capability

Model System Card

# Model System Card — Variance Commentary AFI capability

Version 1.0

Last updated September 2026

Public document

This AFI capability detects budget-vs-actual variances, requests commentary from the responsible budget owner, and records the response as audit evidence.

Variance arithmetic runs as deterministic code. The language model drafts and interprets narrative only.

## 1\. Purpose and intended use

*   Detect budget-vs-actual variances above a customer-defined threshold.
*   Draft a plain-language question for the responsible budget owner about a specific variance.
*   Collect, structure, and store the owner's explanation against the affected line item.
*   Summarise collected commentary into a management-review narrative.

## 2\. Out-of-scope uses

*   Posting journal entries or changing figures in a source system.
*   Producing statutory financial statements or filings.
*   Making accounting policy judgements, including classification or capitalisation decisions.
*   Investment, credit, or hiring decisions.
*   Any use where the output is treated as an audited figure without human review.

## 3\. Data sources and access scope

| Source | Access |
| --- | --- |
| General ledger | Read-only, scoped to the accounts and entities the customer maps during onboarding. |
| Planning platform | Read-only access to budget and forecast versions (Anaplan, Pigment, Workday Adaptive, Oracle EPM). |
| ERP | Read-only access to transactional detail from SAP S/4HANA, Oracle NetSuite, or Workday. |
| Human responses | Commentary supplied by named budget owners, stored with author, timestamp, and the line item it explains. |
| Access control | Role-based access control. An AFI capability cannot read data its assigned role cannot read. |

## 4\. Computation boundary

| Layer | Responsibility |
| --- | --- |
| Deterministic code | Variance calculation, consolidation, currency conversion, threshold evaluation, and reconciliation to the general ledger. |
| Language model | Phrasing the question to the budget owner, interpreting a free-text answer, and drafting narrative. |
| Boundary rule | No figure shown to a user is generated by a language model. The model never computes or restates a number. |

## 5\. Evaluation approach

*   Reconciliation tests: every computed variance is reconciled back to source line items before it is surfaced.
*   Regression suite over historical closed periods, comparing computed variances against the customer's own signed-off figures.
*   Human review sampling: drafted messages are reviewed by a named approver before any outbound send during onboarding and at intervals thereafter.
*   Refusal tests: prompts requesting an unsupported judgement or an out-of-scope figure must produce a refusal and an escalation, not an answer.

## 6\. Known limitations

*   Commentary quality depends on the budget owner's answer. The AFI capability records what it is told; it does not verify the business reason independently.
*   Unmapped or newly created accounts are excluded until the customer maps them.
*   Late or restated source postings change variances; earlier commentary remains logged with its original timestamp rather than being rewritten.
*   Narrative drafting uses a third-party language model and can be imprecise in wording. Figures are unaffected because they are computed separately.

## 7\. Human approval gates

*   Every outbound message to a budget owner is approval-gated. A named human signs off before anything sends.
*   Threshold, distribution list, and message template changes require customer-side approval.
*   Escalation and reminder cadence is configured by the customer, not chosen by the model.

## 8\. Logging and lineage

*   Every action is logged with who, what, and when.
*   Every figure carries line-item lineage back to the general ledger.
*   Approval records, message content, and recipient are retained together as a single evidence chain.
*   Logs are exportable for internal audit and external auditor review.

## 9\. Escalation path

*   Unreconciled or ambiguous data halts the workflow and raises an internal data-quality item instead of publishing a number.
*   A refused or out-of-scope request is routed to the customer's named Octopus AI owner.
*   Service interruptions are communicated on the public status page.

## Questions and answers

### Does the Variance Commentary AFI capability calculate figures with a language model?

No. Arithmetic, consolidation, and currency conversion run as deterministic code against a reconciled data layer. The language model handles wording only.

### What access does the Variance Commentary AFI capability have to source systems?

Read-only access, scoped per source system, with role-based access control applied before data is retrieved.

### Is human approval required before the Variance Commentary AFI capability acts?

Yes. Every outbound message to a person is approval-gated and a named human signs off before anything sends.

### What is logged when the Variance Commentary AFI capability runs?

Every action is logged with who, what, and when, and every figure carries line-item lineage back to the general ledger. Logs are exportable for audit.

This system card describes documented product behaviour and controls. It is not a warranty of outcomes. Figures produced by any AFI capability remain subject to the customer's own review and sign-off.

## Related pages

*   [All Model System Cards](/trust/system-cards)
*   [The Octopus AI Constitution](/trust/constitution)
*   [Transparency Report](/trust/transparency-report)
*   [AI governance in finance (SR 11-7, DORA)](/ai-governance-finance)
*   [Trust Center](/trust)
