Two AIs Talked for Two Hours: Why Bot-to-Bot Call Loops Happen and How to Prevent Them

Two AIs accidentally talked to each other for 2 hours: what actually happened and why it matters

A Redditor tried to book a dentist appointment using a voice AI. The dental practice appears to use its own automated receptionist. Instead of reaching a human, the two bots politely conversed. For two hours. No booking, just endless confirmations and thanks, and a hefty bill for API usage.

They kept politely confirming things, asking for clarification, thanking each other, re-confirming previous confirmations.

You can read the original post here: Two AIs accidentally talked to each other for 2 hours.

It sounds absurd, but it’s a preview of where telephony is heading: AI agents calling AI agents. If you build or buy voice automations in the UK, this is a practical risk to design for, not a meme.

Why bot-to-bot call loops happen in voice AI

Politeness loops and “confirm again” prompts

Many agents are instructed to be helpful and confirm key details. When both sides are doing this without a strict protocol, they can lock into reciprocal “just to confirm…” loops. Ambiguous prompts like “confirm whenever unsure” are a recipe for infinite politeness.

Missing exit conditions and safety caps

Simple but common: no hard limit on call length, number of turns, or time without progress. If the system never reaches a “done” state, it never hangs up.

Mismatched turn-taking and end-of-utterance detection

Voice agents rely on voice activity detection (VAD) and “end-of-turn” heuristics. If both agents pause, barge in, or mis-detect silence at the wrong time, they keep stepping on each other and re-asking.

No shared conversation contract

Two independent bots don’t know when the other has finished an intent. Without a standard “I’m done” signal, each keeps seeking confirmation and never transitions to closure or escalation.

Lack of grounding in real systems

Agents that can’t actually query a calendar or complete a booking will often “slot-fill” forever. If the appointment date never validates, the agent keeps checking and re-checking.

ASR errors that never resolve

Automatic speech recognition (ASR) mistakes can trigger clarification spirals. Two AIs slightly mishearing each other prolong the loop, especially if thresholds for “confidence” are set too high.

Why this matters for UK users and organisations

• Costs can quietly spiral. Per-minute telephony plus LLM/ASR tokens adds up over long calls. Without spend caps and alarms, you can burn budget on nonsense.
• Data protection obligations still apply. Voice calls may capture personal data. Under UK GDPR and the Data Protection Act 2018, you need a lawful basis, clear transparency, appropriate retention, and processor agreements if vendors handle recordings or transcripts. See the ICO’s UK GDPR guidance.
• Call recording and monitoring need care. Inform callers if you record and explain why; don’t keep data longer than necessary.
• Nuisance call risk. Unattended loops may create abandoned or silent calls that frustrate patients and could raise Ofcom issues if it becomes systemic.
• Reputational damage. A “bot arguing with a bot” screenshot travels fast. For clinics, councils, and SMEs, this goes straight to trust.

How to prevent AI-to-AI call loops: practical guardrails

Design your agent like it will meet another agent tomorrow

• Set hard limits: maximum call duration, maximum turns, and maximum clarifications per field. On hitting any cap, escalate to a human or end the call gracefully.
• Adopt a “double-ack then end” rule: after a successful confirmation of all required fields, issue a final summary once, request an explicit confirm, then hang up with a reference number or next step.
• Use slot-filling with progress checks: define required fields (e.g., name, date, time, reason). If no new slot is filled after N turns, trigger a fallback path or hang up.
• Introduce a liveness or human-in-the-loop challenge: request a DTMF key press (“Press 1 to confirm”) or a randomly generated word repeated back. No response in 10 seconds? End call.
• Implement bot detection heuristics: check SIP headers/User-Agent, detect speech synthesis artefacts, or run a preamble handshake. Not perfect, but it reduces risk.
• Tune ASR and turn-taking: enable barge-in appropriately, set conservative end-of-utterance timeouts, and adjust silence detection so you don’t keep interrupting.
• Use a formal conversation contract: have your agent emit and listen for explicit “end_of_turn” or “completed_intent” markers in metadata if your stack supports it (e.g., in vendor-specific events or JSON side-channels).
• Ground actions in real systems: integrate calendars/CRMs so the agent can genuinely book. No write access? Make that clear and escalate early.
• Budget protection: set spend limits and usage alerts across telephony and model providers. Twilio’s Usage Triggers and OpenAI’s billing controls are good starting points.
• Monitoring and loop detection: flag calls with excessive confirmations, repeated phrases, or no state change. Send alerts and automatically terminate.

Prompts and policies that help

• Be explicit: “Never re-confirm the same field more than once unless the user changes it.”
• Prefer short, decisive turns over indefinite courtesy. “Confirm once; move on.”
• Define “done”. Provide a precise end-of-call policy including exact phrase and hang-up conditions.

Implementation notes and useful docs

• If you’re using realtime model APIs, look for native turn detection and event hooks you can assert against (e.g., OpenAI Realtime API).
• On telephony, design hang-up logic and machine detection carefully. See Twilio Programmable Voice for timers, DTMF collection, and call progress features.
• If you track call outcomes in spreadsheets or dashboards, keep PII minimal and access-controlled. For light-weight logging, here’s a practical guide on connecting ChatGPT and Google Sheets for structured records and budget tracking.

Checklist: if you’re buying a voice AI receptionist for the UK

• What are the default caps for call length, number of turns, and clarifications? Are they configurable?
• How does it detect and handle another bot or IVR? Is there a human escalation path?
• Where are audio and transcripts stored? Can data be processed/hosted in the UK or EEA? What’s the retention policy?
• Does the vendor provide UK GDPR-compliant documentation and a Data Processing Agreement?
• Are consent notices for recording and analytics clearly supported?
• What budget controls and alerts exist across both telephony and model usage?
• Can we integrate real booking systems to avoid endless “checking availability” loops?

Bottom line: automation needs guardrails, not just manners

Bot-to-bot call loops are not science fiction; they’re what happens when two polite, unbounded systems meet without a shared protocol. The fix isn’t “be more human” – it’s engineering: limits, state, grounding, and explicit endings.

If you’re rolling out voice AI in the UK, treat loop prevention and data protection as first-class requirements. Do that, and voice agents can save you time and money rather than burning credits on the world’s most courteous stalemate.