← Back to Blog

AI That Navigates Web Apps for Users: How Far It Goes

An AI assistant that moves users through a web app takes a request typed in chat (“where do I export invoices?”), opens the right page, and points to the control, instead of writing out directions for the user to follow. Tools differ a lot in how far they go: some only paste a link, some move the user and mark the button, and a few try to click through whole tasks or fill in forms. Total Chat sits deliberately in the middle. For logged-in users it opens the page, points out the control, clicks it when the feature is a button or dialog, and leaves the typing to the user.

Last updated: September 2026

The change: from directions to doing

For years, in-app help meant text. A help article said “Go to Settings, then Billing, then click Manage.” The user read it, went looking, and got lost when the menu had been renamed last sprint.

An assistant that acts inside the app removes that translation step. The user describes the goal, and the assistant does the moving. That’s a real shift in what “support” means, and it changes which questions need a person at all: a “where is X?” question can be answered by taking the user to X.

But “acts inside the app” covers a wide range. Before you evaluate any tool, it helps to know where on that range it sits.

Five levels of in-app navigation

1. Describe. The chat writes directions. Cheap, and wrong whenever the UI changes.
2. Link. The chat sends a deep link to the right page. Better, but the user still has to find the control once they arrive.
3. Move and mark. The assistant takes the user to the page and points out the exact element.
4. Open. It also clicks the button that opens the feature, so the user lands inside the dialog they asked about.
5. Complete. It fills in fields, submits forms and chains several tasks together on the user’s behalf.

Total Chat covers levels 3 and 4, for one feature at a time. It doesn’t do level 5.

Where Total Chat stops, and why I think that’s right

Here’s my opinion, and you’re free to disagree: in a SaaS product, the assistant should stop one step short of committing anything.

Filling in a form means guessing at the user’s data. Submitting it means acting in their account with consequences they may not have read. A wrong guess on an invoice amount or a permission setting costs far more than the ten seconds it saves. Walking someone to the exact field and letting them type is slower by a hair and safer by a mile.

So a Total Chat walkthrough goes like this:

• It moves the user to the page the feature lives on.
• It waits for that page to load.
• It points out the control they need (or, when the feature is a button or dialog, clicks the control that opens it).
• It explains the final step in chat, and the user does that part.

The walkthrough survives a page reload, so a full page change doesn’t lose the thread. We go deeper on how these work in AI-Powered Walkthroughs: How Chatbots Can Navigate Your App for Users.

The honest limitation is on the other side. Each walkthrough covers one feature. Chaining several into a single workflow (“create the project, add the client, then send the invite”) isn’t built. If your users’ hardest moments are long multi-step flows, the assistant can take them through each piece, but you’ll be asking for it one piece at a time.

How the assistant knows where to go

Moving someone to the right place requires knowing where things are. Total Chat gets that from two sources.

The first is a feature map built by scanning your app’s code, which records routes, features and, where it finds them, the UI selectors needed to point at a specific element. The second is live context. With every message a logged-in user sends, the assistant receives the current page, any open modal, the active tab, up to 10 features available on that page, and whether they’re on mobile or desktop.

It does not take screenshots, copy the full page, or read form contents or filter values. That’s a deliberate boundary, and a useful one to check on any competitor claiming to “see” the screen.

Now the caveat. Page context beyond the URL, walkthroughs and pointing at controls all depend on that feature map, and the codebase scanner that produces it isn’t available as a self-serve download yet. Without it, the assistant knows the route and answers from your knowledge base, but it can’t guide users to specific controls.

Until the scanner is published, there’s no way to build a feature map for your app yourself. Email [email protected] and ask before you plan around it. The scanner also reads Next.js apps only, so if your product runs on another framework, walkthroughs and control-pointing aren’t available for it today.

Marking a specific element also depends on the scan having captured that element’s selector. So “can point out the control” is accurate; “always points out every control” isn’t.

Logged in vs. logged out

All of the moving and pointing happens after login, inside your app. Total Chat’s pre-login widget, the one on your marketing site, is a different mode. It answers from a curated knowledge base, captures leads and can hand a conversation to a person by email. It has no page awareness and doesn’t move visitors around. If a vendor demo shows navigation on a marketing homepage, ask which mode you’re actually looking at. (For the logged-in side more broadly, see Context-Aware Chatbot for Web Apps: What Works in 2026.)

Questions to ask any vendor

Use these whether you’re looking at Total Chat or anything else in this space:

• What does it read from the page on each message? Screenshots, the full page, or structured context?
• Does it click anything, and which kinds of controls?
• Does it ever type into fields or submit forms?
• Can it chain steps across several features?
• How does it learn where your features live, and what happens when you ship a redesign?
• Does it work before login, after login, or both?
• When it can’t help, what happens next?

On that last one: Total Chat can hand an unresolved conversation to a person when you’ve set an escalation email, sending them the full transcript. And when a user says something is broken, it files a structured bug report with the page address, browser and recent JavaScript errors. For a wider view of the support stack around this, read AI Help Desk for Web Apps: SaaS Founder’s Guide.

What this replaces, and what it doesn’t

It’s aimed at the “where is it?” questions. It doesn’t replace a knowledge base; the assistant still answers the “why” and “should I” questions from articles you write or approve. The URL crawler can draft 3 to 8 of those from your live site to start with. And it doesn’t replace pre-scripted tours for the one step every user must take, though Intercom’s Product Tours, for comparison, are pre-scripted and sit on its $85/seat Advanced plan.

Install is a script tag with streaming responses, no SDK. Setup notes are in the docs.

Frequently Asked Questions

Can Total Chat click buttons for my users?

Yes, within limits. During a walkthrough, when the feature a user asked about is a button or a dialog, it clicks the control that opens it. It doesn’t fill in form fields, and the final step is explained in chat for the user to do themselves. This only happens for logged-in users inside your app.

Can it walk a user through a multi-step workflow?

One feature at a time. A walkthrough takes the user to the right page, waits for it to load, points out the control and explains the last step, and it survives a page reload. Chaining several features into one continuous workflow isn’t built, so a longer process means asking for each piece in turn.

How does the assistant know where features are in my app?

From a feature map built by scanning your app’s code, plus live context sent with every message: the current page, any open modal, the active tab and the features available there. The codebase scanner that builds the map isn’t a self-serve download yet, and without the map the assistant knows the route but can’t point to specific controls.

Does it work on my marketing site?

Not for moving visitors around. Total Chat’s pre-login widget answers questions from your knowledge base, captures leads, and can hand a conversation to a person by email. Page awareness, walkthroughs and pointing out controls only run for logged-in users inside your app, where the assistant has the page context it needs.

---