AI Patient Intake Agent for Teleradiology

AI Intake Agent turns website visitors of a remote diagnostic-imaging service into structured, manager-ready requests around the clock.

83% of dialogs end with a request and summary

50% of conversations happen outside work hours

21% click-through to the order page, up from 10%

Partner's Profile

Background

The client is a remote diagnostic-imaging (teleradiology) service for patients. Its website sent visitors straight to an order form. Patients did not understand the difference between CT, MRI and X-ray reading or a second opinion, did not trust a remote service, and left. There was no way to capture and warm up leads, and many questions arrived at night and on weekends, when nobody was there to answer.

Objective

Put a consultant between the visitor and the order form: an AI agent that explains the services in the patient's language, answers from a knowledge base, collects a structured request and passes it to a manager for confirmation. Medical safety had to be enforced by code rather than left to the model, and the agent had to work on the website first and then in messaging channels.

Solution

We built the agent as a custom agentic application on our reusable multi-tenant intake engine: the same core runs 8 configurable verticals, from cargo, aviation and immigration to insurance, travel and now radiology. The agent detects whether the patient writes in Ukrainian, Russian or English, consults from a knowledge base, scores the lead from 0 to 100, gives an order link with UTM tags and sends the request with a summary to a manager through a Telegram alert. The web widget installs with one line of code; Telegram (including voice messages) and email channels are implemented.

The hard part was making an LLM dependable in a medical context. Our principle: the model runs the conversation, code makes the decisions. Validation, statuses and submission are deterministic. The model can change request data only through tool calls, and every value stores its source. The model estimates risk from the patient's words, but the URGENT flag is set only by a deterministic rule, and such cases always go to a human.

A single Pydantic schema is the source of truth: the tool schema, missing fields, progress bar and UI metadata are all derived from it, so the backend, the model and the interface stay in sync. That is why a consultation-only mode was added purely by configuration, without touching the core. The LLM layer is provider-agnostic, with Anthropic Claude and OpenAI GPT models switchable. The admin panel gives the team:

an inbox with operator takeover of a live dialog
prompt versioning and a full LLM call log
a knowledge base that grows from managers' answers to escalated questions

We tuned the agent on 50 reference dialogs and tested it like any other production software: a red-team run on a test stand handled 23 of 23 medical-restriction provocations without a safety violation, the client's QA passed 15 of 15 scenarios, and the codebase carries about 2,700 automated tests. This is how we approach AI software development: the model where language matters, code where correctness matters.

Consultation in the Patient's Language

The agent auto-detects Ukrainian, Russian or English and explains the difference between CT, MRI and X-ray reading and a second opinion. It answers from a knowledge base instead of sending an unsure visitor straight to an order form.

Structured Request and Lead Score

While the conversation runs, the agent fills a structured request with a progress bar, scores the lead from 0 to 100 and gives an order link with UTM tags. The manager receives a summary and a Telegram alert to confirm the request.

The Model Talks, Code Decides

The model changes request data only through tool calls, and each value keeps its source. Validation, statuses and submission are deterministic. The URGENT flag is set only by a rule in code, and such cases always go to a human.

Admin Panel with Operator Takeover

An inbox of all dialogs, takeover of a live conversation by an operator, prompt versioning and a full LLM call log. Managers' answers to escalated questions are added to the knowledge base, so the knowledge base grows with real patient questions.