FirstSeconds

An AI caller that rings your voice bot every night, records what it says in the first few seconds, and tells you whether it admitted to being an AI before it asked anything.

Problem

Since 2 August 2026, EU AI Act Article 50 has required a voice agent to tell callers it is an AI and who it works for, before the first substantive question, on every inbound call, outbound call and callback. Nobody at the company listens to the opening of their own bot after launch. A prompt tweak, a new language variant or an IVR reshuffle can quietly push the disclosure past "Can I take your account number?", and the first person to notice is a regulator or a complaining customer.

Target user

The person who owns the voice agent at a mid-sized company, or the agency that built it: usually a CX ops lead or a Vapi/Retell/Twilio integrator running five to fifty phone lines across EU markets. Their job is to keep the bot answering calls without a compliance incident. What they have today is a disclosure line buried in a system prompt and a launch-day manual test call. Nothing re-checks it after the next deploy, and nothing checks the transfer leg or the Spanish branch.

MVP scope

• Places scheduled test calls to a list of numbers using a scripted caller agent (TTS plus an LLM that plays the customer), with per-number scenarios: plain inbound, ask for billing, ask for a human, request a callback.
• Receives the client's outbound and callback calls on a provisioned test number, so those legs get checked too.
• Transcribes each call with word-level timestamps and runs an LLM judge against a fixed rubric: AI nature stated, principal named, both before the first substantive question, re-disclosed after a handover, a human path offered when asked.
• Stores audio, transcript and the judge's verdict with the exact second each check passed or failed, exportable as a dated PDF evidence pack.
• Runs in at least English, German, French and Spanish, judging disclosure in the language the bot actually spoke.
• Alerts in Slack or email when a check that passed yesterday fails today, with a link to the clip.

Monetization

B2B SaaS priced per monitored number. Roughly $79/mo per number for nightly runs in two languages, $299/mo for ten numbers with all scenarios, and an agency tier around $900/mo with white-label evidence packs. Telephony and model costs per test call run to cents, so margins hold if call volume stays at a few scenarios per night.

Why now

Article 50's disclosure duty took effect on 2 August 2026, and the Digital Omnibus (in force 27 July 2026) postponed the high-risk obligations but left Article 50 alone, so voice operators have no grace period. The 20 July 2026 guidelines are specific in ways a judge can test: disclosure "within the first seconds" on inbound calls, re-disclosure on every callback, the handover named at the moment it happens. In the US the rules are a state patchwork (California SB 1001, Maine, Utah for licensed professions, Texas TRAIGA for government), which makes a per-number, per-jurisdiction rubric more useful. Voice agent platforms are shipping disclosure templates, but a template doesn't test what the deployed bot actually says.

Risks & open questions

• Demand: enforcement of Article 50 runs through national authorities and I have no evidence of a single fine against a voice bot yet. Buyers may wait for the first case before paying for monitoring.
• Platform risk: Vapi, Retell or Twilio could add a built-in "disclosure check" to their test suites and make this a feature, not a product.
• Judge accuracy: "unless this is obvious" and "clearly enough" are legal judgments. The LLM verdict has to be framed as evidence for a human reviewer, never as legal advice, or liability lands on the tool.
• Build: bot-to-bot calls are fragile. The bot may hang up on silence, mishear synthetic speech, or route the test call to fraud screening, and some carriers flag repeated automated calls. Getting stable nightly runs across four languages solo is the hard part.
• Cost control: a customer with 50 numbers and five scenarios each is 250 calls a night; per-minute telephony and transcription have to stay inside the price tier.

Next step

Make 20 manual test calls to publicly listed EU voice bots, log when (or whether) each one discloses, and use the hit rate to pitch five voice-agent agencies.

Sources

• https://www.famulor.io/blog/eu-ai-act-article-50-what-your-ai-phone-agent-must-say (3 August 2026): Article 50 guidance for phone agents, covering timing for inbound vs outbound calls, re-disclosure on callbacks and naming the handover.
• https://www.famulor.io/blog/digital-omnibus-2026-new-ai-act-timeline-for-voice-agents: Digital Omnibus in force 27 July 2026, high-risk deadlines moved to 2 December 2027 while Article 50 disclosure was not delayed.
• https://justcall.io/blog/ai-voice-agent-disclosure-laws.html: US state disclosure requirements for AI callers (Maine, Utah, California SB 1001, Texas TRAIGA).
• https://aiagentsdirectory.com/news/ai-agents-news-brief-october-1-2026: 1 October 2026 agent news, including Classie Supervise for runtime monitoring of enterprise agents; agent-oversight tooling is a category vendors are launching into now.