How we use AI
1. What the AI actually does here
The platform's job is to help your team capture opportunities, research them, score them against your own criteria, and produce documents about them. AI does the research and drafting inside a stage. It does not decide anything.
Feature — What the model does — What it is given
– Research and enrichment — Gathers and summarizes information about a target — The target record, and market reference data
– Scoring — Applies your configured criteria and weights, and explains its reasoning — The target record and your scoring configuration
– Assistant — Answers questions about your pipeline — Your question, plus context the platform assembles server-side from your own workspace
– Guided interview — Interviews you to build a scoring configuration — Your answers in the conversation
– Briefings and narration — Drafts a spoken digest — Your workspace summary data
– Deliverable documents — Drafts structured documents — The target record and the assembled source material
– Chat channel — Answers pipeline questions over Telegram or WhatsApp — Your message, and context assembled server-side
– Voice — Transcribes what you say; speaks replies aloud — Your audio; the reply text
2. Which providers receive what
The current list, with the data categories each receives, is published and maintained at /subprocessors. In summary:
– The AI model provider receives the prompt content for every feature in the table above. Under our commercial agreement it does not use that content to train its models, and inputs and outputs are deleted within 30 days. We do not have a zero-retention arrangement and do not claim one.
– OpenAI acts as the embeddings provider and receives the text being indexed or matched, so retrieval and search can work. When selected for a research beat, OpenAI also receives the scoped research prompt, NAICS scope and date, Census scope titles and exclusions, and any bounded repair instructions. The request carries no tenant, thesis, item, run, or snapshot identifier. Background processing temporarily stores response data with OpenAI for roughly ten minutes so it can be polled, even with store: false; the flow is not Zero Data Retention compatible. API content may separately be retained for abuse monitoring for up to 30 days, and we do not claim Zero Data Retention.
– The voice providers can receive spoken audio for transcription and reply text for synthesis, but voice is disabled for the approved release. Enabling it later requires a new disclosure review.
– Our error-tracking provider can receive scrubbed technical error reports, but Sentry diagnostics are disabled for the approved release. Enabling them later requires a new disclosure review.
The provider evidence reviewed for this release is recorded in the vendor register, the VCF-1783 approval record, and the VCF-2871 approval record. Where a provider's terms do not state a retention or training commitment, this page does not make one on its behalf.
3. What we do not do
– We do not train models on your data. No online training, no continuous training, and no customer data in shared model weights.
– We do not use your data to improve a model for other customers. The only corpus used for any training-adjacent purpose is a consented, anonymized set of design-partner transcripts, and consent is recorded and revocable.
– We do not let the system change its own behaviour. When we improve how the platform prompts a model, the change is proposed, tested against a fixed evaluation set, and promoted by a person. There is no self-modification path.
– We do not sell your data, and we do not use it for advertising.
4. What we do with your assistant history
We keep the questions you ask the assistant, the context the platform assembled, the answer, and any rating you gave it. We use that to measure answer quality and to find where the platform is weak — offline, and gated on evaluation. It is not training data.
5. Where the AI can be wrong, and what we do about it
Models produce confident, fluent text regardless of whether they are right. The platform is built to make that visible rather than to hide it:
– Facts carry their provenance. Measured values are recorded with their source, their unit, how current they are, and how confident we are — reported, estimated, unverified, or pending. An unmatched fact is published as pending rather than estimated into existence.
– Reference data is labelled. Where a value is modelled or unverified, it says so. Do not rely on it as reported fact.
– Scores are reproducible. A score records the criteria, weights, and rules that produced it, so a past score is always explained by the configuration in force at the time and never recomputed from today's settings.
– A person advances the work. Agents research and score within a stage; only a person moves an opportunity between stages, and that action is recorded.
None of that makes output correct. Verify anything you are going to act on. Output is not investment, legal, tax, or professional advice — see section 5 of the Terms.
6. Safety
The platform is a business tool, and its assistant is scoped to your pipeline. It is not an emergency, crisis, medical, or mental-health service.
The assistant does screen the messages you send it for language indicating self-harm, using fixed rules that run in the service handling your request. It does not use a model to do this, it does not read anything you have not sent to the assistant, and it does not build a profile of you. Where that screening matches, the assistant replies with a short, fixed, human-written message listing crisis helplines and answers nothing else for that message. That reply is the entire behaviour.
Nobody is notified. The platform does not alert, escalate, contact anyone on your behalf, or tell your organisation or its administrators. No human at Exonix is told. The message that triggered the reply is not stored with your identity. What is recorded is a count — which rule matched, on which surface, and when — with no message text and, for these categories, no user identity attached.
If a person is at risk, contact your local emergency number or a crisis service. The helplines the assistant shows are the right places to go; the assistant is not.
7. Changes
We may change models and providers. Material changes to who processes your data are notified to customers before they take effect, per the DPA, and this page and /subprocessors are updated at the same time.
Questions: support@exonix.com.