Colonoscopy Prep: a voice patient-guidance agent built with Vocal Bridge

A voice agent that carries a multi-day medical protocol so the patient doesn't have to, with a visual checklist that stays in step.

The problem: the instructions arrive at the worst possible moment

Prep for a colonoscopy is not one instruction. It is a sequence that runs over days, with timing that matters, and it lands as a printed sheet the patient reads once in a clinic and again at home when they are already anxious and hungry.

That is a bad moment to be parsing a document. The questions that come up are small and urgent. Can I still have this. Was I supposed to start yesterday. None of them are answered by scrolling back through a PDF with one hand.

And the failure is expensive on both sides. A missed step means the procedure gets cancelled on the day, so the patient repeats the whole prep and the clinic loses the slot.

What they built

The team built a guidance agent that owns the protocol instead of handing it over. It speaks the current step when that step is due, takes questions in plain language, and confirms what the patient has actually done rather than assuming they read ahead.

Running alongside it is a checklist on screen. The patient can see where they are in the sequence, but never has to drive it. Confirming a step out loud moves it, and tapping it moves the conversation.

What they accomplished is the harder half of patient adherence. Not better instructions, but the right one arriving at the right time and getting acknowledged.

How Colonoscopy Prep uses Vocal Bridge

The build sits on Vocal Bridge's Voice for your app surface, the pattern where a voice agent and a UI talk to each other over one real-time channel. Client actions carry both directions: the agent marks a step complete on screen when the patient confirms it out loud, and the app tells the agent when the patient taps something so the conversation follows.

import { useAgentActions } from "@vocalbridgeai/react";

const { onAction, sendAction } = useAgentActions();

// agent -> app: patient confirmed a step out loud, tick it on the checklist
onAction("complete_step", ({ stepId }) => {
  setChecklist((prev) => ({ ...prev, [stepId]: "done" }));
});

// app -> agent: patient tapped a step, so the agent picks up from there
const handleTap = (stepId) => sendAction("jump_to_step", { stepId });

Illustrative. The shipped build tracks a longer prep sequence and its timing rules.

The split is clean. The team owns the prep protocol, the step timings, and the checklist UI. Vocal Bridge owns the voice channel, streaming, turn-taking, barge-in, and the data channel that keeps speech and screen in agreement.

What's hard about voice here

Two things make this harder than a talking checklist, both handled at the platform layer.

• The patient interrupts, because the questions are urgent. Someone mid-prep does not wait politely for the agent to finish a paragraph before asking whether tea counts. Vocal Bridge's barge-in lets them cut in and get an answer, which is the difference between a guide and a recording.
• Speech and screen have to agree, or the patient trusts neither. If the agent says step three is done and the checklist still shows step two, the whole thing reads as broken at exactly the moment the patient needs to rely on it. Bidirectional client actions on one channel keep both views of the same state in sync, rather than reconciling two sources after the fact.

That gap, between a demo that works when you follow the script and something a patient would trust the night before a procedure, is where the real work sits.