<!-- Canonical: https://myoctopus.ai/trust -->

1.  [Home](/)
2.  /Trust

Trust Center

# Trust, safety, and responsible AI at Octopus AI

Octopus AI is the AI backbone running inside enterprise finance functions — a governed Artificial Financial Intelligence (AFI) layer, not an autonomous chatbot. Finance work is regulated, audited, and unforgiving of unexplained numbers.

This hub documents how the system behaves, what it is allowed to do, and how each output can be verified. Every document here is public.

Three commitments define the architecture: arithmetic runs as deterministic code, every outbound action is approval-gated by a named human, and every figure traces back to its source.

## Safety and alignment

A language model never calculates a number in Octopus AI. Consolidation, currency conversion, and variance arithmetic run as deterministic code against a reconciled data layer. The model handles language only.

This removes the central risk of general-purpose AI in finance: a figure that changes between two identical questions.

*   Deterministic computation. The same inputs return the same figure every time.
*   Scoped, read-only access to source systems, enforced by role-based access control.
*   Documented out-of-scope uses for each AFI capability, with refusal behaviour tested.
*   Model drift cannot change a reported number, because the number is not model output.

## Steerability and human approval

Octopus AI is push-based: it initiates contact with budget owners rather than waiting to be asked. That capability is only safe when a human controls what leaves the system.

*   Every message to a budget owner is approval-gated. A named human signs off before anything sends.
*   Thresholds, recipients, templates, cadence, and assumptions are configured by the customer, not chosen by the model.
*   An AFI capability declines requests outside its documented scope and escalates them to a named human.
*   Customers can pause any AFI capability without pausing the platform.

## Transparency and auditability

*   Every action is logged: who, what, and when.
*   Every figure carries line-item lineage back to the general ledger.
*   Logs, approvals, and version history are exportable for internal audit and external auditor review.
*   Governance changes, incidents, and third-party model providers are published in the transparency report.

## Published documents

| Document | What it covers |
| --- | --- |
| Model System Cards | One card per AFI capability: intended use, out-of-scope uses, data access scope, deterministic boundary, evaluation, limitations, approval gates, logging, and escalation. |
| The Octopus AI Constitution | The operating principles every AFI capability follows, with how each principle is enforced technically and how it is verified. |
| Transparency Report | A dated, periodic report on governance changes, approval gates, incidents, model providers, data residency, and regulatory mapping. |
| Security & Compliance | Controls, certifications, encryption, and access management. |
| System Status | Live service availability and the external communication process during an interruption. |

## Regulatory mapping

*   SR 11-7 model risk: arithmetic runs as deterministic code rather than model output, so model drift cannot change a number.
*   DORA: the governed, logged, approval-gated architecture supports operational-resilience and third-party-risk documentation.
*   COSO and ICFR/SOX: control activities, information and communication, and monitoring map to the approval gates, audit trail, and lineage described here.

## Questions and answers

### Does a language model calculate the numbers Octopus AI reports?

No. Consolidation, currency conversion, and variance arithmetic run as deterministic code against a reconciled data layer. The language model handles language only.

### Can a person stop Octopus AI from contacting a budget owner?

Yes. Every outbound message is approval-gated. A named human signs off before anything sends, and any AFI capability can be paused.

### How is a reported figure verified?

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

### What access does Octopus AI have to source systems?

Read-only access, scoped per source system, with role-based access control. An AFI capability cannot read data its assigned role cannot read.

### Where are the individual model system cards?

Each AFI capability has its own published system card. Current cards cover Variance Commentary AFI capability, Forecasting AFI capability, Knowledge Q&A AFI capability.

## Related pages

*   [Model System Cards for every AFI capability](/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)
*   [AI governance for financial reporting (COSO, ICFR, SOX)](/solutions/ai-governance-for-financial-reporting)
*   [Technical briefs on architecture, cost, and controls](/technical)
*   [Security & Compliance](/security)
*   [System Status](/status)
