How I'm building Remi's medical board with AI

Remi needs a medical board. It's not decoration: I need the best doctors so people trust us and feel safe, and to make sure everything we're doing is medically sound.

To build it faster, I leaned on AI. An AI assistant called the doctors' offices to get their emails. I used AI to draft the invitations, reviewed them, and sent them.

The first six went out on November 10. In less than a day, two doctors had already agreed to a meeting.

Why Remi's medical board needs the best doctors

Remi is a chronic care membership in Mexico for people with diabetes, high blood pressure, or weight issues. It brings together doctors, medications, devices, lab tests, and at-home follow-up. I explained it in What is Remi.

A company like this can't be built on software alone. If the medical board isn't made up of people who know what they're doing, people won't trust it. And neither should I.

The best doctors matter for two things, and both matter equally. The first is trust: anyone who comes to Remi should feel that what's behind it is solid. The second is practice: what we design, decide, and publish should be medically sound, not improvised.

That's why the board comes before locking in the doctors who will see patients. It's the filter for everything we do.

From one lunch to 18 names, then to 52

On October 12 I had lunch with a friend who's a doctor. I asked him for help building the board: contacts for really good doctors in the specialties Remi needs, including endocrinology, nutrition, cardiology, and internal medicine.

On October 25 he sent me a handwritten list: 18 specialists. Nine for their track record and nine opinion leaders.

That settled the list. But I didn't stop there. I asked for research into who the best doctors for this board would be, beyond those 18. The result was 52 candidates.

I hadn't called anyone yet. I had names. What I was missing were emails, a message, and a decision on who to start with.

An AI assistant called the doctors' offices

On November 4 I made it clear: for the ones I already had, there was no need to ask for an email. For the rest, someone had to call their offices and get them.

I gave that job to an AI assistant. The task was narrow: call, ask for the contact email, and write it down. Before the real calls, it did a practice call with me.

This is the real call to one of those offices. I blanked out the doctor's name, the clinic, the email, and the name of the person who answered.

As you can hear, it wasn't perfect: it stumbled quite a bit. But it kept learning, and the following calls went much better. And even so, it got the job done: it got the email.

Transcript (details blanked out):

Office: [clinic], good afternoon.

Assistant: Good afternoon, is this Dr. [name]'s office?

Office: Yes, who's calling?

Assistant: Hi, I'm calling on behalf of Carlos Gotlib, from Remi. We'd like to send Dr. [name] an invitation by email. Could you give me her contact email?

Office: Mmm, it's [email].

Assistant: [email]. Is that correct?

Office: Uh-huh. That's right.

Assistant: Could you tell me your name, please?

Office: [name].

Assistant: Thank you so much, [name]. I really appre...

Office: No, for what?

Assistant: Sor...

Office: Hello, hello. Hello, hello.

Assistant: Sorry! I wanted to...

Office: It's cutting out.

Assistant: Oh, that's too bad! I'm so sorry, [name]. The email is [email]...

Office: Oh, okay.

Assistant: Thanks so m...

Office: Okay.

Assistant: Thank you, [name]! Have a great day. Bye...

Office: No, I said-- thanks.

Assistant: You're welcome! Have an excellent...

Some numbers didn't answer, others couldn't give out the email, and for some offices we had to find another line. I asked them to look for alternate numbers, try again, and not give up.

On November 10 we wrapped up with what we had. From the call list, five emails were confirmed by phone, five came from public sources without being confirmed on the call, and for two we couldn't get an email. We already had another nine from institutional addresses, and we didn't call those.

It's not a clean win. It's a list good enough to start writing. Six days passed between November 4, when the calls started, and November 10, when the invitations went out.

I wrote the emails with AI, then made them mine

I put together the first drafts with AI. They were long. They sounded robotic. Not very personal.

I cut them down. I asked for the close to be a short Zoom meeting to present the project. I rewrote one in my own words and that became the template: in the rest, only the greeting and the first line changed.

On November 10 I went through several versions of the text that same day, until I was happy with it. Then I sent it myself. The first wave went out: six invitations.

A second wave was ready the same day. Before sending it, I'm going to review it with the lawyers.

That's also part of doing it right: not confusing speed with sending everything at once.

Doctors are pretty accessible

That same afternoon, November 10, the first reply came in: happy to, a Zoom meeting starting in late November.

The next morning another reply came in: yes to a meeting, with proposed times. We agreed to meet on Zoom.

I expected it to be harder. Doctors are pretty accessible.

For those conversations I prepared a two-page document: what Remi is and what the board is. It doesn't replace a meeting; it helps the other person come in informed.

The board isn't signed. Nobody's been announced. What there is today are conversations: two doctors who agreed to meet with me, and the rest of that first wave still without a reply. That's enough to keep going.

What I learned doing it this way

I lean heavily on AI to be more efficient and move much faster.

The other thing I saw while rewriting the emails: if the text doesn't sound like you, don't send it. AI gave me a first draft in minutes; my job was to cut it down and put my name on it.

The assistant doesn't decide who to invite or what to say. I do that. What it does do is the work that eats up days: calling offices, getting an email, putting together a first draft. Without that, more weeks would have passed between the October 25 list and the first invitations.

What I don't delegate is judgment. I chose to start with a small group. I rewrote the email until it sounded like me. I sent it. I go to the meetings.

If you're building a board, an advisory board, or any group of people who are hard to reach, the lesson isn't "let a machine send it." It's this: use AI to take the slow stuff off your plate, and keep what only you can do.

What's next for me is simple. Follow up on the conversations already underway. Review the second wave with the lawyers and, when the time comes, send it. The board is set when the right people say yes, not when the first email goes out.