Amplitude shares AI-led homepage experiment with a claimed 160% lift in card-driven messages

Amplitude’s AI product agent found a homepage feature that users largely ignored, then helped its team change it. In an OpenAI customer case study published November 12, Amplitude says personalized suggestion cards increased card-driven messages by 160% after one day. The result measures one route into a conversation—not overall use of the product.

The suggestion cards users skipped

The experiment began with generic homepage cards designed to help users start a conversation with an AI agent. Wave examined a sample of 100 session replays—recordings of how people used the interface. Eighty users typed their own questions; only 12 selected a suggestion.

Wave proposed replacing the generic prompts with suggestions drawn from recent charts, dashboards or the page a user had arrived from. It illustrated the proposed interface and prepared a technical plan. That plan preserved free typing and retained generic suggestions when there was no context for personalization.

After a brief review, the team handed implementation to a coding agent and ran an A/B test, comparing the change with a control. Amplitude’s experiment readout showed the increase in messages sent through the cards after one day. That is the specific outcome behind the case study’s broader description of an agent-engagement lift.

Different models for different parts of the job

The homepage test follows Wave’s wider workflow. Teams supply a product area, goals and metrics. Wave combines analytics, session replays, customer feedback and code to find opportunities and develop solutions. It can send coding agents, including Codex, a technical plan and acceptance criteria—the conditions an implementation should meet—then track the outcome.

Amplitude describes Wave as roughly 80% powered by OpenAI models. GPT-6.1 Sol handles opportunity management and deep investigations. GPT-6 Luna takes frequent, lower-cost work such as code scans and small checks. GPT-Image-2.5 Flare creates mockups using screenshots from session replays as references.

Amplitude’s Opportunity Investigator evaluation tested whether models supported findings with evidence, avoided false positives, incorporated feedback and recovered from tool failures. Sol scored 4.06 out of 5 overall, ahead of three alternatives scoring 3.62, 3.52 and 2.94. It also had the fastest median investigation time, at 54 minutes.

The separate Data Science Agent evaluation shows why the model choice was not simply about winning every score. Sol cost less but finished below another leading model overall. It scored higher on following tool requirements, handling estimates and uncertainty, and grounding answers in evidence. These are Amplitude’s task-specific evaluations, presented in its model supplier’s customer case study.

Teams still decide what ships

Customers across industries are already using Wave in beta, according to Amplitude. Teams choose their level of automation: most send bug fixes straight to coding agents, while new features and bigger bets receive human review. Finding a change and authorizing it are separate steps in that workflow.

Amplitude estimates Wave can shorten shipping a product improvement from roughly a month to one or two days. Its customer shipping observations describe a different measure: among customer-launched experiments, flags and guides, one-quarter went live the day Wave surfaced the opportunity, half in under four days, and two-thirds within a week. Those figures cover changes that reached launch, not every recommendation.

CEO Spenser Skates describes the next ambition as making self-improving products the way companies build software. He expects teams to delegate more investigation and implementation as models improve. That remains a forward-looking goal; the current beta leaves teams in charge of how far automation goes.