The PLG Playbook for AI-Native B2B SaaS: Lessons from Our First Year

By Maxime Champoux

Seventy percent of B2B SaaS users who sign up for a free trial never come back after their first week. That number has held steady for years across categories, company sizes, and price points. Most founders treat it as a fact of life. We decided to treat it as a design problem.

When we started building an AI-native product, we assumed the standard PLG playbook would apply. Give users a free tier, let them discover features on their own, convert the ones who stick. The playbook broke on day one, and it broke in a way that changed how we think about product-led growth entirely.

The Cold-Start Problem at the Individual Level

Traditional SaaS shows value the moment you log in. A project management tool has empty boards, but you immediately see how they work. A CRM presents its interface, and the purpose is self-evident.

AI products have a different constraint. The product needs your data to generate useful output. But you have no reason to hand over your data until the product has proven itself useful. This is the cold-start problem applied to each individual user, and it destroys activation rates.

Our early sign-up flow asked users to connect data sources upfront. The drop-off was brutal. People saw the connection requests, questioned why they should trust a new tool with their CRM, calendar, and contacts, and left. We were demanding commitment before delivering a single insight.

Making Onboarding the Product

The fix was counterintuitive. Instead of simplifying onboarding, we made it longer. We designed 20 gamified activation steps, each earning users bonus AI conversations. Connect your calendar: 10 extra conversations. Import your contacts: 15 more. Complete all 20 steps and you have earned 140 bonus conversations on top of the base allocation.

Here is what this looks like in practice. A sales director signs up on Tuesday morning. She completes three steps in her first session: profile, calendar connection, and company details. She has earned 20 bonus conversations. She asks the AI about her upcoming meetings. Because it now has her calendar data, the response includes prep notes referencing her actual schedule, not a generic template. She connects LinkedIn the next day. Now the AI can cross-reference her network with her meeting attendees and flag warm introductions she did not know existed.

By step 8 or 9, users consistently hit the moment where the AI stops feeling generic and starts referencing their real business context. That is the activation moment. Not a feature click, not a tutorial completion, but the first time the product knows something about your work that you did not explicitly tell it.

The key insight: frame data connection as something the user earns, not something you extract. Traditional onboarding treats setup as friction to minimize. We treated it as a value exchange to maximize.

Ninety Seconds or Nothing

Even with gamified onboarding, we lost users who never reached step 2. The pattern was consistent: if someone did not experience genuine usefulness within the first 90 seconds of their first session, they were gone that day.

This forced a specific design decision. We started pre-loading context from publicly available information about the user's company, so even before they connected a single data source, the AI could produce something specific. A first-time user from a 50-person fintech would see their company's recent funding round, key hires, and competitive landscape referenced in the first response.

The 90-second constraint also killed tutorials. Walk-throughs, tooltips, and guided tours consumed those seconds without delivering output. We replaced every tutorial with an action that produced an immediate, tangible result.

Traditional PLG Metrics Will Mislead You

Here is where AI-native PLG diverges most sharply from conventional wisdom.

DAU is the wrong metric. AI product usage is naturally spiky. A user has an intense session on Monday, gets what she needs, and returns Thursday with a new question. In a DAU dashboard, that looks like declining engagement. In reality, it represents high value delivery. The product solved her problem so effectively she had no reason to return until a new one arose.

Feature adoption rates are meaningless because AI-native products are not feature bundles. The product is a single interaction surface whose quality depends on data depth, not feature breadth. Tracking "feature adoption" in an AI product is like tracking how many different questions someone asks a consultant. It misses the point.

We shifted to Weekly Active Users as our north star, targeting a move from 4% WAU to 30%. That reframing changed everything downstream. Instead of optimizing for daily check-ins through notifications and nudges, we optimized for weekly value delivery. Every week, the product needed to surface at least one insight the user could not get elsewhere.

The metrics framework we settled on after a year of iteration:

Time to First Value measured in seconds, not days. If yours exceeds two minutes, your activation rate will bleed.

Weekly Active Users over daily. WAU captures real engagement patterns. DAU creates false urgency and leads to notification spam.

Conversation Depth instead of session count. Three deep conversations per week signals stronger retention than daily app opens that never go past the home screen.

Data Connection Rate at Day 7 and Day 30. This is the leading indicator. Users who connect three or more sources by day 30 retain above 80%. Users who connect zero churn within 60 days, almost without exception.

Expansion Revenue per WAU rather than per seat. AI products often scale by becoming more embedded in a workflow, not by adding users. Usage-based expansion matters more than seat count.

A GTM Stack for PLG That Is Not Just "Wait and See"

PLG does not mean no sales effort. It means the product does the converting, but someone has to get the right people to the front door. Most PLG companies underbuild their go-to-market and hope organic growth fills the gap. It rarely does, especially pre-product-market fit.

Our stack runs on five tools with a combined cost lower than one junior SDR:

Theirstack finds companies matching our ideal customer profile based on their actual tech stack, not firmographic guesses. A company already using tools that integrate with ours converts at 4x the rate of a cold prospect.

Snitcher identifies which companies visit our website without signing up. Last quarter, Snitcher-identified visitors converted to trials at roughly 3x the rate of cold outbound. These are companies showing intent. The signal is too valuable to ignore.

Reactin runs LinkedIn engagement with prospects who have already visited our site. Not automated connection requests or pitch messages. Genuine engagement with their content so that when they evaluate AI tools, we are already familiar.

Cargo orchestrates data across all four tools above. Before Cargo, we had four separate data streams generating leads with no connection between them. Now, a visitor identified by Snitcher is automatically cross-referenced with Theirstack data, engaged via Reactin, and routed to the right outreach sequence.

Attio is our CRM. We chose it because it was designed for the workflows modern startups actually run, not the processes enterprise sales teams built 15 years ago. Clean API, real automations, no full-time admin required.

The total cost of this five-tool stack is roughly 15% of what we would spend on two SDRs. The qualified pipeline it generates is higher. That is the capital efficiency argument for building GTM infrastructure instead of hiring ahead of product-market fit.

What We Got Wrong

We over-indexed on virality too early. Traditional PLG prizes viral loops: invite your team, share a document, collaborate on a board. For an AI product where value is deeply personalized, virality is a second-order effect. Users need to find the product valuable individually before recommending it. We spent three months building sharing features that went unused because individual value was not strong enough yet.

We also underestimated the unit economics of a free tier when every interaction costs inference compute. Traditional SaaS has near-zero marginal cost per free user. AI products do not. Our free tier went through four redesigns before we found a balance generous enough to activate users but constrained enough to maintain viable unit economics. The 140-conversation bonus structure was the fourth version.

For founders building AI-native products: your PLG playbook needs to account for the cold-start problem, the inference cost structure, and the fact that your users will judge your entire product category based on their first 90 seconds with you. The metrics that worked for the last generation of SaaS will actively mislead you.

At Well, we are still adjusting. The 20 steps keep evolving, the WAU number keeps climbing, and the framework looks nothing like what we started with. That is probably the most honest thing I can say about PLG for AI-native products: the playbook is not a playbook at all. It is a series of bets you validate weekly.