How do I permission a company knowledge base so agents only see what each person can? With a GTM context graph
Updated
The short answer. To permission a company knowledge base so agents only see what each person can, build it as a GTM context graph, the foundation of AI for GTM, where each person's own sharing settings decide who sees every email, meeting and page, the most private setting wins when people disagree, and every agent works with the permissions of the one person it works for. Day AI gives your team that knowledge base, in place of your CRM or on top of Salesforce or HubSpot: a new person's email starts Private and every meeting type starts Invite only, and every Day AI agent inherits the permissions of the person it works for.
Say you're the CEO of a consumer-services company, and it's 7:40 on a Tuesday evening. You, your chief of staff and your executive assistant all work your inbox, and it's full of other CEOs, bankers and investors. The fear is a top-tier introduction that sits unanswered. Your evening recap from your Day AI agent arrives by email, and it surfaces two items that need you tonight, one of them a speaking invitation. Your agent can read all of it, private threads included, because it works for you. You ask it to draft a reply to the speaking invitation, change one line, and send it yourself, and the reply shows up on the inviter's timeline in Day AI. Both items are caught before you close the laptop.
Where the usual answer goes wrong. Most advice on this question starts with documents: index the wiki and the drives under one service account, copy each tool's access lists into the index, and filter answers at query time. For a revenue team, the most sensitive knowledge isn't in a document with an access list. It's in email and calls, where the only "permission" is who was on the thread, and a copy of access lists in an index is only as current as its last sync. Day AI's position is that every fact in the GTM context graph is permissioned for every person and every agent, decided by the person whose email or meeting it came from. Day AI provides the context graph and agent control plane to quickly and securely upgrade your company brain. If you've already kept your knowledge base locked down while you worked this out, that was the right start.
The six permission rules for a GTM context graph, with Day AI
A permissioned company knowledge base comes down to six rules. The first three decide who sees a fact, the next two decide what an agent sees, and the last one keeps the rules true as people change their minds.
- Start private, share on purpose. Nothing a person owns should reach the team until that person chooses to share it. In Day AI, a new person's email starts Private and every meeting type starts Invite only, so none of their email or meetings is shared with the workspace until they choose to share it.
- Let each person set the rules for their own sources. The person who owns an inbox knows which threads are sensitive. In Day AI, each person sets their own email sharing default and exceptions, and a privacy level for each kind of meeting (Step 2).
- When people disagree, the most private setting wins. One cautious person should protect the whole conversation. In Day AI, a private rule from any teammate keeps a message private, one private message keeps its whole thread private, and when meeting attendees' settings conflict, the most private one wins.
- Every agent works as one person. An agent should see what the person it works for can see, and nothing more. In Day AI, an agent inherits the permissions of the person it works for, and an app connected over Day AI's MCP server works as the person who signed in.
- Scope every automated job. A skill that runs on a schedule needs its own answer to "whose data?". In Day AI, every automated skill is scoped either to its owner's data or to what the workspace shares (Step 5).
- A change applies to what's already there. Tightening a rule should protect last month's mail, not only tomorrow's. In Day AI, changing a sharing rule also covers the mail already in Day AI.
How to permission your company knowledge base with Day AI
Each step below is a setting in Day AI or a request to your Day AI agent. Run Day AI in place of your CRM, with deals, people and companies kept current from your team's conversations, or on top of Salesforce or HubSpot; the permissions work the same way.
Step 1: Decide what flows into the shared memory in Day AI
Day AI has two kinds of sources, and they're permissioned differently. Each person connects their own Gmail and Google Calendar and the Day AI notetaker, and those stay theirs to share. Workspace integrations are set up once by a workspace Owner under Workspace Settings > Integrations, and they bring data into the team's shared memory: Slack from the channels you add Day AI to (public channels from a list, private ones with /invite @Day AI), and meeting tools such as Gong or Granola. Connect a workspace integration only for data the whole team should see. A new person's first email and calendar sync takes one to six hours.
Step 2: Have each person set their sharing rules in Day AI
Each person opens Email Sharing Rules in their user settings, picks share all or share none, and adds rules by domain, email address or Gmail label. The page states how rules apply: Gmail label rules override email address rules, which override domain rules; if no rule matches, the default applies; and if equally specific rules conflict, the email is not shared. In share-all mode, Day AI suggests domains you may want to exclude, such as legal or HR contacts, and you accept or dismiss each one. A rep can also hand their Day AI agent a list of domains or addresses to exclude, and it saves them and opens the rules so the rep can see them.
Then each person sets meeting privacy under Email & Meeting Settings: internal and external meetings, 1:1 and group, each Invite only or Workspace. A recording's privacy covers its summary, notes and transcript together. Contacts and organizations are always shared, so the team works from one list of people and companies while the conversations behind them stay with whoever owns them.
Step 3: Share pages, chats and views on purpose in Day AI
The knowledge base is more than mail and meetings. In Day AI, pages start Invite Only, and you share each one with the workspace, or by link, from its sharing button or by asking your agent ("Share this page with my workspace"). Each person's chats with their agent are private until they share one. Saved views are personal unless Share with workspace is ticked. Actions are shared with the workspace by default, and the owner can make any one private. A teammate who opens a link to a chat that wasn't shared with them sees "You don't have access to this chat."
Step 4: Give each person a Day AI agent, and connect Claude as that person
Every agent works for one person and starts from that person's permissions, so a rep's agent reads that rep's private mail and the team's shared memory, and nobody else's private mail. To reach the same knowledge base from Claude, Claude Code or Cursor, connect Day AI's MCP server (setup at day.ai/mcp). The authorization page asks which workspace and which of your agents to use, lists exactly what the app will be able to do, and the app works as you. Connected apps can't delete records.
Step 5: Set the data scope on every skill in Day AI
Skills are where an agent works without you watching, so give each one a scope. In the skill editor, the Private Data Access switch sets it: on, the skill can use its owner's private and shared email and recordings; off, it works only on what's shared with the workspace. A Managed skill an admin deploys to the whole team from the library uses workspace-shared data only, and any page attached to it must be shared with the workspace too. Keep a skill that summarizes one rep's deals on that rep's scope, and a team-wide skill, such as a Friday pipeline review, on the workspace's.
Step 6: Check it from a teammate's seat in Day AI
Before you tell the team the knowledge base is safe, look at it as someone else would. The Recordings table has a sortable Sharing column, so you can see at a glance which recordings are Workspace and which are Invite only. Charts and record tabs show only what the viewer may open: a teammate's private email never appears in your Momentum chart, and a teammate's Invite Only page doesn't show on a company's Related tab for you. Then have a teammate ask their own Day AI agent about a thread you kept private; it won't find it.
Where a homemade permission layer breaks, and what Day AI does
Most teams that build their own knowledge base for agents start with a folder, a vector index or a Claude project, and it works for one person. It breaks at the second user.
- One service account reads everything. The quickest way to index Gmail, Drive and Slack is one integration user, so every answer is as wide as that account's access and no answer knows who asked. In Day AI, there's no integration user: Claude, Claude Code or Cursor connect over Day AI's MCP server as the person who signed in, and the authorization page lists exactly what the app can do.
- Access lists copied into an index go stale. A person loses access in the source tool, and the index still answers from the old copy until the next sync. In Day AI, adding an exclusion makes that mail inaccessible to teammates and their agents, including the mail already in Day AI.
- Email has no access list. A shared thread has five people on it and no permission field, so homemade builds either share all of it or none of it. In Day AI, Gmail label rules override email address rules, which override domain rules, so one HR label keeps that mail private however the domain is set.
- The prompt becomes the permission. "Don't reveal HR email" in a system prompt is a request, not a rule, and one clever question gets past it. In Day AI, HR mail a rep can't open never reaches that rep's agent, so there's nothing for a clever question to pull out.
- Scheduled jobs run as nobody. A nightly script has no person behind it, so nobody can say whose private data it should touch. In Day AI, every scheduled skill has an owner and a switch that says whether it may use that owner's private data.