| Anthropic (Claude) | United States | Loan narrative text, business plan summaries, anonymised borrower questions | Generates appraisal summaries, analyses application evidence, drafts officer responses | Zero retention enterprise tier, no training on customer data, requests not stored beyond response |
| OpenAI (GPT) | United States | Conversational chat turns, voice transcripts, mathematical representations of public documents | Powers chat-style officer assistant and Whisper speech recognition | API requests retained 30 days for abuse monitoring, then deleted, no training on API data |
| DeepSeek | China (Hangzhou) | Anonymised batch inference inputs, KPI commentary prompts, no PII | Cost-optimised batch jobs and background commentary, PII stripped before send | Standard tier 30 day API log, opt-out of training requested via account settings |
| ElevenLabs (Scribe v2 STT) | United States, European Union | Voice audio uploads, transcripts of borrower and officer speech | Speech-to-text for Swahili and English voice notes | Audio not retained beyond transcript return, transcripts deleted within 24 hours |
| ElevenLabs (Eleven v3 TTS) | United States, European Union | Text prompts only, no borrower PII, generated audio returned to caller | Voice synthesis for officer responses and accessibility playback | Generated audio not cached on provider, no training on prompts |
| Brave Search | United States | Public web search queries derived from anonymised research topics | Background research on the public web, borrower data is never sent, only topic keywords | Brave does not log identifiable searches, queries discarded after response |
| SerpAPI | United States | Public web search queries, identical pattern to the Brave path | Background research fallback when Brave is unavailable or rate-limited | Standard SerpAPI log of 30 days, no borrower data in queries |