What Copilot Chat Actually Sees: Data Scope by Licence

What's new — June to August 2026

June 17, 2026 — June 16 GA wave: Copilot Cowork is now generally available worldwide, a new Copilot Credits usage-billing meter has gone live ($0.01 per credit, off by default), and the Work IQ APIs reached GA on June 16, 2026. None of this changes what Copilot Chat sees — your licence still determines the data and your permissions still determine the result — but it does change which surfaces draw a usage charge when they reach into your M365 data. The two new sections below ("Work IQ — the context layer underneath the data scope" and "When access to your data is metered") explain how that fits the data-scope picture in this guide.

July 2026 governance additions: Purview controls now cover Cowork interactions, including sensitivity label inheritance, audit logging, DSPM activity, Insider Risk Management, eDiscovery, and Communication Compliance, all rolled out in June. Copilot-generated files automatically inherit the highest sensitivity label found in their source data, and users are notified if a label cannot be applied. A Purview DLP control to exclude external-sender emails from Copilot grounding is in public preview, with general availability targeted for July. Microsoft is also extending Purview DLP for M365 Copilot into government clouds, letting admins block Copilot and agents from returning a response when a prompt contains sensitive data (Rolling Out, Roadmap 561892).

Mid-July 2026 additions: Content sources in Copilot Chat is rolling out as of 13 July 2026: users can select which content sources ground a response, and answers stay within the chosen sources. The roadmap lists GA for July 2026 (Roadmap 496596). Purview eDiscovery enhancements for Loop and Copilot Pages reached GA on 13 July 2026: eDiscovery now fully indexes Loop and Copilot Pages content for keyword search in review sets, and adds HTML export from search (Roadmap 561492). Purview Data Lifecycle Management is adding insights on Copilot and AI app usage with recommended retention policies for governing those interactions; the roadmap lists preview for August 2026 and GA for September 2026 (Roadmap 561209, status as of 14 July 2026).

Late July 2026 additions: The Microsoft Purview DLP policy that restricts Microsoft 365 Copilot from processing sensitivity-labelled files in Word, Excel, and PowerPoint now applies regardless of where the file is stored (Roadmap 557255, GA). In preview, Purview DLP is being extended to restrict Microsoft 365 Copilot and Copilot Chat from processing emails from senders external to the organization (Roadmap 561552), and a separate in-development control will safeguard web searches that contain sensitive data (Roadmap 565870). Purview auto-labelling capacity for SharePoint and OneDrive is rising from 100,000 to 500,000 files per tenant per day (Roadmap 567890, preview), and Data Lifecycle Management retention-based archiving can move inactive OneDrive and SharePoint files to lower-cost storage while keeping them discoverable to Copilot (Roadmap 561208, preview). With a Microsoft 365 Copilot licence, users can also capture offline discussions as searchable voice notes in the Copilot mobile app (Roadmap 498640, preview).

Early August 2026: the web-search control noted above as in development is now rolling out. Purview DLP for external web search lets a policy stop Microsoft 365 Copilot and Copilot Chat putting sensitive data into a web query, and Microsoft describes the control as operating in real time (Roadmap 565870). Two DLP diagnostics are in preview: DLP Policy Change Insights with Security Copilot, which produces a plain-language explanation of what changed in a policy after each update by comparing the version before and after and reporting what changed, where, the impact, and the enforcement implications (Roadmap 559106), and a guided experience to diagnose and resolve DLP issues, which summarises why a DLP action was taken on a document, which policies were triggered, the order of evaluation, and which conditions evaluated as true or false (Roadmap 561032). Both are aimed at the same recurring problem: a policy blocks something and nobody can say which rule did it. In the other direction, domain exclusion for web grounding has been rolled back. Microsoft's 4 August post says the feature is withdrawn while it evaluates next steps (Roadmap 503144), so blocking named domains from Copilot web results is not an available control right now.

Mid-August 2026: four changes to what Copilot can reach and how quickly it reaches it. Vision in Microsoft 365 Copilot can analyse content shared on a desktop screen and, on mobile during a real-time voice session, a live feed from the device camera, answering questions about what it sees in spoken form and grounded in work and web data; the roadmap entry reads generally available against a June 2026 target date (Roadmap 561037). Treat that as a new input surface rather than a new data source: what reaches Copilot is whatever a user puts on screen or in front of the camera, which sits outside the document permission model the rest of this guide describes. SharePoint Authoritative Sites lets administrators designate specific sites as official, trusted sources, and Microsoft states that classifying a site as authoritative causes its content, such as company news, policies and updates, to be prioritised across Copilot Search experiences (GA). Copilot connectors now run content crawl and identity crawl in parallel rather than sequentially, so ingested content reaches users sooner (GA). And ServiceNow Knowledge and Catalog connectors enforce permissions based on user roles such as admin, knowledge manager and knowledge admin, where permissions were previously determined by user criteria alone without considering roles (GA). Those last three come from Microsoft's 11 August release-notes batch, where general availability means released to Current Channel and still rolling out inside a tenant.

Licence required before you start

The features in this guide require the Microsoft 365 Copilot paid add-on. They are not available on standard M365 plans (E3/E5) or free Copilot Chat (except Copilot Chat within Outlook, which retains inbox and calendar grounding for all users). Verify your licence under Settings → Subscriptions or contact your Microsoft admin or check your organisation's licence portal.

What each tier can access

Data source	Free (E3/E5 included)	M365 Copilot paid add-on
Outlook email (received)	Copilot Chat in Outlook only · inbox grounding · standalone Chat has no access	Up to 10 results per query · recency-weighted
Outlook calendar	Copilot Chat in Outlook only · calendar grounding · standalone Chat has no access	Yes — meetings, invites, attendees
Teams chat messages	No access	Last 30 days · current thread only (in Teams app)
Teams channels	No access	Channels you are a member of only
OneDrive files	No access unless you upload the file manually	Yes — files you have access to · @mention to reference explicitly
SharePoint documents	No access	Yes — sites and files your permissions allow
Meeting transcripts (Teams)	No access	Yes — meetings you attended where transcription was enabled
Web (Bing search)	Yes — primary data source	Yes — Web/Work toggle controls which is queried
Files uploaded to chat	Yes	Yes
Copilot Notebooks	No access	Yes — grounded in references you pinned
SAP · Primavera · external systems	No access	No access without a custom connector. Federated connectors for third-party apps (Canva, HubSpot, Notion, Linear) in Public Preview as of February 2026.

How to confirm which tier you have: open Copilot Chat. If you see a Web / Work toggle in the interface, you have the paid add-on. If you see web results only with no toggle, you are on the free tier.

How the permission model works

Your identity, your data

Copilot queries Microsoft Graph using your identity. It can only retrieve content you already have permission to access. No new access is created. If you cannot open a SharePoint file, Copilot cannot see it either.

Permissions are enforced at retrieval

The semantic index that powers Copilot retrieval enforces your access controls at query time, not at index time. Content is filtered before it reaches the model — the model never "sees" data it could not surface to you.

Sensitivity labels are respected

Emails and files encrypted with Purview sensitivity labels are subject to usage rights. If your rights do not include content extraction, Copilot will not surface that content even if you can open the file.

Overshared content is visible

Copilot surfaces everything your permissions allow — including files shared too broadly. If a SharePoint site grants view access to everyone in the organisation, Copilot can retrieve its content for any user. This is a governance issue, not a Copilot issue.

Work IQ — the context layer underneath the data scope

🧠

What Work IQ is

Everything described in the scope table above — retrieving your email, your Teams threads, your SharePoint files through your permissions — runs on Microsoft's context layer for M365 Copilot, now branded Work IQ. Work IQ has three parts: Data (the permission-trimmed retrieval over Microsoft Graph), Memory (what Copilot retains across sessions), and Inference (how it reasons over what it retrieved). It is the same mechanism this guide has always described — the permission model and recency windows below all live inside Work IQ Data.

🔌

The Work IQ APIs reached GA on June 16, 2026

As of June 16, 2026, the Work IQ APIs are generally available (A2A and remote MCP are GA; a REST API is "coming soon"). This lets custom and third-party agents ground in the same permission-trimmed M365 context that Copilot Chat uses — they see only what the signed-in user can see, exactly as described in the permission model above. The data-scope rules in this guide therefore apply to those agents too, not just to Copilot Chat.

When access to your data is metered

The scope table answers what Copilot can see. The June 16 GA wave adds a second question for some surfaces: does reaching into that data draw a usage charge? Microsoft's new consumption meter is Copilot Credits ($0.01 per credit on pay-as-you-go; commit-for-discount tiers also exist). Credits are off by default — admins switch them on and set spend limits and usage alerts at tenant, group, or user level, with reporting in a new Cost Management dashboard in the M365 admin center.

✓

Pre-built Microsoft Copilot agents — not separately metered

Microsoft's own pre-built Copilot agents, used under an existing M365 Copilot licence, are not separately metered. Your $30/user/month add-on remains the access gate, and using those built-in agents over your M365 data does not draw Credits on top.

Custom & third-party agents via Work IQ — metered

When a custom or third-party agent (built in Copilot Studio, Microsoft Foundry, or directly on the Work IQ APIs) accesses your M365 data through Work IQ, that usage is metered in Copilot Credits. A task's cost reflects model use, context retrieval, tool calls, and runtime. The licence is still the gate; the Credits meter applies on top.

Unlicensed Copilot Chat — a metered surface

Unlicensed Copilot Chat draws Copilot Credits — it is one of the surfaces the meter covers. This does not widen what it can see: the free tier remains web-grounded (plus inbox/calendar grounding inside Outlook), exactly as the scope table shows. Metering governs cost, not data scope.

Scope first, cost second

Credits never grant access your permissions don't already allow. Work IQ still trims every retrieval to what the signed-in user can see. Read the scope table to know what an answer can contain; read this section to know what it costs to produce.

Sources: Copilot Cowork is now generally available · Announcing the new Work IQ APIs · Usage-based billing with Copilot Credits

Why two people get different answers to the same prompt

📧

Different email and Teams histories

Each user's Copilot query runs against their own inbox, their own Teams threads, their own calendar. If User A is on a project email chain and User B is not, they return different results from identical prompts. This is correct behaviour — the answers reflect reality for each person.

🔒

Different SharePoint and OneDrive permissions

A document visible to a project manager may not be visible to a team member not added to that SharePoint site. Copilot returns what each user can see. Two people asking "what does the project charter say" get different answers if only one of them has access to the charter.

⏱

Different recency windows

Copilot prioritises recent content. If User A received a relevant email yesterday and User B received one three months ago, the same prompt surfaces different content — not because the data is missing, but because recency scoring ranks it differently. The hard email cap of 10 results per query amplifies this effect.

Why Copilot ignored that message

Six specific reasons — with what to do about each

Most common

You are using standalone Copilot Chat on the free tier

Standalone Copilot Chat (at m365.cloud.microsoft or in Office apps) accesses web data only on the free tier. It cannot see Teams messages, SharePoint, or OneDrive. Exception: Copilot Chat within Outlook provides inbox and calendar grounding for all M365 users, even without the paid add-on — but only when accessed inside Outlook itself.

What to do

For email queries, open Copilot Chat inside Outlook — it has inbox access without the paid add-on. For Teams, SharePoint, and cross-app queries, the paid M365 Copilot add-on is required.

Hard limit

The email cap returned other messages first

Copilot returns a maximum of 10 emails per query, ranked by relevance and recency. If your query matches 50 emails, 40 are silently dropped. An important but older message may never appear — not because Copilot cannot access it, but because it ranked below the cut-off. This limit is not displayed in the interface.

What to do

Narrow the prompt: add a specific sender, date range, or subject keyword to push the relevant email into the top 10. "Emails from [name] in March about [topic]" outperforms a broad summary request.

Common

The Teams message is older than 30 days

When using Copilot inside Teams, the default search window for chat messages is the last 30 days of the current thread. Messages outside that window, or from other chats and channels, are not searched unless you specify a date range explicitly in the prompt.

What to do

Include the date range in the prompt: "from December 2025" or "in Q4 last year." Copilot can search further back when instructed explicitly.

Common

You are not a member of that Teams channel

Copilot only accesses Teams channels and group chats you are a member of. A message posted in a project channel you were not added to is invisible — even if colleagues tell you it exists. Private channel messages are also restricted to members of that private channel.

What to do

Ask a channel member to forward the message or share the file directly with you. Once it is in your scope, Copilot can reference it.

Frequent in large teams

The file is on SharePoint but you lack access

An email may reference a SharePoint document, but if you were not granted access to that document or site, Copilot cannot retrieve its content. The email itself may appear in results, but the document it references remains inaccessible. Copilot cannot escalate permissions or prompt you to request access.

What to do

Request SharePoint access through the site owner. Once granted, Copilot will include the document in future queries. No re-indexing required — access is checked at query time.

Less common

The content has a sensitivity label that restricts extraction

Files and emails classified with Purview sensitivity labels may have usage rights that block content extraction. Copilot respects these rights — it will not surface the content of a document marked "Confidential — No Copy" even if you can open it in Word. The file exists in your scope; the content does not reach the model.

What to do

Check the sensitivity label on the document. If the label is the constraint, contact your Information Protection admin — the label configuration, not Copilot, is the control point.

Common false assumptions

Assumption

The Copilot Chat included in my M365 E3 or E5 licence searches my work emails and Teams messages.

Reality

Partially. Standalone Copilot Chat (free tier) is web-only — no Teams, SharePoint, or OneDrive access. However, as of early 2026, Copilot Chat within Outlook provides inbox and calendar grounding for all M365 users without the paid add-on. The paid add-on is still required for Teams data, SharePoint, and the full Copilot feature set.

Assumption

Copilot searches my entire email history. If I ask about a project from two years ago, it will find the relevant emails.

Reality

Copilot returns at most 10 email results per query, ranked by recency and relevance. Older emails may rank below the cut-off. Narrow the query with a sender name, date, or subject keyword to improve retrieval.

Assumption

Copilot in Teams searches all my Teams conversations — channels, group chats, and direct messages.

Reality

Within the Teams app, Copilot searches the current chat thread only, defaulting to the last 30 days. To search across channels and messages, use Copilot Chat at m365.cloud.microsoft/chat with the Work mode active.

Assumption

My prompts and Copilot's responses are used to train Microsoft's AI models.

Reality

No. Microsoft explicitly states that prompts and responses from M365 Copilot are not used to train foundation models, including Microsoft's own. Your organisational data does not leave your M365 tenant to train AI.

Assumption

Copilot can access any SharePoint content in my organisation — including sites I have not visited.

Reality

Copilot accesses only what your permissions allow. If you do not have access to a SharePoint site, Copilot cannot retrieve content from it. Overshared sites are the exception — if a site grants everyone access, all users can retrieve its content via Copilot.

Copilot Chat: What It Actually Sees

What each tier can access

How the permission model works

Work IQ — the context layer underneath the data scope

When access to your data is metered

Why two people get different answers to the same prompt

Common false assumptions