AI-Led Growth: The Motion After Product-Led

Product-led growth is having an identity crisis in public. Half the industry says PLG is dead because the next user of your software will be an agent. The other half says nothing fundamental changed. Both sides are arguing about the wrong layer. The economics of PLG survive. The assumptions of its funnel do not.

What was PLG actually solving?

Strip the decade of playbooks away and PLG was a distribution answer: move acquisition, education, and qualification into the product itself, so revenue stops scaling with headcount. That logic is untouched by agents. If anything, it compounds, because the one motion that never needed a human on your side of the table is the one that keeps working when there is no human on the buyer's side either.

What breaks is everything PLG assumed about how value gets discovered: a person signs up, tours the interface, learns the tool, and converts when the learning pays off. Every step of that sentence is now optional.

What are the 3 layers of AI-led growth?

Intelligence inside the experience. The first layer is the product doing growth work per account, in real time, from the account's own data. An in-product agent that reads what this specific user has and has not captured, quantifies the value they are leaving unclaimed, and recommends the single next step. Onboarding stops being a flow everyone gets and becomes a conversation with your own numbers. This is the highest-leverage version because it converts with evidence, and the evidence is personal. It also carries the trust burden I described in The AI Trust Gap, since a recommendation the user cannot verify is worth less than no recommendation at all.

Agents running the growth loop. The second layer is operational: agents executing the experiment cycle, the segmentation, the lifecycle sends, while humans keep problem selection and metric definitions. I wrote about this in The Agentic Growth Team. The short version: the capacity constraint on growth execution is gone, and judgment is the new bottleneck.

Surfaces built for non-human evaluators. The third layer is distribution. Agents now discover tools, authenticate, and complete tasks through protocol registries and machine-readable documentation. A product that only exposes its value through an interface designed for human eyes is invisible to that layer, and the invisibility is silent: no bounce shows up in analytics, because the agent never arrived. Time-to-first-value for this evaluator is measured in API calls, not sessions, which is the funnel I mapped in When Your Next User Is an Agent.

What dies

The onboarding wizard, first. It was already a tax on humans, and an agent will not sit through it at all.

MAU-shaped success metrics, second. When 1 supervisor directs 40 agents, seats and active users stop describing the account. Work completed does. This is the same shift that broke per-seat pricing, and it breaks per-seat measurement for the same reason, which I worked through in Pricing After Seats.

The generous, indefinite free tier, third. Every free interaction now carries inference cost, so free stops being a marketing expense rounding to zero and becomes a real unit-economics decision, gated by value rather than by time alone. That math is the subject of The Free Tier Meets the Inference Bill.

And quietly, the assumption that a signup is a person. Activation logic, fraud logic, and qualification scoring all inherit that assumption from a decade of human-only funnels. All 3 need a second path.

What should you build first?

Start with the in-product growth agent, because it monetizes the users you already have: personalized unclaimed value plus 1 recommended action, computed from the account's own data, validated against production before a single user sees it.

Then the definitions layer underneath it. An agent acting on your metrics needs 1 governed source of truth for what activation, retention, and expansion mean. Loose definitions were an annoyance when humans read the dashboards. They are a production bug when machines act on them.

Then the agent-readable surface: the protocol endpoint, the documentation an agent can execute against, the task-completion path that never touches your UI. Distribution is moving there whether your product is present or not.

The uncomfortable continuity

Teams waiting for the PLG-is-dead debate to resolve are missing that both outcomes point at the same work. If PLG survives, it survives as the version where the product carries more of the motion. If agents take over, the products that win are the ones agents can evaluate and operate. Either way, the build list is identical.

AI-led growth is PLG with the intelligence moved inside. The product was always supposed to sell itself. Now it can also understand who it is selling to, one account at a time, and act on it.