Product-Led Growth becomes AI-Led Growth.

No, it sits on top of it. In a product-led motion the product carries the sales work: people sign up, find value and upgrade largely on their own, and that route stays intact. A predictive layer changes how you steer inside it. The same behavioral signals are used to estimate the next useful step for a given account, and to trigger it without waiting for someone to notice. Product-led growth describes the route revenue travels; AI-led growth decides where you intervene on that route and where you leave the product alone.

Follow the decision the score feeds. Cancellation and expansion are account-level decisions, so predict at account level. Activation is a per-user question: inside the same workspace, a colleague may use the product daily while another never signs in again. An account that looks healthy on average can rest on a single active seat, which is exactly the risk an account average hides. Model behavior where the behavior happens, at user level, then roll it up to the account your commercial action targets.

Then the model learns the tracking change rather than customer behavior. If an event such as 'project created' was recorded differently last year, your history holds a break that has nothing to do with your users. Two habits help. Keep a record of when each key event was redefined, so that period can be excluded or handled separately. And start with signals that have been measured the same way for a long time, even when they are coarse. For this kind of model, consistency counts for more than volume.

Because accuracy answers a different question than impact. Accuracy is the share of cases classified correctly, and only a small fraction of subscriptions cancel in any given period. A model that predicts 'stays' for every account therefore scores well while never pointing at a customer at risk. Judge ranking quality instead, with measures such as AUC or lift, and keep that separate from the business result. Withhold the action from part of the eligible accounts, then compare the two groups on the metric you set out to move, such as activation or retention.

Yes, but on different signals than for established customers. A trial is short and each account carries little history, so the model leans on early behavior: how many colleagues were invited, whether the first real task was completed, whether anyone came back soon after. Do not expect a fine-grained ranking from that. Pick the moment where intervening genuinely changes the outcome, usually the step to that first completed task, and build that intervention properly. Messaging users on every deviation makes each later message count for less.