How I Build & Deploy my first AI Healthcare Chatbot on Azure (2026 Guide)

Ever wondered what it actually takes to deploy an AI chatbot on Azure — not the sanitized tutorial version, but the real one? This build log covers every phase, every field value, and every error along the way, from Custom Question Answering to a live, working widget.

View the source code on Git Hub

Time to build

4-6 hours

Difficulty

Begineer-Intermediate

Cost

Free-tier eligible

Stack

Azure + Netlify

Most Azure chatbot tutorials show you the happy path: click here, click there, done. This one doesn’t. What follows is the complete, unedited build log of a real Healthcare Clinic Assistant chatbot — every phase, every exact field value, every screen it actually took to get from an empty Azure subscription to a live, deployed bot answering real questions on a real website.

It includes the parts most guides leave out: the region restriction that broke the first attempt, the checkbox that’s easy to miss and breaks everything downstream if you do, the exact wording used in the knowledge base, and a full table of test results — including the questions the bot got wrong — so you know what to expect rather than assuming a chatbot project goes perfectly on the first try.

By the end, you’ll have deployed the same stack: Azure AI Language (Custom Question Answering) for the knowledge base, Azure Bot Service and App Service for the backend, Direct Line as the secure channel, and a Web Chat widget hosted on Netlify as the public front end — all on Azure’s free tier.

If you’re a student, a career-switcher building a portfolio, or a developer who wants to ship something real (not just a proof-of-concept), this guide is written to get you from an empty Azure subscription to a deployed, working chatbot — using the same design decisions, the same services, and the same mistakes I ran into, so you don’t have to repeat them.

Architecture overview

Five services hand a message to each other in sequence. Understanding this chain before you open the Azure Portal makes every phase below click into place faster:

In plain language: the visitor types into a chat widget sitting on your website. That widget talks to Azure through a secure pipe called Direct Line. On the other end, Azure Bot Service receives the message and forwards it to Custom Question Answering, which searches the knowledge base you authored and sends back the best-matching answer. The reply travels back the same path to the widget. Every step below builds one link in that chain, in order.

What you'll need

A free Azure account — the services used here fit comfortably inside the free tier for a small project like this.
No prior bot-building experience required. You’ll write very little code — most of the “authoring” happens inside Azure’s visual tools.
A GitHub account and a free Netlify account, for hosting the front-end widget.
A text editor (VS Code is fine) for the small amount of HTML/JS you’ll touch in Step 8.
A clear idea of who your bot is for and what it should know — more on this in Step 1, because skipping this step is the #1 reason chatbot projects stall.

Phase 1 of 9

Azure Account & Subscription Setup

Before anything else, you need an active Azure account with appropriate permissions — and, if you’re on a student or restricted subscription, you need to know which regions you’re actually allowed to deploy into.

Sign in to the Azure Portal

Go to portal.azure.com and sign in with your Azure subscription account.

Check your allowed regions policy

This step is easy to skip and it’s the one that causes the most wasted time. An Azure for Students subscription carries a region restriction policy. Check it under:

Home → Subscriptions → Azure for Students → Policies

On this project, the allowed regions were: austriaeast, uaenorth, centralindia, eastasia, korea.

Critical

Australia East is not allowed on a restricted student subscription. Every resource in this build uses Central India — confirm your own allowed list before you start creating resources, or your first deployment will fail validation.

Phase 2 of 9

Resource Group & Language Resource

Create the resource group

Home → Create a resource → Resource group

FIELD	VALUE
Subscription	Azure for Students
Resource group name	HealthcareBot-RG
Region	Central India

Click Review + Create → Create.

Create the Language resource — with Custom QA enabled

Home → Create a resource → Marketplace → search "Language service" → Create

Critical step — don’t skip this

On the feature selection page, you must tick Custom Question Answering. This is mandatory. Without it, Custom QA will never appear in Language Studio later, and there is no way to turn it on after the fact — you’d have to delete the resource and start over (see Appendix A, this happened for real).

FIELD	VALUE
Subscription	Azure for Students
Resource group name	HealthcareBot-RG
Region	Central India
Name	HealthcareBot-Language
Pricing tier	Free F0 (5K transactions per 30 days)
Azure Search region	Central India
Azure Search pricing	Free F — if that errors out, use Basic
Identity type	SystemAssigned

Phase 3 of 9

Language Studio Setup & Project Creation

Open Language Studio and connect your resource

Go to language.cognitive.azure.com and sign in with the same Microsoft account used for the Azure Portal. On the “Select an Azure resource” popup, choose:

FIELD	VALUE
Azure subscription	Azure for Students
Resource type	Language
Resource name	HealthcareBot-Language
Pricing tier	Free F0
Managed identity	Enabled

Click Done.

Create the Custom QA project

Go to language.cognitive.azure.com and sign in with the same Microsoft account used for the Azure Portal. On the “Select an Azure resource” popup, choose:

PARAMETER	VALUE
Project name	HealthcareAssistant
Description	Custom QA bot for outpatient clinic patient inquiries
Default answer language	English
Default fallback answer	“I’m sorry, I can only assist with clinic-related queries. Please contact us at 1800-XXX-XXXX.”

Click Create project.

Phase 4 of 9

Building the Custom QA Knowledge Base

The knowledge base is the core of the bot. It stores every Q&A pair, every phrasing variation, the chit-chat layer, and the follow-up prompts.

Add the core Q&A pairs via file import

Edit Knowledge Base → Add source → Upload file

Uploading the clinic’s Q&A document auto-imports all 8 core pairs into the knowledge base in one pass:

[View the document]

Add alternate question phrasings

For each pair: click it → expand Alternate questions → Add alternate question. Aim for at least 4 variations per pair — this is what lets the bot recognize a real question, not just the one you typed. A sample of what was actually added:

Tag each pair with metadata

For each pair: expand Metadata → Add metadata → key: category, value as below. This is what lets you filter and analyze the knowledge base by topic later.

Q&A pair	Metadata key	Metadata value
Q1, Q6, Q7	category	general
Q2, Q8	category	appointments
Q3	category	insurance
Q4	category	results
Q5	category	prescriptions

Add follow-up prompts to the booking question

Click How do I book an appointment? → expand Follow up prompts → Add follow up prompt. Add three, one at a time:

Prompt text	Tab	Answer / link
Book Online	Create link to new pair	https://www.yourclinic.com/book-online/
Call Us	Create link to new pair	Call our helpline at 1800-XXX-XXXX
Q3	Create link to new pair	Our reception is open Mon–Fri 8 AM–5 PM.

Leave “Show in contextual flow only” unchecked for all three — this makes the prompts appear as clickable buttons any time this answer is shown, not only mid-conversation.

Add a professional chit-chat personality

Manage sources → Add sources → Add chit-chat

Select the Professional personality and click Add chit-chat. Azure automatically adds roughly 90 social Q&A pairs — greetings, thanks, “who are you,” and similar small talk — in the clinic’s chosen tone, with zero manual authoring required.

Phase 5 of 9

Training & Testing the Model

Save and train the knowledge base

Click the Save icon (💾) in the KB editor toolbar, then click Test (the chat bubble icon) in the top toolbar. The first time you open the Test panel, it triggers training automatically — this takes about 45–60 seconds.

Run test queries in the Test panel

In the Test panel (right side of the KB editor), type each query and press Enter:

TIP

Click the Inspect link after each response to see its confidence score. If a score comes in below target, add more alternate phrasings to that pair and retrain — don’t just accept it and move on.

Phase 6 of 9

Deploying the Knowledge Base to Production

Deploy the knowledge base

Left sidebar → Deploy knowledge base → Deploy → Confirm

Deployment takes roughly 60 seconds. Once done, you’ll see confirmation like this:

Get the prediction URL

Click Get prediction URL on the Deploy page and note down three values — you’ll need all of them for bot configuration in the next phase:

ITEMS	VALUE
Endpoint URL	https://HealthcareBot-Language.cognitiveservices.azure.com/
Project name	HealthcareAssistant
Deployment name	production

Also grab the API key: Azure Portal → HealthcareBot-Language → Keys and Endpoint → Key 1.

Phase 7 of 9

Azure Bot Service & Web App Deployment

Create the bot using the Language Studio shortcut

On the Deploy Knowledge Base page, click the blue Create a bot button. This opens an Azure custom deployment template with several fields already pre-populated.

Basics tab:

FIELD	VALUE
Subscription	Azure for Students
Resource group name	HealthcareBot-RG
Resource group location	(Asia Pacific) Central India
Bot handle	HealthcareBot-Language-bot
Pricing tier	Free
Creation type	Create new User-assigned managed identity

Configure the Web App tab

Click Get prediction URL on the Deploy page and note down three values — you’ll need all of them for bot configuration in the next phase:

FIELD	VALUE
App name	HealthcareBot-Language-bot-6e36 (auto-generated)
App URL	https://HealthcareBot-Language-bot-6e36.azurewebsites.net
SDK language	Node.js
App Service Plan	Create new app service plan
Language Resource Key	Paste Key 1 from Keys and Endpoint
Language project name	HealthcareAssistant (auto-populated)
Language service endpoint	(auto-populated)
Subscription ID	(auto-populated)

CRITICAL The Language Resource Key field shows a red error if left empty and won’t auto-populate. Get Key 1 from the Azure Portal and paste it in manually.

COST WARNING

The App Service Plan defaults to S1 Standard (roughly ₹5,000–6,000/month). After deployment, immediately scale it down: App Service Plan → Scale up → F1 Free → Apply.

Review + Create

Click Review + Create → Create. Deployment takes about 8 minutes and provisions five resources automatically:

Verify the bot in Test in Web Chat

Azure Portal → HealthcareBot-Language-bot(Azure Bot) → Test in Web Chat

Type: “What are your clinic hours?” Expected response: “We are open Monday to Friday from 8:00 AM to 6:00 PM, and Saturday from 9:00 AM to 1:00 PM. We are closed on Sundays and public holidays.”

✅ If this response comes back correctly, the bot is fully deployed and functional — the hardest part is done.

Phase 8 of 9

Direct Line Channel & Web Chat Widget

Enable the Direct Line channel

Azure Portal → HealthcareBot-Language-bot → Channels → Direct Line tab

You’ll see Default Site listed with Enabled: Yes, Authentication: Secrets, Version: 1.1, 3.0. Click Default Site to open the Configure the site panel. Under Secret keys, click the eye icon to reveal Key 1, then click the copy icon and save it somewhere safe immediately.

Key format looks like: A9N1yghpZnN6utWFqkH1PsERd9Y27... — a long alphanumeric string.

Build the clinic website with the widget embedded

Create a single index.html file containing:

A professional clinic homepage (navbar, hero, services, info, footer)
The Bot Framework Web Chat SDK, loaded from CDN
Token generation from the Direct Line API — a security best practice, rather than embedding the raw secret
A floating teal chat bubble in the bottom-right corner
The Web Chat widget styled with the clinic’s brand colors

The key code components:

<!-- Bot Framework Web Chat SDK -->
<script src='https://cdn.botframework.com/botframework-webchat/latest/webchat.js'></script>

// Token generation (security best practice)
const res = await fetch(
  'https://directline.botframework.com/v3/directline/tokens/generate',
  { method: 'POST',
    headers: { 'Authorization': 'Bearer ' + directLineSecret }
  }
);
const data = await res.json();
directLine = window.WebChat.createDirectLine({ token: data.token });

// Render the widget
window.WebChat.renderWebChat({
  directLine: directLine,
  userID: 'patient-' + Math.random().toString(36).substr(2, 9),
  username: 'Patient',
  locale: 'en-IN',
  botAvatarInitials: 'HC',
  styleOptions: {
    bubbleFromUserBackground: '#00838A',
    botAvatarBackgroundColor: '#1F4E79',
    suggestedActionBorderColor: '#00838A',
  }
}, document.getElementById('webchat'));

Phase 9 of 9

Hosting on Netlify (Public HTTPS)

The bot requires HTTPS to connect via Direct Line. Netlify provides free HTTPS hosting in under 60 seconds.

Why Netlify, specifically

Opening the HTML file directly (file://) or via localhost (http://) blocks the Direct Line connection because of the browser’s CORS security policy. Netlify’s free HTTPS is what makes the connection work at all.

Deploy to Netlify Drop

Rename the HTML file to index.html — Netlify requires this exact name for root URL access.
Go to app.netlify.com/drop.
Drag and drop index.html onto the drop zone.
Wait about 30 seconds — Netlify generates a public HTTPS URL, e.g. https://superb-salamander-70b9b9.netlify.app.
Open the URL, click the teal chat bubble — the bot connects and responds instantly.

NOTE

Double-clicking the HTML file locally also happens to work on Mac Safari, thanks to its less restrictive CORS settings. But a real Netlify URL is what you need for sharing with anyone else — or for a portfolio link.

Appendix A — Errors Encountered & Fixes

Nothing here went perfectly on the first try, and that’s normal. These are the exact errors hit during this build and how each was resolved:

Appendix B — All Azure Resources Created

Key URLs & endpoints

ITEMS	VALUE
Language Studio	https://language.cognitive.azure.com
Language endpoint	https://HealthcareBot-Language.cognitiveservices.azure.com/
Azure Web App URL	https://HealthcareBot-Language-bot-6e36.azurewebsites.net
Netlify public URL
Direct Line channel	Enabled — Default Site — version 1.1, 3.0
Project name	HealthcareAssistant
Deployment name	production

Appendix C — Test Results Summary

This is the part most tutorials never show you: what actually happens when you throw 48 varied, real-world queries at the finished bot.

What this actually tells you

Multi-intent queries (asking two things in one message, like “what are your hours and do you take Bupa?”) is where Custom Question Answering genuinely struggles — it’s built to match one question to one answer, not to decompose compound requests. That’s expected, not a bug in the setup, and it’s exactly the kind of gap a true NLU layer like CLU is built to close.

Live deployment spot-checks (on the Netlify URL)

Where to go from here

Close the multi-intent gap with a Conversational Language Understanding (CLU) layer trained on intents and entities, rather than pure Q&A matching.
Add Azure OpenAI as a fallback for genuinely open-ended questions instead of a flat “I don’t know.”
Log conversations to Cosmos DB to see what real users actually ask, and use it to keep growing the knowledge base.
Move the Direct Line secret server-side before using this pattern for anything beyond a personal or academic project.

Add comment Cancel reply

You must be logged in to post a comment.

Recommended for you Refresh