Stop Building Tooltip Tours: Why We Built an AI Agent for User Guidance

The 3-Hour Weekly Tax

Every Tuesday, I'd open Appcues and fix broken flows.

A button moved? Tour step fails silently. A new feature shipped? New walkthrough needed. A tab renamed? Three flows reference the old name. I spent 3 hours per week — 150 hours per year — maintaining tooltip tours for a 4-page SaaS app.

I looked at the market: Pendo ($7K+/year), WalkMe (enterprise pricing), Chameleon, UserGuiding. They all had the same architecture: you author flows manually, and they play them back. The AI revolution happened everywhere in software except here.

So I asked: what if the guide was the AI?

The Architecture Is the Problem

Every digital adoption platform built before 2023 has the same DNA:

Manual flow authoring → Rule-based triggers → Linear tooltip sequences

You build a tour in a visual editor. You set trigger rules (page = /dashboard, user.role = new). The tool plays back your tooltip sequence. When something changes, you fix it.

This worked when products had 5 pages and shipped quarterly. It doesn't work when you deploy daily, have 50+ interactive elements per page, and users ask questions the flow builder didn't anticipate.

The fundamental problem: you're predicting what users will need. And you're always wrong about some of it.

What If the Guide Observed Instead of Followed Scripts?

ActivationKit inverts the model:

DOM observation → AI agent → Structured action execution

Instead of you describing your product to the tool, the tool reads your product. Instead of you authoring flows, the AI generates guidance. Instead of linear tooltip sequences, the AI picks from a palette of actions: highlight an element, navigate to a tab, start a tour, show a nudge, suggest an upgrade.

The design principle is simple: observe the DOM, never require host cooperation.

The AI reads disabled buttons, aria-disabled attributes, locked CSS classes, hidden elements, tab structures — everything your app already communicates to its users. Your app doesn't need to change. We read what's already there.

The Three-Tier Pipeline

Not everything needs an LLM. In fact, most user guidance decisions are predictable.

Tier 1 — Rules (<10ms): 12 deterministic rules handle common scenarios. First visit? Welcome nudge. Idle for 30 seconds? Contextual help. Form abandoned? Encouragement. All tours completed? Celebration. These fire instantly, no LLM.

Tier 2 — Cache (<50ms): When the same question has been asked in a similar context before, return the cached response. Quality scoring ensures bad responses don't persist.

Tier 3 — LLM (500-2000ms): GPT-4o-mini with structured JSON output. Only called when rules don't match and cache misses. The response is cached for next time.

~70% of requests resolve at Tier 1 or 2. The LLM is powerful, but most guidance doesn't need it.

The Feature That Changes Everything

The moment that convinced me this was worth building as a product:

A user typed: "show me how to set up wallet passes."

There was no pre-built wallet tour. I hadn't authored one. The LLM looked at the page elements — Wallet tab, Apple Wallet button, Google Wallet button, Preview Pass, Business Card — and generated a 5-step guided tour on the fly.

The user clicked through it. It worked. I hadn't configured anything for that specific request.

This is the gap in the market: on-demand tour generation. No DAP does this. Pendo can't. Appcues can't. They need you to build every tour in advance. ActivationKit generates them from your actual page structure.

Why Now

Three things aligned:

1. The acquisition vacuum. CommandBar was the only AI-native competitor. Amplitude acquired them in October 2024 for $45M+. Their roadmap is now subordinate to Amplitude's product analytics vision. WalkMe was acquired by SAP for $1.5B. The independent AI-native DAP category is empty.

2. LLM costs collapsed. GPT-4o-mini costs ~$0.001 per conversation. A free-tier user generating 25 conversations per month costs us $0.025. AI-first guidance is economically viable at every pricing tier.

3. AI expectations shifted. In 2024, "AI-powered" was a differentiator. In 2026, it's table stakes. Users expect AI features. But "AI-powered user guidance" — where the AI decides what to show, not just assists in authoring — is still novel.

What We Learned from Dogfooding

ActivationKit runs on a real SaaS product. Real users. Real DOM. Real questions. Here's what we learned:

DOM observation works. The SDK correctly detects disabled states, tier-gated features, hidden elements, and tab structures across React, MUI, and custom components. No data- attributes needed.

Tours need workflow milestones, not form fields. Early auto-generated tours walked users through every input field. Users don't need that. They need: "here's the section, here's the key action, here's how to save." We tuned the prompts to generate milestone-based tours.

70% rule-based is real. Welcome nudges, return visit prompts, idle help, form encouragement — these cover the majority of guidance scenarios. The LLM handles the long tail.

Element resolution by name works better than CSS selectors. The LLM outputs element names ("Apple Wallet", "Background Color"), and the SDK resolves them via text search. More robust than CSS selectors that break on deploy.

Try It

One script tag. Free forever. 2-minute setup.

ActivationKit reads your product so you don't have to describe it. It guides your users so you don't have to author flows. It adapts when your UI changes so you don't have to fix broken tours.

Stop building tooltip tours.