AI you can actually
answer for.
Auditable. Pausable. Private. Here's how Perpetua's agents make decisions — and how we keep humans firmly in the loop.
Six commitments, built into the product.
These aren't policies in a deck. They're enforced in our codebase and in our vendor contracts.
Human oversight by default
Every agent can be paused in one click. Clinical questions are escalated to humans automatically. Nothing irreversible runs without a human in the loop.
Complete audit trail
Every agent action — every call, booking, SMS, decision — is logged with inputs, outputs, reasoning, and the model version used.
No training on patient data
PHI is never used to train models. Zero-retention API contracts with our language model providers — your data isn't retained or reviewed.
Full transparency
See every call transcript, every decision, and why. No black boxes. If you want to know why an agent did something, the reasoning is one click away.
Narrow scope per agent
Each agent does one job. The Receptionist doesn't read chart notes. Collections doesn't access labs. Least-privilege enforced in code.
Continuous evaluation
Automated evals on every agent before deploying updates — hallucination rate, refusal rate, task completion, voice quality. Regressions block the deploy.
How an agent makes a decision.
Receive input
Agent receives a trigger (incoming call, scheduled eligibility check, no-show event) along with minimum context it needs — nothing more.
Reason through policy
The agent consults your practice's configured policy (hours, handoff rules, scripts, allowed actions) and your EHR state. It does not have authority beyond what you've granted.
Act or escalate
If the situation fits the agent's scope and confidence is high, the agent acts. If not, it escalates to the designated human — always visible, always logged.
Log everything
Full input, decision, action, and outcome are written to the audit log. You can replay any decision in the dashboard.
Six safeguards you can test today.
- 01Every agent has an OFF switch — instant, no confirmation dialogs, no delay.
- 02Clinical questions auto-escalate to a designated provider — the AI never gives medical advice.
- 03Agents cannot delete data. Ever. Even with admin permissions.
- 04Billing actions above $500 require human approval by default.
- 05After-hours triage routes emergency calls to the on-call provider within 60 seconds.
- 06Unclear situations trigger a handoff — it's cheaper to bother a human than to guess wrong.
Reducing bias, protecting every patient.
Bias in healthcare AI is well documented. We take it seriously in three specific ways:
- Operational scope only. Our agents handle scheduling, insurance, reminders, collections — not clinical decisions. Bias in those domains is narrower and more measurable.
- Routine disparate-impact testing. We sample outcomes across demographic dimensions (language, insurance type, appointment pattern) and review for unexpected gaps.
- Multilingual by design. Call handling works in English and Spanish at launch, with additional languages rolling out as we validate quality.
- Escalation over guess. Every agent is tuned to escalate when uncertain rather than optimize for completion rate. Patient experience beats agent metrics.
What you can do, right from the dashboard.
Pause any agent
Every agent has a toggle on its detail page and on the Agents Hub. Toggling off is immediate — no config loss, no billing change.
Review every decision
Open /call-log or the audit trail to see every action, who triggered it, and the model's reasoning.
Configure escalation
Decide exactly when an agent hands off to a human — by topic, confidence level, or caller profile.
Export your data
Transcripts, recordings, and audit logs are yours. Export as CSV, PDF, or via API — no retrieval fees.