An AI coaching product without an AI policy is a liability. Here's exactly how we use AI, which models touch your data, what they can and can't do with it, and how we keep a human in the loop.
These aren't marketing claims. They're in our contracts and our architecture. If we ever changed them we'd tell you 90 days before, in writing.
Not our own models. Not our providers'. We use zero-retention, no-training API endpoints in every integration with a frontier model provider. Your content comes in, a result goes out, nothing is kept.
Every coaching suggestion is framed as a recommendation for a manager or rep to weigh — never as an automated decision. No hiring, firing, PIPs, or commission changes should be based solely on model output, and our product never offers to do that for you.
Every AI-generated insight is linked to the specific call moment that produced it. You can replay the audio, see the transcript, and judge the model's reasoning for yourself. No black box.
From the moment a call ends to the moment your manager sees a coaching insight. Five steps, one of which is where the model provider sits.
A meeting or call wraps up in Fathom, Fireflies, Otter, etc. — you own the recording.
We fetch the transcript using the OAuth token you granted. Encrypted in transit with TLS 1.3.
Transcript goes to the LLM provider under a zero-retention contract. Returned annotations come back in seconds.
Results and transcript stored in your tenant-isolated database. AES-256. Backed up hourly.
Manager sees coaching; rep sees feedback. Every insight linked back to the call moment.
We publish the list. When we change a provider — adding one, dropping one, or changing how we use one — account admins get 30 days' notice before the change rolls out.
Every AI product has failure modes. Pretending otherwise is how trust collapses. Here are the four we watch most carefully — what can go wrong, and what we do about it.
Coaching suggestions are grounded in the specific transcript — we quote the exact call moment for every insight. If the model can't cite a transcript snippet, we don't show the insight. Rep score cards show source quotes, not just judgments.
Accents, speech patterns, and cultural communication styles can skew model output. We benchmark methodology scoring across demographic groups quarterly and publish findings to Enterprise customers. Managers see raw transcripts, not just scores.
Our Acceptable Use forbids using the Service as the sole basis for firing, PIPs, or pay changes. Our product UI discourages it: leaderboards show trends, not grades; coaching is framed as suggestion, not directive.
Every call to a model provider is isolated per workspace. No cross-tenant prompting, no retrieval across customers, no shared embeddings. Architecturally, your data cannot appear in another customer's session.
Responsible AI is a system, not a slogan. Here's the operating rhythm behind what we just described.
A cross-functional group (engineering, product, legal, customer success) reviews every material change to how we use models — new providers, new prompting techniques, new automated behaviors.
30-day notice to admins before adding or removing a model provider, or changing how their data is processed.
Quarterly adversarial testing by an independent firm — prompt injection, jailbreak attempts, data-leak probes.
Methodology scoring benchmarked across demographic cohorts quarterly. Material disparities trigger prompt or model changes.
Enterprise customers can disable any AI feature (scoring, suggestions, summarization) workspace-wide if they need to.
Any material change to this page is published with a changelog. Prior versions available on request.