Product-led growth was the dominant motion in B2B SaaS for ten years for a defensible reason: the asset shipped with the seat. A user who signed up for Slack got a workspace they could send messages in immediately. A user who signed up for Figma got a canvas they could draw on immediately. The product delivered value on day one, the user adopted, and the adoption pulled the rest of the team through. PLG worked because the value loop was tight, the asset was self-contained, and the buyer's organization could be entered through any user with a credit card.
Enterprise AI breaks the assumptions PLG relied on. The asset that compounds — the corpus, the model fine-tunes, the editor-in-the-loop signal — does not exist on day one. The free tier of an AI product can deliver the feature, but it cannot deliver the architecture. The adoption that follows is shallow, because the user has access to a model that does not know anything about the buyer's business yet, and the model will not know anything about the buyer's business until the buyer's IT and security teams have permitted the data ingest that PLG was designed to bypass.
What we are watching, in the procurement environments where PLG-style AI products try to land, is a consistent failure pattern. The user signs up. The user has a thin, generic experience. The user finds the experience underwhelming relative to the marketing. The user shrugs and goes back to the legacy workflow. The PLG motion delivered a seat without delivering an asset, and the asset was the value.
Why the asset-first frame breaks PLG.
Three structural reasons PLG underperforms for enterprise AI.
The corpus requires institutional permission. The corpus is the buyer's data. Ingesting it into a vendor's environment requires security review, legal review, and data-classification work that no individual user can authorize. PLG bypasses these reviews structurally; without them, the AI product cannot reach the data that would make it useful. The free tier is, by design, a generic experience.
The compounding loop takes weeks to start. Even if the data ingest happens immediately, the editorial corrections that produce the compounding curve take weeks of editor activity to accumulate. A user evaluating the product on day five has not seen the compounding signal yet. The decision to adopt is being made before the value has been demonstrated.
The procurement audience is not the user. PLG worked because the buyer was reachable through any user. In enterprise AI, the buyer is the architecture committee, the CIO, and the data governance lead. None of those people are signing up for a free tier. The user who does sign up does not have the authority to escalate the evaluation to the people who actually decide.
These three forces compound. The PLG motion brings a user into a thin experience, the compounding signal does not show, and the user does not have authority to escalate. The funnel converts at a fraction of the rate the PLG team modeled, and the cost-per-deployment ends up higher than a properly run architecture-led motion would have been.
What replaces PLG: the architecture-led motion.
The motion that has worked, on our cohort and on the cohorts we benchmark against, is architecture-led. It looks more like the consultative enterprise sales motion of the early 2000s than like the PLG motion of the 2010s, and it has the operational economics to match — higher CAC, longer sales cycles, much higher net retention.
The architecture-led motion has three phases. The first is the architecture call: a 45-minute conversation with the buyer's CIO and operations lead in which the vendor maps the buyer's workflows to the proposed architecture, produces a replacement-math estimate, and identifies the three pillars where the deployment will land first. The second is the install: a defined ninety-day rollout that delivers a working architecture against the three pillars, with a fixed price and a defined deliverable list. The third is the expansion: as the architecture proves itself, the buyer expands depth on the existing pillars rather than adding new vendors.
The architecture-led motion produces deployments that compound. The PLG motion, when applied to enterprise AI, produces seats that decay. The economics favor architecture-led decisively, even though the upfront CAC looks worse on the deck.
What this means if you are building an enterprise AI product.
If you are designing the go-to-market motion of an enterprise AI product, the PLG playbook is the wrong playbook. It will produce a top-of-funnel that looks healthy and a bottom-of-funnel that does not convert. The motion that converts has different ratios, different surfaces, and a different sales team.
Three changes that have worked, in the products we have advised through this transition.
Replace the free tier with an architecture call. A 45-minute conversation, no friction, with the buyer's actual decision-makers. The conversation produces a replacement-math sketch the buyer can take to their board. The free tier — generic experience, anonymous signup — does not produce this output and never converts the actual buyer.
Replace self-serve onboarding with an install. The install is opinionated, fixed-price, fixed-duration, and produces a working architecture at the end. The buyer is not asked to figure out how to deploy; the vendor's install team does it. The install produces the asset that PLG was structurally incapable of producing.
Replace seat-based pricing with [architecture pricing](/dispatch/per-seat-pricing-is-wrong-for-ai). The pricing model has to match the value model. Per-seat encourages PLG; architecture pricing encourages depth.
If you are buying an enterprise AI product that is being sold PLG-style, the question to ask is whether the product can deliver the architecture without the architecture call. In our experience, it cannot. The free tier will be underwhelming for the structural reasons described above. The decision to adopt or not should be based on what the architecture call produces, not on what the free tier delivers in week one.
PLG had a great run. It is not the right motion for the next decade of enterprise AI. The motion that is replacing it is more honest about what the buyer is buying, and the deployments that result are the ones still standing in month eighteen.