AI Use Policy

Plain-English: we use Anthropic’s Claude for proposal drafting, opportunity summarisation, allied-supplier briefing generation, and the intelligence-engine assessments. The CII scoring is deterministic, not AI.

Pathfinder + Enterprise tiers (LLM-backed)

• Proposal-agent. Drafts proposal sections against the SOW you’ve matched to + your cap statement. Claude Sonnet model, prompts versioned and logged.
• Opportunity summarisation. Generates a clean summary of a SAM.gov solicitation + a GO/MONITOR/NO_GO verdict with a short rationale.
• Allied-supplier briefings. Cached “Company Brief” for tech-scout drawer; cached for 30 days per supplier to limit redundant LLM spend.
• Intelligence assessments. The signal-clustering and assessment-generation pipeline uses Claude to compose editorial assessments from clustered raw signals.

All tiers (deterministic, not AI)

• CII (Country Instability Index). Pure scoring engine. No LLM. Sources: OFAC SDN, BIS Entity List, World Bank WGI, GDELT, US State Dept advisories.
• Appropriations relevance. Pure deterministic scoring. No LLM in the rank computation.
• Opportunity matching. Deterministic NAICS / set-aside / value rules. LLM only adds the optional GO/MONITOR/NO_GO verdict if the contractor opts in.

• Train external models on your data. Your cap statement, proposals, messages, watchlists, and pipeline content are never used to train, fine-tune, or improve any third-party LLM (Anthropic or otherwise). We do not opt into model-improvement programs that would expose your data.
• Sell or license your AI inputs/outputs. The content you submit to AI features remains yours; the AI outputs (proposal drafts, summaries) are licensed back to you under the Terms of Service.
• Use AI for fully-automated consequential decisions. AI features are designed as drafting and summary aids; the contractor reviews the output before any external-facing submission.
• Operate undisclosed AI features. Every AI-touching surface is labelled as such inside the product (small “AI” chip in the corner of the panel).

The current LLM provider is Anthropic, PBC. Inference happens under a zero-data-retention configuration: prompts and completions are not retained by Anthropic for model improvement. Anthropic’s commitments to enterprise customers are documented at anthropic.com/legal/commercial-terms.

• Tier gating. Scout-tier accounts have no AI features active. Pathfinder + Enterprise have AI features on by default; org admins can disable per feature in /settings.
• Per-document opt-out. When drafting in the proposal workspace, you can mark a document “no AI assist” to prevent the LLM from receiving it.
• Audit log. Every LLM call from your account is logged in the AI Usage Log (see /settings/billing for token breakdown). You can review what was sent.

LLMs hallucinate, omit, and occasionally produce content that looks plausible but is wrong. Treat AI outputs as drafts. Do not submit AI-generated proposal text to a federal agency without human review. Bridger disclaims liability for consequences of un-reviewed AI output as set out in the Warranty Disclaimer of the Terms of Service.