Agents prepare. People decide.
Agents do the preparation: they read, check, draft, call and chase. People make the decisions that need judgment, a license or accountability. This article names the four decisions that always stay with your people, the laws that require a human route, how staff test agents before go-live, how the pilot is reviewed, and what monitoring looks like afterwards. It ends with published evidence on how often a clinical voice agent hands calls to people.
Four decisions that stay with your people
Clinical calls
When a clinical concern comes up on a call, the agent hands it to your on-call clinician under your own escalation protocol. Agents are configured not to make clinical decisions or give clinical advice, and they are tested for it before go-live.
Code changes
Agents suggest codes with the part of the note that supports them. A certified coder approves any change that raises reimbursement. A clinician signs every medical-necessity appeal.
Sign-offs
Your team signs off each step before it goes live, and each new request type, site or step needs a named sign-off. Any line can switch back to today's process at any time.
The morning review
Each morning you see the night's work: what agents did, what they passed to people, and the evidence for each action.
Rules that require a human route
Some of these checkpoints are also legal requirements. California's AB 3030, in effect since 1 January 2025, requires AI-generated patient messages about clinical information to say they were generated by AI and to explain how to reach a human provider, unless a licensed provider reviewed the message before it was sent (Duane Morris, 2024). Utah's AI Policy Act requires licensed occupations, including many healthcare professions, to disclose generative AI use (Davis Wright Tremaine, 2024); a 2025 amendment limited that duty to "high-risk" interactions, which collect sensitive data such as health information and give personalized advice (Davis Polk, 2025). Your counsel decides which laws apply. Our agents disclose that they are automated everywhere, and a caller can ask for a person at any point.
Billing carries its own risk. CMS says most Medicare fee-for-service improper payments in fiscal year 2024 came from insufficient documentation or documentation that did not show medical necessity (CMS, 2024). A coder who checks each suggested code against the note is the control that keeps an agent's suggestion from becoming a billing error.
Before go-live: staff test the agents
Before any patient reaches the system, your staff listen to and score test calls built from real scenarios. The tests include the limits: a caller asking for clinical advice, a request the agent should pass on, a clinical concern that must reach a clinician. We also test that agents never sign a note and never change a code on their own. A workflow goes live only after the people who will supervise it are satisfied with the results.
During the pilot
The pilot starts with the lowest-risk work, such as after-hours calls, and typically lasts about six weeks. Each week we review a sample of transcripts with your team, along with misrouted calls, booking errors and feedback from patients and staff. Any line can switch back to today's process at once while a problem is fixed. The pilot ends with a go or no-go decision against the baseline you agreed.
After go-live
Scheduled routines watch what was built and alert a named owner on your side when something drifts: a payer portal changes its layout, the share of calls passed to staff jumps, a consent check starts failing. Routines typically run in shadow for 14 days before they go live, so their alerts can be checked against what actually happened. Routines prepare packets and alerts; people make every reply, routing change or correction. You get a monthly scorecard against the baseline.
If something fails, the alert goes to its owner and the affected line can switch back to today's process while it is fixed.
What escalation looks like in practice
Published evidence on clinical voice agents is still thin, so one study is worth reading closely. At a cardiac catheterization laboratory, an LLM voice agent placed 806 pre-procedure calls to 701 patients between January and April 2025, giving instructions and asking a clinical questionnaire. It completed 86.4% of calls successfully, and 98 calls, 12.2%, were converted to a call with a person (European Heart Journal, Digital Health, 2026). It is one study, and only the abstract is public.
The lesson we take from it is about design: a well-run agent should pass a steady share of calls to people, and that share is something to measure. We report it, with the reasons, every week of the pilot.
Related: how we work, where agents run, revenue cycle and clinical audit.