Claude Connectors and Enterprise Search — Connecting Google Drive, Slack and Microsoft 365

How a connector differs from uploading files

There are two fundamentally different routes for getting organisational data in front of Claude, and picking the wrong one is the most common reason enterprise AI pilots stall.

Aspect	Uploaded files / project knowledge	Connector
Freshness	Frozen at the moment of upload	Read from the source at query time, so always current
Access control	Anyone in the room sees every file in the room	Follows each user's existing permissions in the source system
Maintenance burden	Someone has to keep replacing outdated files	None — no copy is made
Best for	Regulations, manuals and templates that rarely change	Email, working files and chat that change daily

One sentence to remember: static documents belong in project knowledge; moving data belongs behind a connector. Exporting a report from a system and re-uploading it every week is doing the machine's job by hand.

What can be connected, and what it can actually do

The list below follows Anthropic's own documentation (checked 1 September 2026). Note that several connectors do more than read — they can draft and send email or create calendar events.

System	What it can do	Documented limits
Google Drive	Search documents, read files, upload, share and save files back	Cannot see images, comments or suggestions inside Google Docs
Gmail	Search, read, draft, send, reply, forward and manage labels	Attachment content is inaccessible — metadata only
Google Calendar	View, create, update and delete events; manage attendees	Complex queries may require multiple API calls
Microsoft 365	Search across SharePoint, OneDrive, Outlook and Teams	An admin must grant Microsoft Entra consent first
Slack	Serves as the "chat" source for Enterprise Search	Only channels the individual user already belongs to
Custom / remote MCP	Connect in-house systems through MCP	The server must be reachable over the public internet, because the connection originates from Anthropic's cloud rather than the user's machine

Enterprise Search: one question across every system

Enterprise Search is a pre-built project called "Ask Your Org" that appears in the sidebar with optimised instructions already in place and your organisation's connectors wired in. One question fans out across documents, chat and email simultaneously.

It is available on Team and Enterprise plans only, and setup requires connectors in at least two categories.

Category	Required?	Examples
Documents	Required	Google Drive or SharePoint
Chat	Required	Slack or Microsoft Teams
Email	Optional but recommended	Gmail or Outlook

Setup always happens in two layers: an owner enables it for the organisation, then each individual authenticates with their own account. That second layer is precisely the mechanism that keeps permissions intact.

Step	Who	Action
1	Organisation owner	Organization settings > Connectors > Browse connectors, then "Add to your team"
2	Organisation owner	Click "Ask Your Org" > "Set up for your org" > pick the documents and chat connectors > name the project and add a description
3	Each user	Authenticate to each service with their own credentials before searching
To disable	Organisation owner	Admin settings > Capabilities turns the feature off organisation-wide

Permissions: read the whole sentence

Anthropic's documentation is explicit: "You only see search results from data you have permission to access in the original systems", and it states that no external indexing of connected data takes place. That is an accurate selling point and genuinely good news for a security team.

A second point for the legal team: connectors reach third-party services that operate under their own terms and may process data outside the country. Organisations with PDPA or similar obligations should scope explicitly which connectors are enabled and over which data sets.

Limits worth knowing before you promise anything

Limit	Practical consequence
Enterprise Search is not available on mobile apps	Web and Claude Desktop only — do not plan the executive demo on a phone
Queries count toward standard plan usage limits	Heavy search teams hit limits sooner; factor this into seat planning
Gmail attachment content is not readable	A supplier quote sent as a PDF attachment still has to be opened manually
Google Docs comments and images are invisible	Documents where the decision lives in the comment thread summarise incompletely
Source-provider API quotas apply	Very large mailboxes search more slowly — expected behaviour, not a fault

How to ask so Enterprise Search answers well

"Ask Your Org" is not a keyword box. It takes questions as sentences and fans them out across systems, so questions that pin down a time window and demand provenance perform noticeably better than bare keywords.

A question that works: "Summarise what we agreed with this supplier about warranty terms, based on email and chat from the last three months, and state which message or file each point came from." That uses all three strengths at once — cross-system reach, a bounded time window, and a requirement to cite the source.

A question that disappoints: "What policies do we have?" The system has to guess where to look, and the answer comes back broad and incomplete. The principle is the same one described in the guide to prompting Claude: the tighter the scope, the more verifiable the result.

What about the ERP?

To be direct: Saeree ERP does not yet publish an official MCP server for Claude to call, and we would not recommend rushing to let an AI read an ERP database directly anyway. The reason is scope rather than technology — an ERP holds payroll, cost prices and approval limits, and opening a broad read channel in round one creates work you will have to undo later.

The sequence we recommend instead: (1) connect the Drive or SharePoint where working documents already live, (2) measure which questions people genuinely ask, (3) only then decide which specific reports from the ERP system to expose through a permission-aware gateway — not the whole database.

The dividing line holds: the ERP stays the system of record with an audit trail; Claude is the search and drafting layer above it, never the place transactions are recorded.

Good fits and poor fits today

Good fit	Poor fit
"Where is the meeting summary for project A, and who edited it last?"	"What is the remaining budget on project A right now?"
Pulling scattered email, chat and drive context into one summary	Work requiring 100% completeness, such as statutory audit
New joiners finding documents they do not know exist	Organisations that have never audited their file-sharing permissions
Drafting a customer reply grounded in the existing thread	Approvals or postings that need an auditable trail

Conclusion

Connectors change the working model from "copy data out so the AI can see it" to "let the AI look where the data already lives" — the right direction for both freshness and security. Enterprise Search bundles those channels into a single question box, requires at least a documents and a chat connector, and is limited to Team and above.

The work that decides whether this succeeds is not on Anthropic's side. It is housekeeping on yours: how tidy your file-sharing permissions are. Enterprise Search grants nobody new access — it simply turns "unfindable because nobody knew where it was" into "found in three seconds".

Enterprise Search does not grant anyone new access. It reveals how loose the access you granted three years ago really was.

- The Saeree ERP team

References

• Anthropic — Use enterprise search (checked 1 Sep 2026)
• Anthropic — Use connectors to extend Claude's capabilities
• Anthropic — Use Google Workspace connectors
• Anthropic — Set up the Microsoft 365 connector