Skip to content

Trust Center

AI Transparency

HelmXP uses AI to draft briefing sections. This statement covers what AI systems we use, what data is sent to AI providers, what is never sent, how data is redacted, citation requirements for AI-generated content, and how we handle hallucinations.

Last reviewed: 2026-05-23. Maintained by HelmXP. Changes are published with a new review date.

Model training commitments

HelmXP customer data is never used to train or fine-tune foundation or third-party AI models.

This is a contractual commitment from Anthropic, enforced by the Azure AI Foundry enterprise agreement under which HelmXP operates. Anthropic's API terms explicitly prohibit using API inputs and outputs to train models. Azure AI Foundry's enterprise data protection terms additionally prohibit Microsoft from using customer data sent through the service for model training. Claude — the only foundation model HelmXP uses — performs inference only.

The full AI Transparency Statement document — including the specific contract clauses — is available on the Documents page.

Separately, the optional Voice feature builds an in-tenant writing-style model — private to your hub.

If a hub enables Voice, HelmXP builds a writing-style profile from that hub's own approved meeting minutes — and nothing else. The style model is private to the hub that built it: it is never shared across hubs, never pooled, and never used to improve Claude or any shared model. It is used for inference only — to match drafts to the hub's established voice — and the hub can delete it at any time, which removes the profile immediately.

AI systems used

Claude Sonnet

claude-sonnet-4-5 (pinned; updated only with explicit HelmXP approval)
Provider
Anthropic via Azure AI Foundry
Role
Primary drafting model — AI-drafted briefing sections, Q&A, and risk-scan analysis.
Data access
Receives the section context, the hub voice profile, peer NCUA 5300 statistics, and the operator prompt. Does not receive other hubs' data, raw financial ledger detail, or member-level PII.

Claude Haiku

claude-haiku-4-5 (pinned)
Provider
Anthropic via Azure AI Foundry
Role
Background tasks — smart summary generation, search index pre-processing, low-latency completions.
Data access
Receives condensed section or query context only. Same data-access constraints as Claude Sonnet.

What data is sent to AI providers

Each category is listed with whether it is included in AI model requests. Data marked Not sent never enters an AI prompt.

CategoryExamplesSent to AI?

Section text and prior memo context

Content of the briefing section being drafted, plus up to 5 prior approved memos for voice calibration.Sent

Hub voice profile

Aggregated stylistic features from approved memos — tone, sentence length, preferred terminology. Not the raw memo text.Sent

NCUA peer statistics

Aggregate NCUA 5300 peer benchmarks for the hub's asset tier and peer group. No individual peer data.Sent

AI model system prompt

The HelmXP-authored system prompt, including the seat label and any seat-specific voice instructions.Sent

Individual member records

HelmXP works with aggregated financial metrics, not individual member records. Member-level data is never included in AI prompts.

Member account data, loan records, names, account numbers, payment histories.Not sent

Full GL trial balance detail

Only summary-level financial metrics (net interest margin, loan-to-share ratio, asset totals) are included in prompts.

Line-item general ledger entries, individual account balances below summary level.Not sent

Personnel or HR data

HR data is not ingested by HelmXP and is never sent to any AI provider.

Employee salaries, performance reviews, hiring records.Not sent

Data from other hubs

Each hub's RAG corpus is isolated by tenant prefix and RLS. No cross-hub data is ever included in a prompt.

Any data from institutions other than the hub whose section is being drafted.Not sent

Redaction practices

PII in uploaded documents

When a seat holder uploads a document to the hub knowledge base, HelmXP applies a pattern-based redaction pass before chunking for RAG. Detected patterns include US Social Security Numbers, 16-digit card numbers, and common financial-account number formats. Redacted with [REDACTED-<type>] markers.

System prompt separation

Seat-specific voice and style instructions are placed in the system prompt turn, not in the user-turn context. This prevents the model from treating operator instructions as user-generated content.

Prompt logging

Every AI request is logged: model name, prompt hash (SHA-256 of the prompt — not the text), token count, and the seat ID that initiated the request. The full prompt text is not stored; only the hash, for audit deduplication.

Error traces

If an AI request fails, the error trace logged to Azure Application Insights is scrubbed of prompt content. No partial completions are retained in error logs.

Citation requirements

Every AI-drafted briefing section must carry citations grounding its claims in the hub's actual data before a seat holder can sign it off. Citation enforcement is built into the AI prompt contract and validated at the application layer.

  • Every financial claim cites its source. Sources include: NCUA 5300 call report line items (with line number), the GL trial balance, the ALM model output, or a named integration export.
  • Citations are machine-readable. Citations are structured objects in the AI output — not inline footnotes. The application validates that each claim resolves to a named source.
  • Uncited drafts are not accepted. If the model returns claims that cannot be grounded, the application discards the draft and prompts the seat holder to re-run with narrower context.
  • AI drafts are marked visibly. Every AI-drafted section displays a persistent “AI draft · review before signing off” marker, which clears only when the seat holder explicitly signs off.

Hallucination handling

HelmXP applies three layers of hallucination mitigation. None are a guarantee — the seat holder's review is the final control.

  1. Grounding in source data

    AI prompts are constructed from the hub's actual data — NCUA 5300 figures, ALM outputs, GL summaries. The model drafts commentary on provided numbers, not from parametric memory. Specific figures the model cannot source are flagged with [ESTIMATE] markers.

  2. Citation validation at application layer

    The application validates each citation in the draft against the source data that was included in the prompt. Claims citing a line item not provided are rejected before the draft reaches the seat holder.

  3. Seat holder sign-off

    No AI-drafted content reaches the briefing without a seat holder explicitly signing off. The sign-off UI reminds the seat holder to verify all numeric claims. The Logbook records who signed off and when.

To report a specific case of inaccurate or harmful AI-generated content, contact security@helmxp.com. HelmXP reviews all flagged AI outputs within 2 business days.

The full AI Transparency Statement — including specific contractual clauses from Anthropic and Microsoft governing the model training commitments — is available as a downloadable document on the Documents page.