AI scheduling assistants

Q4 rock · AI voice agent · Prepared 10/29/26

What we're asking for

Approval to build and test an AI scheduling assistant for each of our Medicare agents, piloted in Florida first. Approval covers test leads only. No real lead talks to the AI until Sergio answers the compliance questions in writing.

  • Justin: approve the build, the ten compliance rules and the consent wording.
  • Sergio: review the consent wording, the call script and the Q&A, and answer the questions in the Sergio tab.

Medicare leads, Aug 1 to Oct 28, 2026

The leads arrive. The conversations don't.

Every alert reaches the agent within a minute. Most leads still never get a real conversation.

Qualified leads31from our Facebook formsTracked
Alerts on time31every one reached the agent within 60 secondsTracked
Reached513 never worked, 14 marked no answerTracked
Met with the agent03 booked, 2 of those canceledTracked
  • The alert isn’t the problem. The gap is between the alert and a real conversation.
  • One agent can’t call every lead within minutes while also running appointments. An assistant can.

One job

An assistant that books a time with the agent

Each agent gets their own AI scheduling assistant. It says it's the agent's AI assistant, never the agent, and it only sets up a time for the agent to call.

  • It answers right away. A lead taps “call now” on the thank-you page and the assistant picks up, day or night.
  • It calls back fast, once consent allows. Within minutes of the form, for leads whose form carries the new consent wording.
  • It books on the agent's own calendar. Through the booking engine we already run. Our confirmation and follow-up emails take it from there.
  • It's a layer, not a new system. Agents get the same lead alert and lead sheet they use today, with a short call summary added.

Hard limits

What it never does

These are built in, tested on every test call, and checked on every pilot call.

  • No plan talk. It never discusses plans, prices, benefits, drugs or doctors, and never enrolls anyone.
  • No pretending. It never claims to be human, licensed, Medicare or the government.
  • No sensitive numbers. It never asks for a Medicare or Social Security number, bank details or health conditions.
  • No pressure. It never pushes past a no, never calls outside allowed hours, and stops everything the moment someone says stop.

For Justin

Five decisions, with our recommendation

Each one is recommended and agreed by Nick. Approving the proposal approves these.

  • 1. Scope of Appointment: the agent takes it, on the appointment. The assistant only confirms the form's “What would you like to go over first?” We're open to discussing the assistant taking the SOA instead.
  • 2. Outbound calls: inbound first. Outbound starts once Sergio approves the new consent wording and it's on new forms.
  • 3. Recordings: in our own storage. Recording starts only after the lead says yes. Each recording is saved to HPM's own storage and linked to the lead. See How we handle recordings.
  • 4. Phone numbers: one local number per agent's assistant. A local number can show a caller name: it reads “Ryan's Assistant” (each agent's own name). Toll-free numbers can't show a name. The 833 keeps sending texts.
  • 5. Vendor: Telnyx. Our numbers and texting already run there. Their security report and a signed health-data agreement (BAA) come before we build.

The flow

How a call works

About a minute, start to finish. The full wording is in the Call script tab.

  • Opens honestly. Names itself as the agent's AI assistant with Health Plan Markets, offers “say stop at any time,” and asks before recording.
  • Offers real times. Two or three open times from the agent's own Google Calendar, inside the agent's booking hours.
  • Reads it back. Day, date, time and time zone, then the agent's own number to save, and the agent's contact card by email.
  • Hands off to the right agent. A plan question means a warm transfer to the assigned agent, who hears a short summary first. If the agent can't pick up, a booked time. Never a different agent.

The agent's own calls still come from the agent's own number, so the promise on each contact card holds: Ryan’s card says he’ll only call from (407) 718-5060, and the assistant isn’t Ryan.

Recordings and transcripts

How we handle recordings

Only with the lead's yes, kept in our own storage, seen only by the people who need it.

  • Only after a yes. The assistant asks first. Recording starts only after the lead agrees. If they say no, nothing is recorded and the agent calls them instead.
  • Saved where we control it. Telnyx can write each recording straight to a storage bucket we own. We'd use the Google Cloud project our lead relay already runs on, so there's no new vendor. Telnyx charges $0.002 a minute to record, and storage at Telnyx is free.
  • Linked to the lead. The relay files each recording and its transcript against the lead. The agent sees a short call summary in the alert, with a link to the recording.
  • Private by default. Only the assigned agent and admins can open a recording. Transcripts have personal details such as numbers blanked out automatically.
  • Kept 6 years, then deleted. That's how long Medicare sales and marketing calls must be kept. Sergio confirms whether that applies to scheduling calls (question 4).
  • HQ gets the outcome. HQ receives the booking and the call result, not the audio, unless Mike wants it. We agree that with Mike before the build.

Before any real lead

Four small builds on the lead relay

Each gets its own spec and review before the prototype.

  • One stop list for every channel. Today only email has one. A stop by voice, text or email must stop everything, and HQ hears about it.
  • An “AI spoke with them” mark. So the assistant, the agent's call and the emails don't pile onto the same person.
  • AI calls in the speed report. The report times first contact from the alert's call button today. AI calls need their own time stamp.
  • Booking without the SOA. Our booking page still records a Scope of Appointment. The assistant's bookings skip that step.

Florida first, then expansion

Plan and cost

We pilot in Florida, and expand to our other states once it's proven there. These are target dates; we move sooner wherever we're ready.

  • 10/30, approved. Justin approves; questions go to Sergio; Telnyx paperwork requested; records home agreed with Mike.
  • By 11/20, prototype passing test calls. A test number, test leads, staff phones only. At least 40 scripted calls covering every hard case.
  • By 12/11, Florida pilot with one agent. The agent is picked later. Inbound first. Nick reads every transcript.
  • By 1/7/27, launch. Agents added one at a time, each after Nick tests every step as the lead. Other states follow once their rules are confirmed.
  • What the pilot proves. That the assistant works end to end, every time, and stays compliant on every call.
  • Cost. Telnyx’s pricing is about 6 cents a minute for the AI voice, $0.002 a minute to record, and $1 a month per number.

Feedback on this page

Tell us what works and what to change.
It comes straight to Nick.

Optional. Add it and we'll write back when your note changes something.

What’s off? Paste it, link it, or add a file. Rough is fine.

Drag files here, or click to choose.Any type, up to 2 GB each, 10 per note.

Go-live: 7/17/26 · Last updated: 9/28/26