AI voice agents your compliance team can approve
Build, test, version and publish voice agents that answer and place calls within limits you define, and hand over to a person when it matters.
Transcript
auth: allowed · source: policy system
Call
Live · 01:42
Illustrative interface with sample data.
Voice AI needs more than a good voice
A convincing voice is the easy part. Production voice agents need to know when to stop, what they are allowed to say, which systems they may query and how to prove afterwards what happened on the call.
EAIBAP separates the agent’s identity, instructions, knowledge, tools, voice and policies, and ties them together in a versioned, immutable snapshot. The same discipline applies to telephony: tenant policy is kept apart from media handling, and outbound calls are checked against consent and policy before they dial.
From draft to production
Configure
Define identity, instructions, language, voice, knowledge sources and allowed tools.
Test
Run scripted scenarios in the test console. Inspect tool calls and expected outcomes before publishing.
Publish a version
Published versions are immutable snapshots. Deploy them to numbers, campaigns or workflows, with staged rollout.
Operate
Monitor quality, handoffs and cost. Pause or roll back instantly.
Guardrails built into every call
Tool allowlist
The agent can only call tools you approve. Arguments are validated against a schema and authorized by the server, independent of what the model asks for.
No invented facts
Pricing, balances, coverage and eligibility are fetched from approved systems. If data is unavailable, the agent falls back safely and escalates.
Human handoff
Transfer to a person or queue with a structured summary, a reason code and a recorded outcome.
Disclosure and recording prompts
Configure disclosure and consent prompts per region and workflow. Recording and transcription can be switched independently or disabled.
Kill switches
Stop one agent, a campaign, a provider, a tenant or all outbound calling in one step.
Call quality diagnostics
Latency, time to first audio, interruptions, recognition confidence and disconnect causes per call.
What happens when something fails
| If this fails | The agent |
|---|---|
| Speech recognition | Retries briefly, then offers keypad options or transfers to a person |
| Language model | Switches to a configured fallback; if unavailable, uses a scripted safe response and transfers |
| Speech synthesis | Uses a fallback voice or a recorded prompt, then transfers |
| Your system of record | Says it cannot confirm the detail, never guesses, and escalates |
| Policy service | Outbound calls are blocked until policy can be read (fail closed) |
Behavior is configurable per tenant and verified in test before production.
Telephony designed for operators
EAIBAP uses a FreeSWITCH-based media layer behind an adapter, so SIP trunks, caller-ID policy, number mapping and concurrency limits are configured without touching business logic.
- Dial permissions, destination restrictions and spend caps for fraud control
- Correlation IDs across SIP, AI runs, workflow events and CRM activities
- Real-PSTN testing for codecs, jitter, DTMF, transfer and hangup behavior
Web app (PWA) and public API
Role-based, tenant-scopedCore modules
Identity · CRM · Communications · AI orchestration · Knowledge · Workflow · Campaigns · Integrations · Billing · AuditData and messaging
PostgreSQL · Redis · RabbitMQ · object storageProviders
LLM · speech · telephony (SIP) · messaging · external systems of recordConceptual diagram.
Frequently asked questions
Can the voice agent take payments or approve claims?
No. Voice agents do not approve credit, adjudicate claims, change balances or take regulated decisions unless a separately approved integration and authorization model explicitly delegates that action. People and systems of record remain authoritative.
How do you stop an agent from saying the wrong thing?
Through layers: constrained instructions, retrieval from approved sources, tool allowlists with server-side authorization, evaluation suites before publishing, immutable versions, human handoff and kill switches. No single control is relied on alone.
Are calls recorded?
Only if you configure it. Recording and transcription are independent options and can be disabled by region, workflow or compliance profile.
Can we use our own SIP trunks?
The telephony adapter supports configurable SIP trunks and routes. Specific carrier support is confirmed during onboarding.
Do you support multiple languages?
Language, voice and pronunciation settings are configurable where the selected providers support them. Quality varies by language, so we evaluate each language before launch.
See it handle a real workflow.
Book a 30-minute walkthrough tailored to your industry and your systems.