PostHog will begin training its own machine‑learning models on customer data to make existing analytics tools smarter and to launch new capabilities like automated session‑replay insights and synthetic user testing. The program starts June 29, is opt‑out for EU customers, and promises full data anonymization and no third‑party sharing.
PostHog’s New AI Ambition
PostHog is moving from adding isolated AI helpers to building self‑driving product tools that act on data automatically. The company’s latest beta, PostHog Code, is a product editor that aims to surface answers, suggest improvements, and even execute changes without manual prompting. To make that vision practical, PostHog plans to train its own machine‑learning models on the data already collected in customers’ PostHog instances.
What the Team Wants to Build
- Smarter existing features – Enhancements to the AI installation wizard, session replay analysis, and the broader PostHog AI suite will become more proactive, delivering insights before a human even looks at the data.
- Whole new products – The flagship of this effort is PostHog Code, a code‑assistant that helps teams ship faster by automatically detecting broken flows, suggesting UI tweaks, and even generating synthetic user tests.
Two concrete use cases illustrate the direction:
- Scalable session‑replay analysis – Today, PostHog AI can flag issues in individual replays, but the process is costly and doesn’t scale. A model trained on the underlying event streams could flag patterns across thousands of sessions in near‑real time.
- Synthetic user testing – By learning typical user journeys, a model could simulate how a new feature might be used, predict where users could get stuck, and recommend adjustments before the code reaches production.
The goal is to reduce the manual effort that developers currently spend on debugging and A/B‑testing, while also cutting the number of AI tokens burned in the process.
How the Training Will Work
PostHog has sketched a straightforward pipeline:
- Data selection – Only data that already lives in a customer’s PostHog instance (event streams, session replays, feature flag logs, etc.) will be used.
- Anonymization – Before any model sees the data, personally identifiable information is stripped out. The resulting dataset is a collection of abstracted user actions, not raw user profiles.
- In‑house training – All model training happens on PostHog’s own infrastructure. The company explicitly states it will not ship data to external model providers, nor will it sell the trained models.
- Opt‑out defaults – EU‑hosted customers are opted out by default, as are any accounts bound by agreements that prohibit data training (e.g., BAA, MSA). All other US‑hosted customers are opted in, but can opt out at any time via org settings.
- Launch timeline – Training will not begin until June 29, giving customers a window to review the policy and adjust settings if needed.
The company will announce the change through multiple channels: a dedicated email, in‑app notifications, and public blog posts like this one. Transparency is positioned as a core value, contrasting with the “boring T&C update” many firms use for similar changes.
Why an Opt‑Out Model?
PostHog argues that a sufficient volume of data is essential for the models to be useful. If too many customers opt out, the resulting dataset would be too small to capture the diversity of real‑world user behavior, limiting the accuracy of the insights.
Customers who stay opted out will continue to receive the current PostHog feature set, but they will miss out on the upcoming AI‑driven capabilities that depend on the trained models. Conversely, those who stay opted in will benefit from faster, more automated analysis and the new synthetic testing tools.
What This Means for Users
- Data security: Anonymization and in‑house training keep the raw data away from third‑party clouds.
- Feature access: Opt‑in customers will see new AI‑powered suggestions in real time, while opt‑out users will retain the existing manual workflow.
- Control: Organizations can toggle their participation at any time through the admin console.
PostHog’s approach reflects a broader trend among analytics platforms: shifting from reactive dashboards to proactive assistants that act on data automatically. The success of this experiment will hinge on the quality of the training data, the robustness of the anonymization pipeline, and how well the new models integrate with existing developer workflows.
Looking Ahead
PostHog is also hiring AI researchers, indicating a long‑term commitment to building these capabilities internally rather than licensing external APIs. If the models deliver on their promise, we could see a future where product teams spend less time digging through logs and more time iterating on features based on AI‑generated recommendations.
For anyone interested in the technical details, the company has promised to publish a follow‑up post once training begins, likely including model architecture choices, evaluation metrics, and early performance numbers.
PostHog remains an all‑in‑one developer platform offering product analytics, session replay, feature flags, experiments, logs, data warehouse, CDP, and an AI product assistant. The new training initiative is the latest step toward turning that stack into a self‑optimizing product engine.
Comments
Please log in or register to join the discussion