How to use the Predict agent (beta)

The Predict agent is an AI-powered conversational assistant built into every Predict model. After your model is trained, the agent helps you validate it, explore data and predictor patterns, decide which records to focus on, and get recommended next steps based on your own playbook.

This article explains what the Predict agent is, when to use it, and how it works.

Beta: The Predict agent is in beta. Functionality and limits described in this article may change before general availability.

Prerequisites

To use the Predict agent:

• A trained Predict model in your project. The agent appears as a tab inside the model's analysis once training is complete.
• Project access. Anyone with access to the project that contains the model also has access to its agent.

Predict agent use cases

Once your model is trained, the Predict agent supports four areas:

The agent is most helpful when you want to:

• Decide whether the model's scores can be trusted before acting on them.
• Explore predictor patterns.
• Triage a large set of predictions down to the records that matter right now.
• Diagnose why a specific record received its score, with the reasoning in clear, straightforward language.
• Generate a save plan tied to a single record, drafted against your own playbook.
• Iterate on a plan or a draft communication.

How the Predict agent processes your query

The agent uses large language models hosted by Pendo. It interprets your prompt, retrieves grounding from the sources below, and returns a structured, streamed response.

Sources include:

• Your trained model. Score, predictor contributions, training and audience statistics, and per-record signals.
• Connected playbooks. PDFs, slide decks, and documents you've uploaded to the model's knowledge base. The agent retrieves relevant sections using RAG (retrieval-augmented generation) and cites them in its responses.
• Custom instructions. Project-level text you write to tell the agent how to explain scores, how to generate plans, and any global context it should always apply.
• Skills catalog. Customer-defined plays (for example, "Executive Business Review" or "Win-Back Outreach") the agent can choose from when building a plan.

Note: We don't use your data to train our AI models. For more information, see Pendo's Trust Center.

What the Predict agent doesn't do

• It doesn't send communications. The agent drafts emails, meeting agendas, and other artifacts, but it never sends them. You copy each draft into your own tool to send.
• It doesn't simulate alternative scenarios. The agent doesn't answer questions like, "What if Acme were in a different region?" or, "What if their ARR doubled?" — predictions are grounded in actual training data.
• It doesn't modify model settings through chat. Settings changes are made on the Settings page and the agent will redirect you there.

Use the Predict agent

After your model finishes training, the agent will appear as a new Agent tab inside the model's analysis evaluation section.

To use the agent:

1. Go to your trained model's analysis.
2. Select Evaluation on the left.
3. Select the Agent tab.

The first time you open the agent, it'll start with a generated insight about your data. You don't have to know what to ask. The agent suggests follow-up questions you can pick from, or you can type your own.

When you return later, you can pick up from where you left off and continue your conversation. Scroll up to revisit earlier messages.

Reference data in your prompt

In the chat, you can reference any data point from your model's decision base by name, such as accounts, opportunities, predictors, fields, segments, or other entities the model knows about. Type the name into your prompt and the agent will resolve it. It matches partial names and asks for clarification when there's ambiguity.

Tip: To get deeper insights into a specific record, reference it by name in your question (for example, "Why is Acme Healthcare at risk?"). The agent opens the record widget on the right and explains the score in context.

Note: To pin a specific data point that the agent should always know about, add it to your Custom instructions in Settings. Custom instructions support @ mentions for inline references.

Viewing record details

Open a record's full view from a chat or from any list the agent returns. The view includes reasoning, recommended actions, and drafts.

When you select a record, the record widget opens as a sidebar. The widget contains:

• What's driving this prediction. A headline reason for the score (for example, "High risk of churn at renewal: exec sponsor is gone and billing concerns are surfacing") followed by a clear, straightforward summary, key risk factors, positive indicators, and business impact. These all grounded in the signals the model used.
• Success plan. A short, ordered list of next steps tied to the record. Each action card shows a title, description, recommended timing (relative to renewal or another anchor date), the skill that produced it, and a quote from the playbook section it draws from.
• Drafts. If an action produces a draftable artifact such as an email or agenda, you can select it on the action card and view them as read-only from the widget. You can edit drafts using the agent chat. 

The widget is the same view your CSMs see in the CRM widget once the model is activated, so the conversation you have in the agent translates directly into the surface they work in.

Note: When a record is in a closed state (already renewed or already churned), the widget shows only what drove the prediction, no recommended actions, no drafts.

Refine plans and drafts through chat

You can edit plans and drafts by asking the agent. 

Examples:

• "Change action 2, we already tried that. Suggest a different approach."
• "Make the email shorter and don't mention pricing."
• "More formal tone for the meeting agenda."

The agent regenerates the action or draft. The conversation keeps a full history and past responses aren't rewritten when you make a change.

Configure the agent

The agent's behavior is controlled by the Settings tab inside the analysis. Anyone with project access can view and edit settings; configuration applies to every user in the project.

There are four sections:

Playbooks

A list of all playbooks published in the project's playbook library. Toggle a playbook to connect or disconnect it from the agent. Connected playbooks are made available to the agent's RAG retrieval.

Each row shows the playbook name, last-published date, and description. For details on uploading and managing playbooks, including supported file formats and size limits, see Use Playbooks in Predict.

Skills

The catalog of skills the agent can choose from when building a plan. Each skill defines a type of action, for example, "Executive Business Review", "Win-Back Outreach", or "Adoption Push".

A skill has:

Select Extract from playbook to have the agent read a connected playbook and propose candidate skills. The button is enabled only when a new playbook has been connected, or when a connected playbook has been updated. Extracted skills appear as Active.

Custom instructions

Free-text project-level context you want the agent to use. There are options:

Score labels

Rename the four scores so they match your internal language. Defaults are inherited from the model's use case (for example, Critical / High / Low / Minimal for churn).

About the model

The About the model view is a read-only summary of the model's training and evaluation results. It includes:

• Use case and description. From the model's business logic.
• Headline accuracy. Overall accuracy on the held-out test set.
• Accuracy breakdown. Performance on records that actually succeeded vs. failed (for example, churned vs. renewed).
• Propensity by score. For each score (Critical / High / Low / Minimal), the percentage of records that actually had the outcome in the test set.
• Top model features. Ranked list of predictors by importance.
• Audience distribution. Count and share of records in each score band, shown as a pie chart.
• Data sources. The systems that contributed data to the model (for example, Salesforce, Pendo).
• Training metadata. Training record count, prediction window, last trained date, model version.
• Record drill-down. Zoom into the records behind the model. Search, filter, group, edit columns, and export to CSV. Available for both the test set and the audience.

Manage conversations

Each user has a single rolling conversation per model and there's no per-session boundary and no timeout. Messages show in chronological order, oldest at top, newest at bottom.

You can:

• Reset the conversation. Select the New Chat button in the top right to clear your chat history (and any plans or drafts tied to it). Settings and model state are unaffected. After a reset, the agent opens fresh.
• See past messages as immutable. The agent never rewrites earlier responses when you change a setting or when the model is retrained. If a setting changes mid-conversation, a brief system message is inserted at the timestamp of the change so you can see what was different.

Note: When a model is retrained, all users' conversations are automatically reset. Past responses were grounded in the previous model; resetting prevents stale or contradictory answers.

Sharing

The Predict agent inherits its sharing model from the project that contains the model. Anyone with access to the project has access to the agent.

When a shared user opens the agent for the first time, they start with a fresh conversation. Conversations are per-user, not shared.

Best practices and troubleshooting

For the most reliable results:

• Start specific. Ask about one record, one predictor, or one segment at a time. Use follow-up prompts to broaden.
• Reference records by name. The agent matches partial names and surfaces close matches when there's ambiguity.
• Use English-language playbooks. Non-English content isn't blocked, but response quality may vary.
• Use thumbs-down to flag drift. If a response looks wrong, mark it so the team can investigate.
• Reset the conversation when context drifts. If you've moved across several unrelated topics, a fresh conversation produces sharper answers.

If a response looks wrong

• The agent's data is current to the model's last training run. Recent activity outside that window isn't reflected.
• If you push back on a score, the agent will list the top signals that drove it. If those signals aren't accurate for your record, that's a data-quality flag worth raising with your Pendo team.
• If the agent says it can't answer a question (for example, a counterfactual or an out-of-scope ask), the limit is intentional, see What the Predict agent doesn't do section.