The short answer
Most serious enterprise search tools respect source permissions, but they do it in three different ways, and only one of them is safe to switch on for the whole company in week one. Microsoft 365 Copilot and Search show only what the signed in person can already open in Microsoft 365. Glean and Gemini Enterprise connect to other apps and check each source's access rules for the person asking. InSearch does that for Google Drive, Slack and Gmail and lets the admin pick everyone or admin only for its other connectors. Elasticsearch gives you the parts, document level security on roles, and leaves the mapping to your engineers. The only way to know is to plant a restricted file and search for it with a second account, which takes about 20 minutes and is laid out below.
Vendor statements are from Microsoft Learn and Elastic documentation, read November 19, 2026. The test and the table verdicts are our own judgment.
If you are shortlisting search for a company of 50 to 500 people, "does it respect permissions" is the question that decides whether security signs off. Every vendor answers yes. The useful question is how, because the mechanism decides what breaks when someone changes teams, a Slack channel goes private or a Drive folder gets reshared.
We sell InSearch, so we have a side. We keep the comparison to what each vendor documents, and we say where each one is the better pick.
Which enterprise search tools respect permissions?
All of the tools below claim to. They differ in where permissions come from and when they are checked.
| Tool | Where permissions come from | When they are checked | Best fit | Watch out for |
|---|---|---|---|---|
| Microsoft 365 Copilot and Copilot Search | Microsoft 365 itself, SharePoint, OneDrive, Exchange, Teams | At query time, against the user's identity in Microsoft Graph | Companies whose work lives almost entirely in Microsoft 365 | Microsoft says Copilot only surfaces data a user has at least view access to, so oversharing in SharePoint becomes oversharing in answers |
| Glean | Each connected app's own access rules | Vendor states it enforces them per user at query time | Large organizations with many apps and a budget for a quote based contract | No public price list, see Glean pricing |
| Gemini Enterprise | Google Workspace plus connected sources through identity sync | Per user, depends on connector and identity setup | Google Workspace companies | Third party sources need identity sync, covered in Gemini Enterprise permissions and identity sync |
| Elasticsearch / Elastic Cloud | Roles you define on indices, with document level and field level security | At query time, once you have indexed the source ACLs yourself | Engineering teams that want to own the build | Connectors, ACL mapping and sync are your work, see Elastic Enterprise Search pricing |
| InSearch | Drive, Slack and Gmail item by item. SharePoint, Notion, Confluence, Jira and Salesforce are shared with everyone in the workspace or only the admin, chosen at connection | Per person at query time, against access read at the last sync a few minutes earlier | Teams of 20 to 500 that want a finished product without a build | Eight connector types only, no iManage or NetDocuments, see InSearch pricing |
The Microsoft statement is the plainest one in the market: Copilot only surfaces organizational data to which individual users have at least view permissions, and it uses the same underlying access controls as other Microsoft 365 services (Microsoft Learn, Data, Privacy, and Security for Microsoft Copilot). That is a strength and a warning. The tool is not leaking, your sharing settings were already loose.
Elastic documents document level security as a role setting that restricts which documents a user can read (Elastic docs, controlling access at the document and field level). It is a real mechanism and it is only half of the job. Something still has to read the ACL from Slack or Drive for every file, store it next to the content, and keep it fresh.
What does "respects permissions" actually have to mean?
Three properties separate a tool that is safe from one that only says so.
- Inherited, not rebuilt. The tool reads the source app's access rules. If it asks an admin to recreate groups and folders inside the search product, you now maintain two permission systems and the second one is always stale.
- Checked per person at query time. A nightly snapshot of who could see what leaves a gap of up to a day after someone leaves a channel or a contractor is removed.
- Applied to AI answers, not only to the result list. A restricted document must not be quoted in a generated answer or in the citations under it. Ask the vendor directly, because some products filter links but summarize from a wider set.
Our own page on permission-aware AI search goes through the mechanism in more depth, and the permission-aware enterprise search article explains why item level control is what lets a pilot become a company rollout.
The 20 minute permission test to run in any trial
Do this before you sign. It works on every product in the table and it is the one thing no vendor page can tell you.
- Plant three canaries. Create one Google Doc or SharePoint file shared with only you, one Slack private channel message with a unique made up word, and one Confluence page restricted to a single space. Use something that would hurt if it leaked, such as a fake "2027 compensation bands" sheet. Finance material is the best canary because it is what people most fear exposing. A finance team that chases invoices in a tool like accountsreceivable.ai still keeps its payment notes and approvals in Drive and Slack, and those are the files to plant.
- Search as yourself. Confirm all three canaries appear. If they do not, the connector is not indexing them and the test proves nothing yet.
- Search as a second account. Sign in as an ordinary employee with no access. Search the unique word, then ask a natural question that the canary answers. Neither the result list nor the AI answer should mention it.
- Change access and search again. Share the Doc with the second account, wait for the tool's sync, and confirm it now appears. Remove access and confirm it disappears. Record how long each change took. This is the number to put in front of your security team.
- Remove the user from a group. Take the second account out of a Slack channel or a Google group that granted access. A product that checks identity live should drop the content at once. A product that caches ACLs will keep showing it until the next sync.
- Ask the AI to guess. Type "what is the salary band for senior engineers" as the second account. A safe tool says it found no accessible source. An unsafe one answers from the canary or hints at it.
Score each tool out of six. Anything below six of six is not ready for a whole company, whatever the sales deck says. The same thinking applies to AI products that are not search tools at all: the ChatGPT Enterprise connectors article covers what a connector inherits and what it does not.
What breaks permissions in practice
In the projects we see stall, the cause is rarely a bug in the search tool. It is one of four things.
- Link sharing. "Anyone in the company with the link" files are readable by everyone, so search returns them to everyone. A permission-aware tool is correct here, and the fix is a sharing cleanup.
- Shared inboxes and mailboxes. Gmail and Exchange permissions are per mailbox. Indexing a mailbox for the wrong people is a bigger risk than any document.
- Identity mismatch. The person is [email protected] in Google and jane.doe in Slack. Tools that cannot match identities across apps either hide too much or show too much.
- Ethical walls and legal holds. Screened matters and deal rooms often rely on folder level restrictions that a coarse tool flattens. Law firms should read knowledge management for law firms for how to test this.
Which one should you pick?
If you live in Microsoft 365, run the sharing cleanup and pilot Copilot Search first, because nothing else matches its native identity. If you are on Google Workspace, price Gemini Enterprise against the alternatives. If you have engineers and want control, Elastic is a sound base and you should budget the connector and ACL work honestly. If you are a 20 to 500 person company that uses a mix of Drive, Slack, Notion, Confluence, Jira and Gmail and wants permission-aware cited answers working this week, that is the case InSearch is built for. Compare it against the field on enterprise search tools or start from pricing, where Team is $20 per user a month billed annually and Business, with single sign on, is $32.
Permission questions buyers ask
Which enterprise search tools respect permissions?
Microsoft 365 Copilot and Search, Glean, Gemini Enterprise and Elasticsearch with document level security all state that they respect source permissions, and InSearch does so for Google Drive, Slack and Gmail. They differ in who does the mapping and when access is checked, so run the canary test in a trial before rollout rather than relying on the claim.
Can AI search respect access controls?
Yes, when the tool inherits each source's access rules and checks them per person at query time, for the answer as well as the result list. Microsoft states Copilot only surfaces data a user already has view access to. Test it with a restricted file and a second account.
Does permission-aware search prevent unauthorized access?
It prevents the search tool from becoming a new way around your access rules, but it does not fix bad rules. Files shared with the whole company through a link will still appear for everyone. Clean up link sharing first, then search is a safe layer on top.
How fast do permission changes show up in search?
It depends on the product. Live identity checks apply a removed permission immediately. Products that cache access lists apply it at the next sync, which can be hours. Measure the delay in your own trial with the removal step above and ask the vendor for the documented figure.