Glean and the workplace AI search problem

Ask where the latest pricing policy lives and you may get four answers: a wiki page, a shared slide deck, a pinned chat message and a ticket comment. Each sits in a different app with its own rules about who may see what.

That is Glean's problem to solve. The company says it was founded in 2019 to build enterprise search, then added an assistant and, in 2025, AI agents.[1] Its pitch: it supplies the company context AI needs, with each user's access rights enforced.

This profile covers what Glean makes, how it relates to retrieval-augmented generation, how it charges and what to weigh before connecting it to company data. It draws on Glean's documentation and announcements, with independent reporting for the financial figures. We have not tested the product.

The problem: knowledge in many apps, each with its own rules

Most organisations have a document store, a chat tool, a ticketing system, a CRM and an HR system, not one knowledge base. A general chatbot cannot see any of them.

The hard part is access. A support agent should find the runbook but not the pay-review folder. A system that pulls every app into one place must also pull in who may see each item, keep that current and apply it to every answer. Otherwise an AI assistant becomes a fast way to read documents you were never meant to see.

What Glean makes

Connectors. Each connector links Glean to one app and, its documentation says, fetches content and "the permissions map from each source".[2] Glean's site claims more than 250 connectors;[1] other sources can push data in through an Indexing API.[2]

An index and a graph. Content, metadata, identities and access rules go into a search index in the customer's own tenant, with a graph of relationships between people, documents and projects on top. Some connectors query the source live instead, or mix both methods.[3]

Permission-aware retrieval. Glean's security documentation describes "permission mirroring": the connector copies each source's access controls, and at query time Glean checks the signed-in user against them before returning anything.[4] The same check covers search results, AI answers and citations, agents and Glean's MCP servers.[4]

"Glean never grants broader access than the source system provides."

— Glean, Core security principles documentation.[4]

Assistant and agents. On top sit Glean Assistant, a chat interface, and Glean Agents, multi-step workflows that can call tools such as updating a ticket.[3] Agents run with the permissions of whoever triggered them, and web-app write actions wait for user approval by default unless an admin exempts them.[4]

AI Chronicle

How it relates to RAG

Glean is a commercial, company-wide implementation of retrieval-augmented generation. Its own glossary defines RAG as retrieving "permission-aware context from indexed or live sources" and passing it to a language model to ground an answer.[5]

The difference from a do-it-yourself pipeline is plumbing. A team building its own must write connectors, sync access rules, match one person's identity across apps and keep it all fresh. Glean sells that layer ready-made. The limits are the same: it can only answer from what it retrieves, and a model can still misread a correct passage. Pasting every file into a prompt instead runs into the limits in our context windows explainer.

The business model

Glean does not publish list prices; its pricing web address showed the general homepage when we checked.[1] Its chief executive, Arvind Jain, told TechCrunch in May 2026 that customers can pay on consumption, or on a hybrid of a fixed fee per active user plus separate charges for model usage.[6]

Announced funding and revenue figures, as stated by Glean:

TechCrunch noted that, with consumption pricing, part of the $300 million is closer to an annualised run rate than recurring revenue.[6] Glean also says over 85% of customers use it in five or more departments.[9] We could not verify these figures independently.

What to weigh

Permissions hygiene comes first. Mirroring permissions means mirroring mistakes. A pay spreadsheet shared with "All staff" is visible to every employee in the source app, and a good AI search tool will find it for them. Glean's documentation adds the qualifier "assuming the connector and configuration are correct."[3] Its developer guide is blunter:

"Permissions are the boundary between a useful connector and a data leak."

— Glean Indexing SDK documentation.[10]

Oversharing risk changes shape. Glean's docs list deliberate exceptions. Google Drive files shared domain-wide or by link are hidden from search unless certain conditions apply, such as the link being posted in a Slack channel the user belongs to (admins can disable that rule).[11] Glean sells Glean Protect, which it says can detect sensitive content and hide overshared documents.[12] Check which defaults apply to your sources.

Freshness is a security property. For indexed content, Glean applies the permission snapshot taken at the last crawl.[2] Ask how quickly a revoked share disappears from results for each source you plan to connect.

Cost is usage-linked. Under consumption or hybrid pricing, spend rises with use.[6] Glean says its context layer cuts the tokens a model consumes, citing its own benchmark.[9] Measure your own bill in a pilot.

Lock-in sits in the configuration. Glean describes itself as open, with model choice and an MCP gateway.[9] The switching cost lies in connector setup, tuned permissions and agents your teams build on it.

Alternatives exist. TechCrunch lists Google, Microsoft, OpenAI, Anthropic, Salesforce and Atlassian among companies building comparable tools.[6] Assistants built into productivity suites reuse the suite's permissions; Microsoft's documentation says its assistant surfaces only data a user can view, and tells admins to ensure "the right users or groups have the right access."[13] The trade-off is breadth across vendors against depth inside one suite. The hygiene work is the same either way.

A worked example (hypothetical)

Everything here is invented for illustration. No product was tested and no results are claimed.

A 600-person software company wants one place to ask questions across its wiki, drive, tickets and chat. It is weighing Glean against its office suite's bundled assistant.

Before connecting anything, the security lead audits sharing and finds a "2027 pay review" folder shared with all staff. The fix belongs in the drive; no AI tool would correct it. The team writes 20 test questions, each naming who should and should not get an answer, following Glean's advice to test the negative case.[10]

In the pilot, it revokes one person's access to a document and times how long it takes to vanish from answers, source by source. It runs the same questions through the suite's assistant and tracks usage charges weekly.

The decision rests on three checks: did any answer cite a document its asker should not see? How stale were permissions after a change? What did a month cost per active user? The pilot would have to establish these; none are reported here.

Questions readers commonly ask

Can Glean show me documents I can't open?

Its documentation says no: results, answers and citations are limited to what the user can open in the source.[4] Documented exceptions cover some domain- or link-shared files.[11] Over-broad source permissions will be reflected.

Is Glean just RAG?

Retrieval-augmented generation is the core pattern, and Glean's glossary uses the term.[5] The product adds connectors, identity mapping, permission syncing, a graph of relationships, an assistant and agents on top.

Does Glean train models on our data?

Glean's docs say model processing depends on your deployment and contract.[3] Get the answer in writing during procurement.

How much does it cost?

There is no public price list; expect a quote under consumption or hybrid terms.[6]

The next step

Every AI-at-work rollout meets this problem: the model needs company context, and that context has owners. Before choosing a vendor, understand the pattern underneath. Read what RAG is and how retrieval-augmented generation works, then return to the checklist above.

Earlier in the process? Our guide to starting with AI at work covers choosing a first use case, and What workplace AI studies actually show sets vendor time-saving claims against independent evidence. Our SaaS onboarding investigation agent tutorial shows an agent reading internal records under tight permissions.

Sources and evidence note

This profile uses Glean's documentation, website and press releases for product behaviour and company figures. It uses TechCrunch reporting to confirm funding and revenue figures and for the pricing description, and Microsoft's documentation for one alternative's permission model. Financial and adoption figures are company-stated unless noted. Documentation describes intended behaviour; it does not prove behaviour in any particular deployment. All pages were accessed on 16 October 2026.

Related reading

• LangChain's role in AI agent development
• Best AI research tools for evidence-led work