Compliance officer reviewing an access policy binder in a quiet office

Security · Permission-aware search

Permission-aware AI search that respects every access control

The number one blocker to company-wide AI search is the fear that it will leak something. The moment a search tool can surface any document, someone worries it will show an employee the salary spreadsheet or the unreleased roadmap. That fear is what keeps powerful search confined to a pilot team.

or ask a sample question ↓

One search · cited answers · permission-aware

Answer Console

Your files only

Drop your documents here PDF, DOCX, TXT, MD, CSV or a photo of a page. Up to 3 files.

Reading your files…

That request did not get through. Tap Ask InSearch again.

No account needed. Your files are read once to answer, then deleted, and never used to train any model.

Short answer

The enterprise search tools that respect permissions are the ones that check each connected app's own access rights at the moment a question is asked, per person, rather than once at indexing time. InSearch, Glean, Coveo, Slack enterprise search and Microsoft 365 Copilot all enforce source permissions; they differ in which sources they cover, whether an admin or each user connects them, and what happens to items with unusual sharing. The test that separates them is not the checkbox on the security questionnaire but the revocation window: remove someone's access to a document, ask the question again, and see whether the answer changes immediately or after the next crawl.

InSearch is permission-aware by design, and that is the unlock. It inherits the native, item-level permissions of every connected source and enforces them at query time, so each person only ever gets answers built from content they are already authorized to open. It creates no new permissions and grants no new access. Combined with no training on your data and encryption under SOC 2 practices, that is what makes a safe, company-wide rollout possible.

Permissions are the part security teams probe hardest, and our deeper write-up on permission-aware enterprise search covers how item-level access is enforced at query time. It is the foundation of secure enterprise AI search, and how AI enterprise search works shows where the check sits in the pipeline.

Drive Slack Notion Confluence Jira Salesforce

Cited to source docs

Permission-aware · never trains on your data

Why it works

What your team gets with Permission-aware search

Inherits every permission

It uses the native, item-level access controls of each source, so nothing new is exposed.

Enforced at query time

Permissions are checked as the answer is built, so each person only sees what they can already open.

The rollout unlock

Because it leaks nothing, a company-wide deployment is finally safe, not a permanent pilot.

What it handles

One search, a cited answer, scoped to you

InSearch searches across every connected app, checks your permissions, and writes one clear answer with citations back to the exact source documents.

Inherits each source's native access controls
Enforces permissions at query time, per person
Creates no new permissions or access
Makes company-wide rollout safe
Never trains on your data, encrypted in transit and at rest

Answer Cited

Written answer

Full-time employees get 20 days of PTO per year plus 10 company holidays¹, requested in Workday with manager approval².

Only sources you can access

CF People Handbook · Time Off [1]

DR PTO-request-process.pdf [2]

Drive · Slack · Confluence Never trains on your data

Why InSearch

One search across every app, with a cited answer

Not eight separate search boxes, not a wall of blue links. InSearch unifies your apps and returns one written answer with citations, scoped to exactly what you can see.

One search, every app

Drive, Slack, Notion, Confluence, Gmail, Jira and Salesforce searched in a single query, so you stop opening a different search box for each tool.

A cited answer

A written answer in plain language with inline citations back to the exact source docs, so you can trust it and verify it instead of reading a page of links.

Permission-aware

It inherits and enforces each source's access controls at query time, so every person only ever sees answers from content they can already open. It never trains on your data.

Which enterprise search tools respect permissions?

Most serious enterprise search products now respect permissions in some form, so the useful question is how, not whether. The table below sets out the published permission model for the tools buyers shortlist most often, with the caveat each vendor documents about its own approach. Every row was checked against the vendor's own documentation within the last six weeks, and the caveats are the details that decide security reviews.

Tool	How permissions are enforced	Published caveat
InSearch	Item-level permissions inherited from every connected source and evaluated per person at query time. Sources are connected once by an admin; each answer cites only documents the asker can open	Read-only connector scopes. No new access is ever created
Glean	Crawls each connected app with an administrator credential and builds one permission-aware index across the company	Sold by quote with seat minimums; the index is only as current as the last crawl
Microsoft 365 Copilot and SharePoint agents	Enforce Microsoft Graph permissions; a SharePoint agent answers only from what the user can open	Microsoft documents that an item shared with more than 10,000 distinct users or security groups is not searchable by anyone. Non-Microsoft content is outside the Graph
Slack enterprise search	Query-time permission handling across its connectors, which is genuinely good	Available only on Enterprise+ at $45 per user per month billed annually, disabled by default, 13 connectors with per-connector caps such as five GitHub organizations
ChatGPT Business and Enterprise	OpenAI states a member can retrieve only information they are already allowed to access, and that the Company Knowledge plugin does not override app permissions	Outside admin-managed Google Drive, SharePoint and Teams, each person authorizes their own apps, so coverage is per user and opt-in
Coveo	Document-level security on both the Pro and Enterprise plans, per Coveo's product entitlements	Cloud connectors are limited to three on Pro; SAP, Salesforce, ServiceNow and Shopify connectors are sold separately
Onyx	Permission-syncing connectors mirror source access into the index	Onyx documents permission syncing as an Enterprise Edition feature, so the open source build does not include it

Two patterns fall out of the table. Tools that index centrally with an admin credential (InSearch, Glean, Coveo) can answer for everyone from day one and have to get revocation right. Tools that connect per user (ChatGPT, and Copilot outside the Graph) get permissions right almost for free, because they only ever see what that person connected, and pay for it in coverage: nobody can find what nobody connected. The right choice depends on which failure you would rather explain.

Which AI assistant enforces document permissions?

Every major work assistant enforces document permissions for the sources it can reach, but they reach different sources in different ways. Microsoft 365 Copilot enforces Graph permissions across SharePoint, OneDrive, Teams and Exchange. ChatGPT Business and Enterprise enforce the permissions of each app a member authorizes, plus admin-managed Drive, SharePoint and Teams. Claude Enterprise connectors are enabled org-wide by an admin and then authorized by each user, so the same per-person rule applies.

The gap in all three is the same: they are assistants, not an organization-wide index. An assistant enforces permissions on what it was given; it cannot answer from a Confluence page or a Slack thread that the person asking never connected, even when that person has every right to read it. That is the case for a dedicated permission-aware search layer under the assistant, and it is what InSearch does: $20 per seat per month, every source connected once, every answer checked against the asker's access at query time and cited back to the document it came from.

How to test a vendor's permission claims in an hour

Three tests, run in a trial workspace with two accounts, tell you more than any questionnaire. Each takes minutes and each has caught a vendor that passed the paperwork.

The revocation test. Share a document with account B, confirm B can find it through the search tool, then remove B's access at the source. Ask the same question from B again. Query-time enforcement changes the answer immediately. Crawl-time enforcement keeps returning the document until the next sync, and the vendor should be able to tell you how long that is.

The private channel test. Post a distinctive phrase in a Slack channel or Teams chat that only account A belongs to. Search for it from B. Anything other than no result is a finding, and it is worth repeating for a private Confluence space and a Drive folder shared with a group B is not in.

The link-sharing test. Set a document to anyone with the link and search for it from an account that was never sent the link. Vendors treat this case differently and most do not document which way, so ask, and check that the answer matches what your own information security policy expects. The secure enterprise search page covers the rest of the review, including encryption, data residency and the training question.

Good questions

Questions about Permission-aware search

It means InSearch enforces the exact access controls of every connected source. If you cannot open a document in Drive or a channel in Slack today, InSearch will never use it to answer your question. Each answer is built only from what you are already allowed to see.

No. InSearch grants no new permissions and changes nothing about who can access what. It simply respects the access that already exists, which is what makes a broad rollout safe.

Not noticeably. The access check runs as the candidate documents are gathered, before the answer is written, and it filters a short list rather than the whole index. The alternative, checking access once at crawl time, is faster on paper and wrong the moment a permission changes.

It depends on the source app and the vendor, and most vendors do not document the case. A link-shared file is technically open to anyone holding the link, which is not the same as everyone in the company. Test it with a second account during the trial and confirm the behavior matches your information security policy before rollout.

Read every enterprise search question

Explore more

Give your team one place to find every answer.

One search across all your company apps, a clear cited answer, scoped to exactly what each person can see, and never trained on your data.

See pricing

One search across every app · cited answers · permission-aware · never trains on your data