SAFEGUARDS

Four safeguards. Non-negotiable. Included in every build.

Standard · Included in every build · never toggleable

Every phone agent from Lucid AI Labs has these four safeguards built in. Not as a premium add-on, not as an optional configuration, not as an upsell. They are the difference between an agent your callers tolerate and one your practice can put on the public number.

  • < 2s warm transfer
  • 0.75 confidence gate
  • 100% transcripts logged
  • Bot identity disclosed first sentence

01 · Human handoff

Always one transfer from a person

Every caller eventually reaches a moment the agent should not own alone: acute pain, an insurance dispute, a sensitive question. It does not stutter, apologise, or loop. It hands off.

A warm transfer to your front desk in under two seconds. The caller is never left alone, the reason is logged, and the agent stays on the line until the human picks up.

02 · Uncertainty stop

A guessing agent is worse than none

Patients act on what they hear. A confident wrong answer about availability, a protocol, or reimbursement does real harm. So every answer is scored against a calibrated confidence gate, default 0.75, adjustable per practice.

Below the line, the agent does not hedge into something that still sounds like an answer. It says it cannot answer reliably and offers the next real step: handoff, callback, or a concrete source.

03 · No gaslighting

It declines clearly instead of stalling

"I can help with that," followed by failure, is not a neutral outcome. It burns the caller's trust and lands as a review on your practice, not on the agent.

Scope is defined explicitly at build time. Anything outside it triggers a clean decline with a named alternative. Endless hold loops are ruled out architecturally, and out-of-scope promises are never made.

04 · Transparency

Identity disclosed first, everything logged

The GDPR requires it and so does trust. A caller who only realises at the end that they spoke to an AI feels deceived, and that feeling sticks to your practice. So the agent says what it is in the first sentence, naturally, not apologetically.

Full transcripts, per-turn confidence scores, and booking confirmations land in your admin dashboard in real time. Nothing hidden, no black-box decisions.

Why exactly these four. Not three. Not five.

Not safety features bolted on after an incident. Each one is a failure pattern that surfaced, was analysed, and was engineered out over nine months of building and running Lucid, a production AI system under daily real load.

Trust is fragile, reputation is permanent

A single bad call becomes a Google review. The safeguards make sure the agent never creates that moment, however unusual the request.

GDPR compliance is not optional

Safeguards 01 and 04 map onto GDPR Art. 22 (automated decisions) and the BNetzA rules for AI-assisted phone services. Legal obligations for German practices, not nice-to-haves.

Non-negotiable means non-negotiable

Asked to switch one off, the answer is no. An agent without them is a liability, not a product. They ship together or not at all.

  • 9 mo — building Lucid under daily real load
  • 240,049 — lines of code behind the system
  • ~8,000 — memory nodes (768d vectors)
  • BNetzA — registered telecoms operator, hosted in the EU

Ready for an agent on the public number? Let's scope it.

Send a briefing about your practice: current phone setup, call volume, where patients are lost. I reply with scope, questions, and an honest assessment of fit within one working day.

Send a briefing · To the phone agent