The problem
A wrong price or promise on a live call reaches the customer instantly
There's no review step between the agent generating a response and the customer hearing it. A misquoted price or an unauthorized promise is out before anyone can catch it — unless the check happens before the words are spoken.
Proxy-based guardrails add a pause a caller can hear
Most guardrail tools route the request through a separate server before and after the model call. On a chat interface that's a couple hundred milliseconds nobody notices. On a live call, it's dead air — the caller can tell something's off even if they can't say what.
You can't review a call transcript before the customer hears it
Text-based review workflows — flag, queue, have someone check it — don't work when the medium is a live conversation. Whatever's going to catch a bad response has to run in the same real-time path the call is already on.
How Zespan helps
Rules compile into a bundle your service evaluates locally
Guardrail rules — a price ceiling, a list of commitments the agent isn't authorized to make, a required disclosure — compile into a bundle synced to your own service ahead of time. No per-request call to an external server for a decision.
The check runs in-process, with no added latency
When the agent generates a response, the rule is evaluated inside your own service — the same process, no network round trip. The response is checked before it reaches the text-to-speech step, without the delay a proxy architecture would add.
Backtest against real call transcripts before going live
Before a new rule enforces anything on a live call, replay it against calls you've already handled. See what it would have caught — a wrong price, an unauthorized promise — and what it would have wrongly flagged, with the transcript behind every result.
A blocked or redacted response never reaches the caller
If a generated response violates a rule, it's blocked or corrected before the voice pipeline speaks it, not flagged for review after the fact. The call continues on a compliant response instead of pausing to hand off to a human.
Every call keeps a trace, cost, and quality record
Alongside enforcement, every call is traced end to end — latency per turn, cost, and quality score — so a pattern of near-misses or a drift in response quality shows up before it becomes a compliance issue.
Start free — 50K traces/month, no card needed
See every agent decision, tool call, and handoff in production. Setup takes a few minutes.
Start free →