How Noesis CFO uses AI

Noesis CFO uses large language models (“LLMs”) to produce narrative commentary that accompanies deterministic financial calculations. This page discloses which model family is used, what data reaches the model, what the model is not permitted to do, and how you can opt out of AI-generated outputs. It is designed to satisfy B2B diligence (SIG Lite, CAIQ), the EU AI Act transparency obligations for limited-risk systems, California AB 2013 training-data provenance disclosure, and Colorado SB 24-205 consumer-facing AI disclosure.

1. AI provider and model family

The Platform uses the Anthropic Claude 4.7 model family from Anthropic, PBC (United States) via the official @anthropic-ai/sdk client. Specific model identifiers in active use are listed below. The current set is reflected in lib/ai/models.ts and can be rolled forward or pinned by operations without a code change via environment variables.

Default reasoning tier (tax-ask, insights, narratives): claude-sonnet-4-7
Fast tier (classification, blog drafting, access triage): claude-haiku-4-7
Vision tier (scanned tax-return OCR fallback): claude-sonnet-4-7
Founder tier (internal high-accuracy tools only): claude-opus-4-7

Our Anthropic account is configured with zero-retention terms: prompt inputs and completions are not retained by Anthropic beyond the request lifecycle and are not used to train Anthropic models. A second Anthropic-compatible endpoint (Vercel AI Gateway) is available as an auth/quota fallback; it is configured with the same no-training posture.

2. What data reaches the model

The Platform sends the model only the minimum context required to produce the requested output. Prompts are assembled server-side from your own workspace data, with the following classes of content permitted:

Your entity profile (entity type, jurisdiction, ownership) and your recent normalized P&L categories.
Filenames (not contents) of files you have uploaded, for context grounding.
Your free-form question or request, if you initiated the interaction (for example, the tax-ask engine).
Structured numeric results produced by our deterministic calculation engines, passed through read-only for narration.

The model does not receive raw uploaded file bytes (except the single vision-tier pass for scanned tax-return OCR, which is ephemeral and returns only recovered text), integration credentials, OAuth tokens, session cookies, HMAC secrets, bank account numbers, SSNs, or any customer personal identifiers beyond what appears in the entity profile.

3. What the model is not permitted to do

The three-layer output model (Computed Analysis | LLM Interpretation | Experienced CFO Review) is enforced on every page that renders AI commentary. The Experienced CFO Review layer is human advisory provided by Athena Consulting LLC (contracted separately) - not by the Platform operator and not by the LLM. The LLM layer is prompt-constrained and server-validated to:

Not compute or modify financial figures. Every number on the Platform originates from a deterministic, version-pinned engine and is passed through to the AI commentary read-only.
Not give legal, tax, investment, accounting, or fiduciary advice. AI commentary is informational narrative only. Methodology choices (inventory method, UNICAP scope, lease classification, ETR reasonability) are routed to a human CPA / advisor handoff path.
Not act on instructions embedded in uploaded documents, emails, or other untrusted sources. The injection-defense layer treats all non-chat content as data, never as instructions, and verifies any instruction-like content with the user.
Not cite tax authorities it cannot ground. Uncited tax claims (ASC 740/718/842, IRC sections, etc.) are automatically suppressed.
Not retain or train on your data. See Data Handling Terms Section 2 for the full contractual commitment.

4. How AI outputs are labeled

Every AI-generated surface in the Platform is visually labeled so you can tell deterministic calculation from model commentary:

A Computed Analysis badge on values produced by our deterministic engines.
An LLM Interpretation badge on any narrative or commentary block, paired with a link back to this page.
An Ask Your CPA follow-up line on every tax-ask answer and tax-intelligence commentary, so you know where the Platform stops and professional judgment starts.

5. Your rights and opt-out

You can opt out of AI-generated commentary at any time from workspace settings (Settings -> Profile). Opting out disables LLM Interpretation blocks across the Platform; Computed Analysis figures and reports remain available.

You may also request, at any time:

A list of AI-generated outputs associated with your workspace (ask-tax logs are retained in the SecurityEvent audit table).
Deletion of AI-generated commentary together with the source analyses via the self-serve purge flow at Settings -> Danger Zone.
A copy of the system prompt, guardrails, and validation logic that applied to a specific output, for diligence or audit purposes. Email support@noesiscfo-io.us.

6. Regulatory classification

Noesis CFO is an informational and analytics tool for business users. It does not take custody of funds, execute trades, originate credit, issue financial instruments, or render automated decisions with legal or similarly significant effects on natural persons. Under the EU AI Act we classify the LLM interpretation feature as a limited-risk system subject to transparency obligations, which this page satisfies. Under California AB 2013 and Colorado SB 24-205, we disclose AI use in this page and label outputs in-product.

Security questionnaires, SIG Lite, CAIQ, ISO 27001 vendor assessments, and custom DPA requests are accepted at support@noesiscfo-io.us. We counter-sign the DPA in one click at /legal/dpa.

This page is a policy disclosure and is incorporated by reference into the Terms of Service and Data Handling Terms. For the contractual formulations of the no-training commitment, the sub-processor list, and audit rights, see the Terms of Service, Data Handling Terms, and DPA.