Every customer success team has lived through the same uncomfortable meeting. A renewal is coming up, and the CSM swears the account is fine. "They're using the product," they say. "We had a good call last quarter." And then the customer churns — or worse, they quietly downgrade, and nobody catches it until the number hits the board deck.
The problem isn't that your CSMs don't care. The problem is that they're working from incomplete signals. Product usage data lives in one system. CRM notes live in another. Support tickets are somewhere else. And the mental model any given CSM carries for an account is assembled from all of these disparate sources, filtered through memory, filtered through optimism, and never quite complete enough to act on with real confidence.
That's the problem Pendo Predict was built to solve. In this episode of RevOps Demos That Don't Suck, we sat down with Kobi Stok, who leads predictive AI and employee-related product functionality at Pendo — and who most recently served as CEO and founder of Forwrd AI before it was acquired — to walk through how Pendo Predict turns raw product and business data into actionable, AI-generated signals your team can actually use.
Pendo is best known as a product analytics platform — the layer that tells you what your users are actually doing inside your product. But Pendo Predict extends that foundation into predictive territory, combining product usage data with business data from your CRM, marketing automation, data warehouse, and other sources to generate account-level predictions about churn, upsell, cross-sell, and expansion risk.
What makes Predict genuinely different from a standard health score is the depth of the underlying model. This isn't a weighted formula someone built in a spreadsheet. It's a deep learning and boosted trees model that runs cross-correlation analysis across your entire historical dataset to identify which signals actually predict the outcomes you care about — not which signals you assumed would matter when you built the dashboard.
As Kobi put it:
"We don't sell models. We sell you the ability to build your own AI agents from your data." — Kobi Stok
Churn prediction tools have been around for years. The reason most of them disappoint isn't the model — it's the workflow. A score that lives in a health dashboard nobody checks is not a tool. It's a report. And reports don't prevent churn.
The clean data problem is also real: most organizations have years of product and business data sitting in warehouses and CRM fields that have never been properly normalized or structured for AI. Predict addresses this with an automated data preparation layer that classifies inputs by intent, demographics, and activity, then cleans and normalizes the dataset before any model training begins.
But the more important gap Predict fills is the adoption gap. If your team has to log into a separate tool to see churn risk scores, most of them won't. The insight has to live where the work happens. That's the design principle running through everything in this demo — and it's the reason the Predict architecture is built around integrations and overlays rather than standalone dashboards.
This connects directly to a challenge RevOps teams know well: the ticket-taker mindset that comes from building tools nobody uses. Predict's approach flips the equation — rather than asking CSMs to change their workflow, it brings the intelligence into the workflow they already have.
Pendo Predict connects natively to product usage data — naturally — but it goes substantially further than that. The platform integrates with major data warehouses, CRM systems, and marketing automation platforms to pull in the business context that product data alone can't provide.
If your organization is already a Pendo customer, that integration is immediate. If not, Predict can connect to other product analytics tools and data sources to reconstruct the same signal set. The point is that churn and expansion risk are never purely product phenomena — they're shaped by support ticket volume, contract size, marketing engagement, lifecycle stage, and a dozen other variables that live outside your product instrumentation.
Kobi described the core challenge this way: "Combining those two together [product data and business data] is very, very challenging, and this is one edge that we offer as Predict."
Raw data doesn't go directly into a model. Before any prediction runs, Predict's AI agents automatically classify, clean, and normalize incoming data — differentiating between intent signals, demographic attributes, and activity patterns to separate signal from noise.
After data preparation, the system runs cross-correlation analysis historically across your dataset to identify which variables actually predict the business outcome you've defined. This isn't a one-time setup. Predict retrains its models daily, pulling fresh data and updating predictions so your team is always working from current signals rather than a snapshot that's six weeks old.
This is where Predict becomes operationally useful rather than just analytically interesting. RevOps and CS leadership define the business outcome they want to predict — churn, upsell, cross-sell, expansion, any outcome for which you have historical samples — and Predict builds the model around that objective.
More importantly, Predict supports the integration of existing playbooks. PDFs, presentations, Google Docs — any format that describes how your team handles risk, opportunity, or account escalation can be ingested directly. Confluence integrations are in development. Once playbooks are loaded, Predict doesn't just flag at-risk accounts: it generates a recommended success plan for each one, drawn directly from your own documented procedures.
The implication for RevOps teams responsible for process consistency is significant. This is standardized execution at scale — not every CSM interpreting a health score differently and taking different actions, but every CSM working from the same playbook, prompted by the same signal, in the same workflow.
The most practically useful part of the demo was a live walkthrough of how Predict surfaces inside Salesforce — without any Salesforce development work required.
In the demo, a customer success manager's standard Salesforce dashboard is augmented in two ways. First, a custom field on the opportunity object shows the Pendo Predict churn classification directly in the pipeline view — so when a CSM reviews their book of business, churn risk is already there, sliceable and filterable without leaving the report they already use.
Second, and more interestingly, a notification widget appears at the bottom of the screen — surfaced via Pendo for Employees, Pendo's in-app guidance layer for internal tooling. This overlay requires no Salesforce admin involvement to deploy. It appears automatically and shows the CSM how many tasks require their attention based on the AI's current predictions.
Clicking into that task list reveals a prioritized view of accounts organized by risk category — in the demo, these were "critical high churn," "high churn," and "absent." Drilling into a specific account surfaces a natural-language explanation of why Predict flagged it, followed by a step-by-step success plan generated from the integrated playbook.
This is the workflow vision fully realized: the AI generates the insight, the playbook generates the plan, and the CSM sees both inside the tool they were already using. The insight-to-action path has no handoff gaps.
For organizations building their own AI infrastructure, Predict is accessible via Model Context Protocol (MCP) — meaning Predict's predictions and agent outputs can be queried through Claude or other large language model (LLM) platforms your organization already works with.
This matters for RevOps teams thinking about AI strategy at the infrastructure layer rather than the tool layer. Predict isn't asking you to bet on a single AI interface. It exposes its outputs to whatever AI environment your organization is building toward. As Kobi noted: "It doesn't have to be Claude. It can be any of your favorite LLM vendors out there."
The output layer also includes native Slack alerts, Salesforce object creation, in-app and out-of-app notifications via Pendo Orchestrate, and direct sync to your data warehouse — so whatever your team's system of record is, Predict can write to it.
One nuance worth flagging for RevOps practitioners designing the governance model: Predict supports a tiered automation approach. Not every action has to be fully automated. The platform includes a permission model that lets you configure where Predict runs end-to-end automation and where it stops to let a human decide.
For enterprise or high-touch accounts where a CSM should review and approve before any outreach happens, you can configure that. For lower-tier accounts where the economics of high-touch review don't make sense, you can let Predict execute the playbook automatically — sending emails, scheduling calls, triggering in-app notifications — without human intervention.
This is the kind of design choice that separates tools built for RevOps practitioners from tools built for demos. The operators who have to explain their automation governance to legal and compliance will appreciate that this is configurable, not assumed.
Company size: Mid-market to enterprise. Predict is designed for organizations with enough historical data to train meaningful predictive models and enough account volume to make automated prioritization valuable. Very early-stage companies may not yet have the data volume to generate reliable predictions.
Primary use case: Customer success teams managing renewals, churn prevention, and expansion. The demo is centered entirely on post-sale motions — this is not a top-of-funnel tool.
Tech stack signal: Organizations already using Pendo for product analytics will get the fastest and deepest integration. But Predict is explicitly designed to work with other product analytics tools and third-party data sources. The requirement is data, not a specific existing tool.
Existing Pendo customers: If your organization already uses Pendo, Predict represents a meaningful extension of infrastructure you've already deployed. The Pendo for Employees overlay and Pendo Orchestrate communication layer are already part of your stack.
RevOps + CS alignment: This product is equally relevant to RevOps teams designing the systems and CS leaders managing the workflow. The configuration and model-building work lives in a back-office UI that RevOps would naturally own; the CSM-facing experience is designed for zero additional training.
Pricing details were not discussed in the demo. For current pricing information, reach out directly through Pendo's website.
The primary outcome Predict is designed to deliver is churn prevention — specifically, earlier identification of at-risk accounts at a time when intervention can still change the outcome. The secondary outcome is standardized execution: every CSM working from the same AI-generated success plan, based on the same playbook, rather than each person interpreting a health score and deciding independently what to do.
The Salesforce overlay in the demo illustrates the adoption argument: when the tool lives inside the system your team already uses every day, adoption isn't a change management problem. The insight finds the CSM. The CSM doesn't have to find the insight.
For RevOps teams thinking about how AI fits into the broader revenue system, this is a useful case study. The model is sophisticated, but the real leverage is in the deployment architecture — making sure the output of the model lands somewhere useful, not just somewhere impressive.
The churn prediction problem isn't new. What's new is having a system that can ingest product data and business data, train its own models on your historical outcomes, generate natural-language explanations a CSM can actually act on, and surface all of it inside Salesforce without a custom development project.
For RevOps and CS operations leaders who have watched health score initiatives stall because nobody changed their Salesforce workflow to check a new dashboard, Pendo Predict takes a meaningfully different approach. The intelligence comes to the rep. The plan comes with the flag. And the governance layer gives you control over where automation runs and where a human stays in the loop.
If your organization is already a Pendo customer, this is a compelling reason to look closely at what Predict can add on top of infrastructure you've already paid for. If you're not, it's worth understanding what the combined Pendo platform actually offers — because the sum is considerably more useful than any of its individual parts.
Request a personalized demo or learn more on their website.
Check out our blog, join our community, and subscribe to our YouTube Channel for more insights.
