Using the Copilot

The AI Copilot is BPP's in-product assistant. It answers questions about your workspace and performs platform operations for you — always scoped to your customer workspace, and never writing anything without your approval.

Where to find it

• Chat drawer — click the sparkle button in the bottom-right corner of any page to open the Copilot in a side drawer, without leaving what you're doing.
• Copilot workspace — open the AI Copilot section from the navigation for a full-page chat with your conversation history. Previous conversations are saved and can be resumed.

How it works

You write in natural language; the Copilot decides which platform operations (tools) it needs, runs them, and streams the answer back. Tool calls are shown inline in the conversation, so you can always see what the assistant looked at to produce an answer.

The Copilot can work with:

• Your data — tables, fields, data quality, field value previews, and read-only BigQuery queries.
• Audiences and signals — listing, inspecting, creating, updating, cloning, and running them, including previewing signal size before saving.
• The identity graph — reconciled user counts, identifier coverage, single-profile lookups.
• Connections and destinations — configured platforms and where each audience or signal is pushed.
• AI models — model instances, model types, and guided setup of new instances.

Human-in-the-loop approvals

Read operations run immediately. Write operations never do.

When the Copilot wants to change something — create, update, clone, or run an audience or signal, bind a destination, create or delete an AI model instance, or edit field descriptions — the conversation pauses and shows an action card describing exactly what is about to happen. You choose:

• Approve — the action is executed and the conversation continues with the result.
• Reject — nothing is changed, and the Copilot moves on.

info

Entities created by the Copilot start in draft status, exactly as if you had created them in the UI — they don't run or push anywhere until activated.

Example prompts

Explore your data

Which tables do I have in my datasource, and what's in the event table?

Check the data quality of the email field in my user table — how many nulls and distinct values?

Run a query: how many orders per month did we record in the last 6 months?

Identity graph

How many reconciled users do we have, and what share of them has an email identifier?

Audiences & signals

Create a draft audience of users with at least one purchase in the last 30 days.

Preview how many users a cart-abandonment signal would match if the trigger is "added to cart but no purchase within 3 days".

Clone the "High LTV customers" audience and restrict it to users acquired this year.

Connections & AI models

Which destinations is the "Newsletter subscribers" audience pushed to?

List my AI model instances and tell me which ones produced predictions this week.

A worked example

1. You ask: "Create a draft audience of users with an order in the last 30 days."
2. The Copilot inspects your tables and fields to build the right filter, then shows an action card: create_audience — "Users with an order in the last 30 days", with the full filter definition.
3. You click Approve.
4. The audience is created in draft status and the Copilot replies with its ID and a link — you can review it in the Audiences section, attach destinations, and activate it whenever you're ready.

Limits

• Approval required for every write — there is no way to bypass the action card.
• Draft-first — created audiences, signals, and AI model instances always start as drafts.
• On-demand runs are throttled — manually triggering an audience or signal run is subject to a per-customer daily cooldown.
• Conversation limits — requests are rate-limited, and a single answer can only take a bounded number of tool steps; very broad requests work better split into smaller ones.