Control & safety, in full

Polylog operates a real person's account and talks to real leads. That is exactly why the system is designed so that a human is always one tap away from taking over — and why oversight is the default, not an add-on. This page is the long-form version of the safety model summarized on thelanding page.

Shadow review

Every AI-generated reply can be held in a review window — 30 seconds by default — before it is sent. A reviewer sees the reply in context and can let it pass, rewrite it, or cancel it. In the first week of a pilot this applies to every single message. As trust builds, review narrows to the conversations that matter most: sensitive topics, high-value leads, edge cases. The team decides the split, and can widen it again at any moment.

Automatic kill switch

Polylog continuously watches its own behavior against human baselines: reply speed, message volume, repetition patterns, activity hours. If any signal starts to look non-human, the system pauses itself — automatically, without waiting for anyone to notice — and stays paused until a person reviews what happened and explicitly resumes it.

Guardrails, twice

Constraints on what a persona may say — no promised outcomes, no medical or financial claims, no commitments the creator hasn't approved — are enforced in two independent places: in the instructions the model receivesand in a filter applied to its output after generation. Either layer alone can block a reply; both must pass for a message to send.

Full audit trail

Every message sent, every reviewer edit, every takeover, every pause and resume, every knowledge-base change — logged with timestamps and attribution. Agencies can reconstruct exactly what happened in any conversation at any point. The audit log is available to the client for the duration of the engagement.

Human takeover

An operator can take over any conversation instantly, mid-thread, from the same interface — the lead sees no seam. The AI stands down for that conversation until it is explicitly handed back.

Data handling

Conversation data is processed to run the persona and stored for the engagement plus a defined wind-down window. It is not used to train shared models. Deletion requests are honored. Details, retention periods and the data-processing terms live on the legal page.

Questions this page doesn't answer are usually pilot-specific — ask them directly via the pilot request form.