← Back to Blog

Auto-Navigate Users Through Complex Workflows: The Basics

To move users through a complex workflow automatically, don’t try to automate the whole thing in one run. Break the workflow into feature-sized steps, have the assistant take the user to each step’s page and open the right control, and leave the data entry and decisions to the user. That’s also where Total Chat sits today: its walkthroughs cover one feature at a time, for logged-in users on the Starter plan and up. Chaining several features into one continuous workflow isn’t built yet, so a five-step process runs as five walkthroughs the user asks for in turn.

Last updated: October 2026

Complex-workflow guidance got overbuilt

Somewhere along the way, guiding users through a long process turned into a project of its own. Branching tour builders. Trigger rules for every step. Conditional paths for every role. A spreadsheet to track which tooltips broke after the last redesign.

No wonder teams give up on it. Tired of rebuilding the same tour every quarter? You’re not alone. The tour that took two sprints to build is out of date by the third, and the users it was meant to help clicked “Skip” on step two anyway.

If you’ve suspected most of that machinery was for the vendor’s demo rather than your users, you were mostly right. Strip it back and there are only three things an assistant has to do well to get someone through a hard workflow. Everything else is decoration.

The three fundamentals

1. Know where the user is. You can’t guide someone to step four if you don’t know they’re stuck on step two. Total Chat sends the current page with every message. With a feature map from its codebase scanner, it also knows any open modal, the active tab, and which features are available on that page. So “what do I do now?” gets an answer about the screen the user is actually on.

2. Get them to the next screen and open the right thing. This is the navigation part, and it’s where most of the time savings live. A Total Chat walkthrough moves the user to the feature’s page and waits for it to load. If the feature is a button or a dialog, it clicks the control that opens it; otherwise it points out the element. Then it explains the final step in chat. It survives the page reload, and if the user clicks or types while it’s running, it pauses and asks whether to continue.

3. Leave the data and the decisions with the user. Total Chat never fills in forms. I think that’s the right call for a support tool, and AI That Navigates Web Apps for Users: How Far It Goes makes the full case, so I won’t repeat it here.

That’s the whole model. Know where they are, move them forward, let them finish.

Turn one complex workflow into feature-sized steps

Here’s the practical part. Take your hardest workflow, the one that generates the most “how do I…” tickets, and write it down as a list of features. Below is a made-up example for a project-management app’s “set up a new client” flow.

Step	Feature in your app	What a walkthrough can do	What the user does
1	Create client	Go to Clients, open the “New client” dialog	Types the name and details
2	Connect an integration	Go to Integrations, point out the right card	Signs in to the other service
3	Set permissions	Go to the client’s Access tab	Picks roles
4	Invite the team	Open the “Invite” dialog	Enters email addresses
5	Create the first project	Open the “New project” dialog	Names it and picks a template

Five walkthroughs, each one short. The user asks for each one when they’re ready. That’s slower than an imaginary agent that does it all, but it’s how it actually works today, and in my experience it’s closer to how people work anyway: they stop between steps to check something, get a coffee, or ask a colleague which roles to pick.

Three rules make this table work.

Each step should be one feature with one way in. If “Invite” lives behind a menu, then a submenu, then a tab, no walkthrough will open it cleanly. Neither will your users. A step the assistant struggles with is usually a step people struggle with too.

Tag the controls that open each step. Total Chat’s docs recommend data-guide attributes on key elements, such as <button data-guide="add-team-member">Invite</button>, and the scanner suggests which ones to add first. Pointing at a control depends on the feature map having that control’s selector. If it isn’t captured, don’t count on it being marked.

Write the order down once. The assistant answers from your knowledge base and your app’s feature map. A single article called “Setting up a new client” that lists the five steps in order means a user who asks “what’s next?” after step two has something to get an answer from. Without it, every step is an island.

For a deeper look at the single-step mechanics, see AI-Powered Walkthroughs: How Chatbots Can Navigate Your App for Users.

What breaks complex workflows, and the fixes

Most failures aren’t model failures. They’re setup failures that look like model failures.

• The user isn’t recognized as logged in. Your server has to sign a token for the user. Without a valid one, the widget runs in visitor mode, which has no navigation or walkthroughs.
• The feature map is stale. A renamed button or moved page strands the walkthrough. Run the scanner (npx @totalchat/cli scan --upload) as a GitHub Actions step so the map keeps up with your code.
• The app isn’t Next.js. The scanner reads Next.js App Router and Pages Router apps only. On anything else, walkthroughs and page context beyond the route won’t have the map they need.
• A step depends on data from an earlier step. The assistant can open the dialog, but only the user knows the value that goes in it. Design for that; don’t fight it.

Setup details for all four are in the Total Chat docs.

Where a scripted flow still wins

Fair warning: this approach is worse for some workflows, and I’d rather say so than pretend otherwise.

If the process is mandatory and has to run start to finish in a fixed order (a compliance checklist, say, or a setup wizard that blocks access until it’s done), a scripted flow or a purpose-built wizard does that job better. A chat assistant waits to be asked. A wizard doesn’t. And if the hard part of your workflow is a twenty-field form, opening that form for the user saves a few seconds out of ten minutes.

The chat approach wins everywhere else: workflows people do occasionally, workflows they forget between uses, and workflows where the order depends on the customer. For how this plays out in first-week onboarding specifically, AI User Onboarding Walkthrough: What Actually Works covers it.

Navigation and walkthroughs are on Starter ($49 a month) and every plan above it, per the plans table in Total Chat’s docs. On the Free plan, the assistant gives directions in words instead of moving the user.

Frequently Asked Questions

Can an AI assistant guide users through a multi-step workflow automatically?

Partly. Total Chat runs walkthroughs one feature at a time: it moves a logged-in user to the right page, clicks the control that opens a button or dialog feature, and explains the final step in chat. Chaining several features into one continuous workflow isn’t built yet, so a longer process runs as separate walkthroughs that the user asks for one after another.

Does Total Chat fill in forms during a walkthrough?

No. Total Chat never fills in forms. It takes the user to the right page, opens the control for button or dialog features, can point out the element they need, and explains the final step in chat. The user enters the values and finishes the task, which keeps decisions about their own data in their hands.

What happens if the user clicks something mid-walkthrough?

The walkthrough pauses and asks whether to continue. It also survives the page reload that happens when it moves the user to a new page. That matters on busy screens, where users often click around before the assistant has finished its steps, so the walkthrough doesn’t fight them.

What does my app need for walkthroughs to work?

Three things. A feature map from Total Chat’s codebase scanner (npx @totalchat/cli scan --upload, Next.js apps only), a signed token from your server so logged-in users get the in-app mode, and the Starter plan or above. Adding data-guide attributes to key controls makes pointing at them more reliable.

---