AI Disclosure
How IMFA Agents uses AI models, which providers process which data, and what AI cannot promise.
Last updated:
1. Scope
This page explains how IMFA Agents, operated by IMFA Labs LLC ("IMFA," "we"), uses artificial intelligence: which models process which data, through which gateway, with which protections, and what AI cannot promise. It adds detail to our Privacy Policy and Data Processing Agreement.
2. Where AI is used
- Agent replies: what a visitor writes, the agent's instructions, and the relevant passages of its knowledge are sent to an AI model, which writes the reply.
- The builder assistant: what you ask it, and the parts of your agent it needs, are sent to an AI model, which proposes or makes the changes.
- Decisions: a decision model makes short, bounded judgments: which way to route a message, whether a reply matches its sources, how warm a lead is, how a conversation ended, and the safety checks in section 6.
- Voice: during a voice call, audio is streamed in real time to a voice model that listens and speaks. During builder dictation, your voice is turned into text.
- Knowledge: documents and web pages are turned into numeric representations (embeddings) so the agent can find relevant passages, and the decision model keeps the passages that answer the question.
3. Gateways and providers
We reach AI models through two separate gateways:
| Gateway | Used for | Data retention |
|---|---|---|
| Convex AI Gateway (Convex, Inc.) | Agent replies, the builder assistant, decisions (including ranking knowledge search results), query translation, builder dictation (ElevenLabs), knowledge embeddings (Google Gemini) | We ask for zero data retention on every request: it may only reach a provider endpoint that keeps nothing after answering. There are two exceptions: builder dictation (the gateway cannot take that request for audio) and the embeddings of current knowledge indexes (we checked on November 11, 2026: that model has no zero data retention endpoint, so the provider's own data terms apply) |
| Vercel AI Gateway (Vercel Inc.) | Live voice calls (Google Gemini) | We checked on October 31, 2026: this model has no zero data retention route, so the provider's own data terms apply. We switch it on as soon as one is offered |
Models available on the Convex AI Gateway: Anthropic (Claude), DeepSeek, Qwen (Alibaba Cloud), Z.ai (GLM), OpenAI, Google (Gemini), Moonshot AI (Kimi), Mistral AI, and xAI (Grok). The decision model is TypeSafe's Jev, and builder dictation uses ElevenLabs Scribe.
Which model is used. By default, agents answer with Anthropic's Claude Haiku. If it is unavailable, the reply may be generated by DeepSeek, Qwen, or Google Gemini instead. An owner can choose another supported model for an agent. The builder assistant uses DeepSeek first on standard accounts and Claude Haiku on Pro accounts or when you attach a file. The full list of providers and their locations is on our Subprocessors page.
4. Data protection
- Zero data retention applies to the Convex AI Gateway requests above, except builder dictation and the embeddings of current knowledge indexes. It does not cover live voice on the Vercel AI Gateway, the providers' abuse monitoring where their terms require it, or what IMFA itself stores to run the service (for example the conversation, kept as described in the Privacy Policy).
- No training. We do not use your content or your visitors' conversations to train AI models, and we choose provider options that exclude training where providers let us choose.
- Less data to the decision model. Contact details are removed before a lead is qualified or a conversation is summarized. Decisions are stored as scores and labels, without the conversation text.
- Instructions stay private. Agents are told never to reveal their instructions, and replies that would quote them are cut.
- Locations vary. Providers process requests in their own data centers, which may be in the United States, China (for some models), or elsewhere. See section 10 of the Privacy Policy.
5. What AI cannot promise
AI output can be wrong, incomplete, outdated, or biased, even when it sounds confident. Agents are designed to answer only from their business's own information and to say when they do not know, but they can still make mistakes. Agents and the builder do not give legal, medical, financial, or other professional advice; do not rely on AI output alone for decisions that matter. Owners are responsible for reviewing what their agents say and do, and for telling people they are talking to an AI where the law requires it (agents say they are an automated assistant when asked, and can add an AI notice to the first reply on messaging channels).
6. Safety checks
The decision model also checks content against our Acceptable Use Policy: agent instructions, skills, products, knowledge, builder requests, and visitor messages. It can refuse a request with a short fixed reply, hide a product, pause publishing, and warn or block an abusive visitor. These checks are automated and can be wrong in both directions; people review account holds, and owners can lift visitor blocks. The checks are weakest on Moroccan Darija written in Latin letters, where the agent's own rules still apply.
7. Human oversight
You can join any conversation and take it over from the agent; while you do, the agent stops replying. You can review conversations, correct leads, and change or pause an agent at any time. The builder assistant asks for your approval before changes that need it.
8. Contact
Questions about AI at IMFA: privacy@imfa.io.
