Radar / Ideas / Disclose.ai: provable AI…

Disclose.ai: provable AI self-disclosure for every outbound voice call

daily ideaambitiousJEV confidence 0.572026-10-09
OutcomeEvery outbound AI voice call provably discloses that it is AI, with a tamper-proof audit trail.

The problem

Enterprises deploy AI voice agents at scale, but disclosure is a policy setting on the agent platform, not a verified fact. A viral video showed a hotel AI agent 'Jolene' insisting she was human until cornered; the FCC's 2024 AI-robocall ruling exposes enterprises to $500-$1,500 per-call liability for undisclosed AI voices. No middleware verifies disclosure on the enterprise's own outbound stack.

The idea

A compliance layer that sits between the enterprise's voice stack and the carrier. On outbound calls it verifies the AI agent actually played its disclosure script (audio fingerprint check), stamps a signed disclosure receipt, and logs it to an append-only audit trail. On inbound, it fingerprints the caller against known AI-voice models (fleming-style detection) to flag undisclosed agent callers spoofing humans. Regulators, auditors, and lawyers get one ledger that proves compliance per call.

Why now

Sierra's fleming-1 just made AI-agent phone detection a dedicated model; Cisco's Dialog is about to put persistent Claude agents on enterprise voice calls en masse; and the Personal Agent Protocol (Sierra, Meta, Genesys, NiCE) standardizes how cooperating agents identify themselves, leaving a gap for verifying the ones that don't.

What it combines

Sierra fleming-1 AI-agent phone detection (sierra-fleming-1-model-that-detects-ai-agents-calling-by-phone) + Cisco Dialog's persistent voice agents (cisco-dialog-claude-managed-agents-for-webex) as the enforcement surface + JEV typed judgments (jev-typesafe-ai) to decide disclosure-compliant vs violation per call. The mix matters because detection without enforcement is analytics, and enforcement without an audit trail is unverifiable.

MVP

Weekend MVP: a SIP/Twilio middleware that records outbound agent calls, checks for the presence of a configured disclosure phrase via speech-to-text, and writes signed pass/fail receipts to a simple log with a call-ID lookup page. Deliberately skip: inbound spoof detection, legal-grade immutable storage.

Distribution

B2B2C: embed in contact-center platforms (CCaaS vendors) and telephony aggregators that already own the enterprise voice stack; compliance buyers in banking, insurance, and healthcare pay per-call or per-seat.

Why it wins

Pindrop BotStopper and Reality Defender are detection-only tools for the receiving side (fraud defense); they don't sit on the enterprise's outbound stack or produce disclosure receipts. Balto coaches human agents, not AI ones. Nobody sells 'prove your AI agents disclosed' as a product.

Risks

Biggest risk: regulatory requirements differ per jurisdiction and shift often, so the product can lag the law. The MVP de-risks by targeting one jurisdiction (US TCPA disclosure rule) with a configurable rule pack, adding more rules as buyers ask.

Build it with

Repo to start from

disclose-ai: a Twilio/SIP middleware with the disclosure checker, signed receipt logger, and a minimal audit dashboard.

Evidence

Get the week's best AI launches, plus 3 ideas worth building

One email every Saturday. Ranked by traction, not hype. Free.