Deck Intelligence

Deck Intelligence is the in-app AI PM surface. Ask plain-English questions about your product signal — themes, opportunities, NPS, tickets — and get decision-oriented answers with links to the underlying feedback. Instead of navigating reports manually, you can ask questions like "What are enterprise customers complaining about?" or "How has NPS changed for churned accounts?" and get a direct, source-backed answer.

Availability

Deck Intelligence is available on every plan. Chat appears in the app shell for every authenticated workspace. Asking questions consumes credits from your organization's monthly allocation. Each answer is charged for the AI work it actually does, so a multi-step research answer costs more than a quick lookup. If you stop an answer or it fails partway, the work done up to that point is still charged. Deck checks your balance when you send a message. An answer that has already started runs to the end, including the steps after you approve an action. If your organization runs out of credits and overages are disabled, Deck pauses new questions until an admin enables overages or the next billing period starts.

The first-run onboarding chat is a separate surface. It is free for seven days from the start of that onboarding session, then uses credits the same way Deck Intelligence does. See Getting Started.

Opening Deck Intelligence

Chat appears inside the app shell on product pages (Visualize, Build, Manage, and full-page chat). Settings pages do not show this panel. On desktop, a docked panel opens beside your workspace by default so you can keep browsing Deck while the assistant works. Collapse it if you want more room — Deck remembers that choice in this browser. You can resize the panel, press Cmd+J (Mac) or Ctrl+J (Windows) to toggle it, or expand it into a full chat page.

If you just finished onboarding, Deck Intelligence greets you and suggests leftover setup tasks you can start from chat. Clicking a suggestion sends it as your message. The greeting itself is not saved: ignoring it does not save an empty conversation, and once you reply, the saved conversation starts with your first message. Expanding to the full chat page continues the same thread instead of starting a new one.

Expanding opens a bookmarkable route (/app/{your-org}/chat) with conversation history on the left: start a new chat, jump between recent conversations, and (when enabled) manage connectors. The New chat icon at the top of the conversation also starts a fresh chat. The full chat page stays inside the window: the transcript and answer board scroll in place, and the message box stays on screen. On mobile, the same full-page route is used; history opens as a sheet. Closing or going back returns you to the product page you expanded from, and Deck keeps that page’s context for the assistant.

Some screens may still show the compact floating chat button when Deck cannot reserve a docked panel, including on a phone. The compact launcher uses a sparkle icon.

Conversation history is private to your account in this workspace and saved to Deck, so recent chats come back when you sign in on another browser or device. Teammates cannot see your conversations. Empty new chats are not saved until you send a message.

An answer keeps running even if you close the panel, leave the page, or reload while Deck is still working. Deck saves it when it finishes, and when you come back, the conversation picks up where the answer is. If the connection drops for a few minutes, Deck asks you to reload the page to see the latest reply. To end an answer early, press the stop button (it replaces Send while Deck answers). What Deck wrote up to that point stays in the conversation. If a GitHub repository review is running, Stop asks you to confirm first, because that review takes several minutes and uses credits.

You can continue any saved conversation, including older ones. When Deck picks up an older conversation, it works from the saved messages rather than the full results of earlier lookups, so it may look data up again before it answers.

Asking Questions

Type your question in the input field and press Enter (or click Send). The assistant reads your organization's feedback data in real time and streams a response back. While it works, a thinking mark appears next to the working label; when the step finishes, that mark settles into the Deck logo. It selects the most relevant tools for the page you are on and the question you asked, so a question from a theme page can focus on themes and insights while a question from settings can focus on integrations or feature setup.

Deck Intelligence knows who is signed in: your name, whether you are an organization admin or a member, and (when set) your job title and function. It does not put your email in the conversation context. You can ask it who you are.

Deck Intelligence is framed as an embedded product manager: for roadmap and prioritization questions it aims for decision-oriented answers (bottom line, evidence, trade-offs, and a recommended next step), using signals such as volume, severity, segment concentration, NPS impact, and recency. Factual lookups stay concise. When evidence supports it, the assistant may recommend creating Opportunities or Projects — write actions still require your approval.

When it proposes solutions (feature ideas, Opportunity directions, or Project approaches), Deck Intelligence makes them AI-native by default. It says what the AI does, what data or signal it uses, and how your users stay in control. Problem statements stay free of solutions, and Projects it drafts say where AI is built into the approach. Bugs, broken flows, and reliability or performance problems get the direct fix first. It adds an AI-native follow-up idea only when that helps.

Good questions to try:

• "What themes are most common among detractors this month?"
• "Summarize the top complaints from Zendesk tickets in Q1"
• "Which customer segment has the highest NPS?"
• "What should we prioritize on the roadmap this quarter?"
• "Which problems look like strong Opportunity candidates?"
• "Are there any insights about the onboarding experience?"
• "What subthemes do you see in the Pricing theme?"
• "Which subthemes have the most negative sentiment right now?"

Working With Responses

You can copy completed user or assistant messages from the message actions that appear next to each bubble. Copied assistant answers include only the visible response text, so hidden follow-up metadata and source markers are left out.

You cannot edit a message after you send it. To change a question, send a new message (for example, a corrected version of the question), and Deck Intelligence answers from there.

On the full chat page, longer conversations include a compact turn navigator anchored to your prompts. Use it to jump between major turns without losing the latest-response jump control.

While a new answer streams, Deck keeps the matching question anchored so the response grows underneath it. If you scroll manually or jump elsewhere in the transcript, Deck clears that anchor and lets you browse freely.

Citations

Every factual claim in the assistant's response is backed by a citation — a direct link to the insight, theme, NPS response, customer service ticket, or competitor finding the answer draws from. Citations appear as numbered references in the text.

While Deck Intelligence works, a Working… status shows instead of step-by-step reasoning or tool steps. You only see the finished answer. On the full chat page, open the Sources chip under the answer to see every source it used; that page also drops progress notes such as “I will search…” between tool calls. Pending approve/deny cards still appear when a write needs your confirmation.

Click any citation to open the original piece of feedback in a new tab.

Answers that use your team sources cite them too. A code citation names the file and opens it on GitHub at the exact version Deck read. An analytics citation names the PostHog query Deck ran, such as "PostHog · Onboarding Started to Onboarding Job Completed, last 30 days". It has no link.

When Deck Intelligence creates a new Project, Initiative, Opportunity, or Theme for you, a compact link card appears near the response so you can open the newly created asset directly from the chat. The link card stays outside the sources list so it is never buried.

Charts, Tables, and Media

For visual questions, Deck Intelligence can show trusted tables, charts, and media. These visuals come from Deck’s own read tools (not free-form AI drawings), so they stay aligned with the same data the Sources list cites.

Examples of what can appear:

• Charts for NPS category breakdowns and similar distributions
• Tables for theme or segment volume, insight lists, quote samples, and category overviews
• Video clips when an insight already has a trusted recording snippet

On the full chat page, those visuals appear on the answer board to the right of the conversation (a drawer on a phone), not inside the transcript. The board stacks every answer newest first, and tables the assistant writes in its reply move there too. File downloads stay in the conversation. Changing board filters refreshes the latest agent board once the chat is idle. In the docked or floating panel, tables, charts, and media still appear inline in the conversation.

Opening an older conversation replays the same snapshot that was saved with the message—Deck does not re-run the query just to redraw the chart or table.

Approval requests stay visible as inline cards so you can approve or deny changes. On the full chat page, the outcome stays visible after you answer.

Page Context

Deck Intelligence is aware of which page you are on. When you open the chat while viewing a specific theme, subtheme, insight, interview, NPS view, Project, Initiative, Opportunity, or Visualize → Competitors, the assistant automatically focuses on that topic without you needing to repeat the context. On Competitors, starter questions cover rival gaps, disliked patterns, promoting findings, and the last check-in. When a specific entity is active, Deck shows its name above the input so you can see what context will be sent. You can still ask general questions — page context only narrows the default scope.

Using Connectors

When connectors are enabled for your workspace, use the Connectors button beside the chat input to connect supported external services. The menu lists team sources (GitHub and PostHog) above services connected just for you. Choose Add a connector to search the personal catalog, then complete the provider's authorization flow in the secure popup. Deck confirms the connection before continuing the conversation.

The Connectors menu has two parts. Team lists GitHub and PostHog, which an admin connects once for everyone (see Team Sources below). Just you lists the services you connected for your own chats. You can disconnect one of yours from its toggle, reopen the catalog to reconnect or remove it, or choose Manage connectors to open your workspace integration settings.

You can also ask Deck Intelligence to use an external service directly. If a connection is missing or needs attention, the assistant shows a Connect or Reconnect card in the conversation. After you finish connecting, Deck adds a visible confirmation message and resumes your request automatically.

Team GitHub and PostHog

Ask Deck Intelligence to connect GitHub or PostHog for the team, or use the team rows in Connectors. The assistant shows a connect card in the chat:

• Admins connect and finish setup inside the card. For GitHub, pick the repository and branch from searchable lists (repositories you recently pushed to, with each default branch filled in). If a list cannot load, you can type the names instead.
• Members see which admins to ask. They cannot connect GitHub or PostHog for the team.

Team PostHog in chat is product analytics (events and journeys). It is separate from the PostHog survey import in Settings → Integrations.

The first time an admin saves a GitHub repository and branch, or when they change them, Deck starts a repository review. A Reviewing your repository row appears while it runs (about five minutes). Keep the chat open. When the review finishes, Deck saves the product map immediately and starts using it — there is no approval step. Deck shows what it mapped (including features added and removed) and asks what is wrong or missing. Later corrections replace the saved map.

If you press Stop while a review is running, Deck asks you to confirm. Stopping does not refund credits already used. If Deck had already saved a map before you stopped, that map stays.

External actions use the same approval experience as Deck changes. Read-only actions can run without interruption when Deck can verify that they are safe. Actions that may create, update, delete, send, or otherwise change external data require your explicit approval in chat. If Deck cannot confidently classify an action as read-only, it asks for approval. If such an action is interrupted partway, Deck does not run it again. It tells you the action may or may not have completed, so you can check the connected app before you try again.

Team Sources: GitHub and PostHog

Admins can connect two team sources that everyone in your organization can use in Deck Intelligence: GitHub, so Deck can read your product's code, and PostHog, so Deck can read your product analytics. Deck only reads from them. An admin connects each one once, and teammates don't need to connect anything.

With them connected, Deck Intelligence can check what customers say against how your product works and how people use it. Customer feedback still comes first, and every claim names its source.

Connecting a Team Source

Only organization admins can connect, set up, change, or disconnect team sources.

1. Click the Connectors button beside the chat input. Team sources are listed under Team.
2. Choose Connect for the team next to GitHub or PostHog and finish connecting in the popup.
3. Deck opens setup. For GitHub, enter the Repository (owner/repo, or paste a github.com link) and the Branch. For PostHog, enter the Project ID: the number after /project/ in your PostHog URL.
4. Choose Check and save. Deck checks that it can read that repository or project before it saves.

Step-by-step guides: GitHub and PostHog.

Each team source shows its state. Members see whether it is connected, waiting for an admin to finish setup, or needs an admin to reconnect, and which admins can connect it. Admins get Finish setup, Settings to change the repository, branch, or project, and Reconnect when the connection stops working, for example after the admin who connected it leaves the organization. To disconnect a team source, choose Add a connector and then Disconnect under For the whole team. Deck keeps the setup, so connecting again restores it.

GitHub and PostHog are team sources only, so they don't appear in the personal connector catalog.

Questions to try once they are connected:

• "Where do people drop off between signup and creating their first project?"
• "What do people do after they open the pricing page?"
• "Customers say CSV exports time out. What does the export code do?"
• "Which pages make up the billing feature, and how much are they used?"

Product Map

Deck keeps a map of your product, with its pages, features, and docs, so Deck Intelligence can say what exists in your product and where before it answers. The map comes from up to three places:

• Your public website. When Deck knows your company's website (for example from the company research it runs when you sign up with a work email), it reads your site's sitemap each day (until an admin saves a map from GitHub) and records up to 200 pages with their titles and descriptions. It follows your robots.txt, identifies itself as DeckBot/1.0, and never reads pages behind a login. This needs no team source.
• PostHog. With PostHog connected, pages people visit (three or more views in the last 30 days) are added with their view counts.
• GitHub. With GitHub connected, an admin reviews and saves a map built from your code. Pages found in the code then replace the matching website and PostHog pages.

Reviewing your repository. Right after an admin sets up GitHub, or changes its repository or branch, the chat sends "Review our GitHub repo and recommend a product map setup." on that admin's behalf, and Deck reviews the repository. Admins can also ask for a review at any time. The review takes a few minutes. Deck then shows a recommended map: your apps, features, docs, the apps it left out, and any questions. Reply with corrections and Deck revises it. When it looks right, ask Deck to save it and approve the Product Map Needs Confirmation card. Nothing from your code becomes part of the map until an admin approves it.

Keeping it current. Each day, Deck updates the saved map from the latest version of your branch: new pages are added and described, and pages that no longer exist drop off. The features, docs, and product terms you confirmed stay as you saved them. Ask for a new review to change them. Docs you confirmed (in your repository or on a public help site) become searchable, so Deck can answer how a feature is meant to work.

Repository reviews, and the descriptions Deck writes for new or changed pages, use credits from your organization's allocation. If credits run out during a review, the review stops. Reading your website and PostHog pages uses no credits.

Confirming Changes

Deck Intelligence can help with a limited set of workspace changes, such as updating Build records, managing segments, toggling supported settings, or working with Competitors. It can also perform approved actions in connected external services. A GitHub repository review saves the product map as soon as it finishes, with no approval card. Other write actions are gated:

• When the assistant proposes a write, an inline approve/deny card appears in the chat with a summary of the exact action, target, and fields that would change.
• You must approve the action inline before it runs. Denying cancels the action and the assistant offers alternatives.
• If you reload the page while a card is waiting, the card is still there and you can answer it.
• If you send a new message instead of answering the card, Deck cancels that action (it does not run) and answers your new message.
• Members can approve competitor research, adding a competitor, promoting a finding, and assigning competitor insights onto an opportunity. Changing the Competitors check-in schedule, and other admin-only writes, will not run for members.

Suggested Questions

When the chat is empty, a set of Starter Questions appears. These are tailored to the page you are viewing and help you get started quickly. Click any suggestion to send it immediately.

Limitations

• The assistant only has access to data within your organization and services you explicitly connect. It cannot browse the open web. It knows your public website only through the product map.
• Team sources cover one GitHub repository and branch, and one PostHog project, per organization. Deck never reads secret files (such as .env files and private keys), binary files, or files larger than 1 MB.
• A repository review can take up to about 10 minutes. Repositories larger than 200 MB can't be reviewed yet.
• Each analytics question covers up to one year of data, and journey samples up to 90 days.
• Available external actions depend on the connected service and the permissions granted during authorization.
• Very large date ranges may take a few extra seconds to process
• Each answer can run up to 25 AI steps (each step can look data up), including the steps after you approve an action. A very broad request can reach that limit before Deck writes a summary. If that happens, ask it to summarize what it found, or narrow the question.
• New questions require available credits or enabled overages
• Deck Intelligence only modifies workspace data after someone approves a supported write action inline, except the product map: a GitHub repository review saves the map as soon as it finishes. Most questions and analysis tools are read-only. Members can approve competitor research, add, promote, and assign competitor insights; check-in schedule changes stay admin-only.

Privacy

All questions and responses are scoped to your organization. Connector identities are also separated by user and organization, so a connection is not shared across organizations. External-service credentials are managed by the connector provider rather than stored by Deck.

Team sources are read-only. PostHog answers describe what people do, never who they are: Deck won't list people, emails, names, or user IDs, and removes personal-looking values from results. Sampled journeys show people as codes such as P1 and P2. Deck stores a daily list of your PostHog event names and page paths with their 30-day counts, with no information about people, and deletes entries not seen for 90 days. A repository review copies one version of your repository into a temporary, isolated workspace with no internet access, and deletes it when the review ends. When the admin who connected a team source leaves the organization, Deck removes that connection, and another admin needs to reconnect it. Queries are processed by Deck's AI infrastructure and are subject to the same data handling policies as all other AI features in the product.

Your saved conversations are stored with the rest of your workspace data in its hosting region. So that answers keep running when you reload or leave the page, Deck's agent hosting provider also keeps a working copy of each chat session outside your workspace's regional database. That copy holds your messages, the assistant's replies, and the results of lookups and connector actions.