Your agent went off-script three weeks ago. Nobody can prove it.

Voiciee scores production voice-agent calls against the script and knowledge base they were supposed to follow. Every flagged call comes back with an attributed cause and a cost.

Have a live agent you’re slightly nervous about? I’ll score last week’s calls.

hello@voiciee.ai

01 Example

The same call, read two ways.

Transcript pass

Caller: No
Agent: Sorry, no to what?

Audio flag

Caller: No, I don’t want the extended warranty.
Agent: Sorry, no to what?

A transcript-only tool scores this as a pass. The agent asked a reasonable question given the input it received. The caller hung up.

The endpointer cut the turn at 400ms of silence. The failure happened two stages upstream of anything the transcript can see.

02 Output

The three outputs

Policy failure: The agent said or did something the script prohibits, or skipped something the script requires. Disclosures, refund authority, escalation triggers.
Unsupported claim: The agent asserted something the knowledge base does not support. Prices, availability, policy details, commitments it had no basis to make.
Cost per call: What each call cost, broken down by transcription, model, and speech. Which call types are quietly expensive.

03 Attribution

A failure count is not useful. A failure cause is.

Heard wrong: Transcription diverged from what the caller said. The agent behaved correctly on bad input. Fix: endpointing and vocabulary.
Retrieved wrong: Transcription was accurate; the knowledge base had a gap or returned the wrong passage. Fix: the KB, and often it’s the client’s.
Reasoned wrong: Input and retrieval were both fine and the agent still went off-script. Fix: the prompt.

Telling you there were twelve failures is worth nothing. Telling you nine were transcription, two were knowledge-base gaps, and one was the prompt tells you what to go fix on Monday.

04 Scope

What this is not

This is not pre-deployment testing. Simulated calls only contain the failures someone thought to script. Voiciee reads the calls that actually happened.

It runs across Vapi, Retell, Bland, and LiveKit. Your QA layer shouldn’t be owned by the platform you’re testing.

05 Setup

A webhook and a copy of your knowledge base.

Webhook: The end-of-call webhook you already have. Audio, transcript, and whatever you retrieved.
Knowledge base: A copy of the script and the documents the agent was supposed to be working from.

Voiciee does not sit in the live path. It does not add latency to the call.

Ambiguous calls go to a human. Never a silent pass.

06 Pricing

Pricing

Pilot: $2,000 for 30 days, one agent, up to 2,000 scored calls.
After that: $3,000/month.

Agencies get an extra seat so you can put the reports in front of your client.

No setup fee. No annual contract. Cancel after the pilot. Billed by invoice.