Halcyon Health keeps a human in every refund

A support team with rules it cannot bend

Halcyon Health runs patient billing support for 140 clinics across three states. Every refund it issues is subject to audit, and the rules are strict: a person must approve any money that leaves the business, and the reason must be on record. When the team started looking at AI agents last spring, that rule was not up for discussion.

“We were never going to let software refund patients on its own,” says Ana Ruiz, who leads support operations. “The question was whether an agent could do everything else, so our people only spend time on the decision.”

The setup

Halcyon’s agent handles inbound billing tickets from start to finish, except for one step. It reads the ticket, pulls the patient’s balance and payment history from Payly, checks the clinic’s refund policy and drafts a resolution. If the resolution includes a refund, it stops and asks.

• Approval card. The reviewer sees the amount, the policy clause the agent relied on, the payment it would reverse and a two-line reason.
• One click, or a note. Approve, reject, or send it back with a comment the agent reads before trying again.
• Hard limits outside the prompt. Refunds over $500 need a second approver. The agent cannot change that, however a ticket is worded.

The agent does the reading. Our people do the deciding. That split is why compliance signed off in two weeks instead of two quarters.

Fast, without acting alone

The worry with approvals is that they turn an agent back into a queue. Halcyon measured this carefully during a six-week canary on 10% of tickets.

1. Median time from ticket to resolution fell from 26 hours to 3 hours 40 minutes.
2. Median time a refund waited for approval was 11 minutes during business hours.
3. Reviewers approved 93% of proposed refunds unchanged, edited 5% and rejected 2%.
4. Zero refunds were issued without a named approver, which matched the audit requirement exactly.

The 2% of rejections turned out to be the most useful part. Most traced back to two clinics with older, informal refund rules that were never written down. Halcyon wrote them down, added them to the policy source and the rejection rate on those clinics dropped to zero.

What they would tell another regulated team

Ruiz is careful not to oversell it. The agent is good at gathering facts and following written policy. It is not good at policies that live in someone’s head, and it was never meant to be.

• Put approval rules in the platform, not the prompt.
• Show reviewers the evidence, not just the answer.
• Treat every rejection as a gap in your policy or your tools.

Halcyon now runs the agent on all billing tickets. If you are planning something similar, our guide to guardrails your agents can’t talk their way around covers the approval setup in detail.