How I Used AI to Build an AI Chatbot That Helps Medical Interpreter-Coordinators

I built an AI chatbot for MyClinic that supports medical interpreter-coordinators during consultations with international patients. Its role is to help a medical interpreter-coordinator, during a consultation with an international patient, find clinic information, prices, and procedure descriptions, and answer more accurately. International patients don’t talk to the AI chatbot; a person does the consulting, with help from the AI chatbot. I named the AI chatbot Chiron. I’m told it is the name of the teacher who taught countless heroes in Greek mythology.

International patients need accurate information.

International patients ask whether the price is really this amount, how long recovery takes, and whether they can have the procedure given their condition. When they ask, we check everything with the clinic, from the price to whether the procedure is possible given the client’s condition. If we give an outdated price or an inaccurate recovery period, it can also disrupt the client’s visit plans. So rather than having the AI chatbot talk directly with clients, I made it something coordinators use to prepare their answers. It puts a person in place to check the information before it reaches the client.

But I had to cut the time it takes to find answers.

When I calculated how much time I spent communicating with one client, from the first inquiry until their treatment finished, it came to nearly 10 hours. I also kept answering similar questions about procedure costs and recovery periods, clinic locations and opening hours, and the languages available for consultation. A less experienced coordinator has to look things up or ask someone else to answer questions like these. If a price is quoted from memory, or several procedures are done at once, the amount can differ from the one on the site, so the checking repeats. I wanted to cut that time with an AI chatbot.

The AI chatbot can answer some things, but others must be checked with the medical team.

Even after building Chiron, I test it repeatedly. It’s an internal test: I enter questions a coordinator might ask during a consultation and review the answers.

Recovery period. When I asked, “A client is asking how many days they need to rest after double eyelid surgery,” Chiron searched the dataset of procedure information and answered. It explained how long it takes to return to daily life and to go home, described the difference in swelling between the buried-suture and incisional methods, and cited its source. It explained that this period does not mean full recovery.

Medical judgment. For the question “A client who takes blood pressure medication is asking whether they can get ultrasound lifting,” Chiron did not judge whether it was possible. Because the medication and health condition need review, it advised checking with the medical team in charge. It also provided a sentence for asking the client to tell the medical team which medications they are taking. This is something an AI chatbot should not answer, so Chiron’s answer was the right one.

Prices by clinic. When I asked, “Tell me which clinics do Botox and their prices,” it showed only prices that the clinics had confirmed. The amounts were regular prices for a single session, including VAT. It also added a note that each clinic has different standards for product, dose, and area, so the prices may be hard to compare as they are. Clinics that had not confirmed yet were marked as not yet confirmed, with no amount.

Message wording. When I said, “Please write a LASIK consultation booking confirmation message in English,” it wrote an English message. It left blanks for the parts a medical interpreter-coordinator needs to fill in, such as the clinic name and address.

Four things I needed to build the AI chatbot Chiron

1. Verified data

I gathered clinic information, procedure prices, and procedure descriptions, and had the clinics confirm them again. This took the longest. I set Chiron up to query the curated dataset and the procedure information database, and to show prices only for clinics that confirmed them. Keeping the data up to date takes an enormous amount of work.

2. An AI model that can use tools

Chiron uses the Google Gemini API. During development, I compared several models, including ChatGPT and Claude, by asking them the same questions, and then chose one. Based on the quality of its answers, I’m also considering switching later.

3. Tools that read data

I connected 11 of MyClinic’s internal functions: quote calculation, clinic search and clinic details, price search, two Procedure Wiki functions, inquiries and work status, coordinator search and summaries, and goal progress lookup. All of them can only read data, not write it. I did not include any function that changes bookings or sends messages to clients.

4. Tests checked against correct answers

I set up a separate harness that tests it with the actual instructions and questions.

If developing it yourself is hard, you could also upload your price list and procedure description files to a general-purpose AI service and test the answers with questions whose correct answers you already know. That said, I’m not sure improvement can be sustained this way. If you use files, you have to update them whenever prices change, and you also have to check the number of clinics and the price comparison results against the source data.

Managing the data is the hardest part. Confirming clinic information and keeping it up to date takes constant work.