How does it avoid showing people things they should not see?

By recording each document's existing access rules when it indexes and applying them per person at the moment of asking. Nobody gains access they did not already have, which is the whole design constraint, and keeping those rules current as they change in the source systems is the part that makes this genuinely hard engineering rather than a search feature.

What decides whether the answers are any good?

The state of your material, far more than the quality of the product. A tidy estate produces answers that feel like magic and a contradictory one produces well-referenced confusion, because the tool reports what it finds rather than judging it. Budgeting for curation alongside the licence is the difference between a successful rollout and a disappointing one.

Do we need this if our work lives in one workspace?

Usually not, and that is worth establishing before a procurement starts. The value comes from spanning systems that were never chosen together, so if most material already sits in one place, that platform's own search plus its AI layer covers a large share of the need at a fraction of the effort.

How is this different from a general assistant that can search?

Scope and permissions, rather than the interaction. An assistant searching the open web answers from what is published; this answers from what your organisation holds, and must respect who is allowed to see each part. The pattern feels identical to use, which is exactly why the difference gets missed in demonstrations.

Glossary

Enterprise search

Enterprise search reaches across the systems an organisation already runs and answers from all of them at once, showing each person only the material their existing access permits.

In plain terms

One place to ask a question about anything your company knows. It reaches into the drives, the wikis, the tickets and the chats rather than making you remember which of them held the thing you half remember reading. The catch is that it must never show you something you were not already allowed to open, and that constraint is most of the difficulty.

Why it matters

Because the alternative is an organisation that cannot find what it already knows. Material accumulates across systems nobody chose together, each with its own search that only covers itself, so the practical answer to most internal questions becomes asking a colleague. That is expensive in a way that never appears on a budget line.

How it works

It indexes rather than searches on demand, which is what makes an answer possible at all. Content is pulled from each connected system in advance and held in one place that can be queried quickly, because reaching out to a dozen systems per question would be far too slow to feel like search. That index is a copy, which is precisely why the permissions question follows immediately.

Permissions are the hard part and they have to hold at answer time, not at index time. Every document already carries its own access rules from whichever system owns it, so the index must record them, keep them current as they change, and apply them per person on every query. Getting this wrong is not a bad result but a disclosure, which is why these products are bought slowly and by committee.

Answer quality is inherited from the estate rather than produced by the tool. The retrieval and the writing can be excellent while the underlying material is contradictory, superseded or years stale, and the result is a confident, well-referenced answer built on a document nobody has believed since a reorganisation. One vendor's own page puts it plainly: stale documents produce well-cited stale answers.

Maintenance has become part of the product rather than a prerequisite for it. Rather than assuming a tidy corpus, several tools now detect conflicting content, merge duplicates, flag material whose sources have changed and identify gaps where no answer exists, which is an admission that no real organisation has a tidy corpus. Treating curation as ongoing rather than as a migration step is the more honest posture.

The scale assumption is real and cuts both ways. Connectors, permission modelling and governance are built for organisations with hundreds of seats and an IT function to own them, so the procurement matches, while smaller teams are usually served well enough by the search already inside whichever workspace holds most of their material. Buying the enterprise shape too early costs more in effort than in money.

The two halves, and which one is priced

The reason this distinction matters during an evaluation is that the left column is what a demonstration can show in twenty minutes and the right column is what decides whether the deployment survives its first year. Any of these products will answer a question about a document beautifully when the person asking has access to everything, which is the situation in every demonstration environment ever built. The interesting cases are the ones nobody demonstrates: the contractor who should see three folders and not the fourth, the document whose permissions were tightened last week, the team drive that changed owner during a reorganisation, the person who moved between departments and kept access they should have lost. Each of those is a correctness problem where the failure mode is not an unhelpful answer but a disclosure, and it is the reason this category is bought by committee and priced accordingly. The question worth asking a vendor is not how good the answers are but how quickly a permission change in the source system takes effect here.

Seen in the wild

A product that indexes the organisation's applications and answers with references, showing each person only what their permissions already allow.
Glean
A curated company knowledge layer whose agents give cited, permission-aware answers and actively maintain the material behind them as sources change.
Guru
A work-management platform whose AI layer searches across connected applications from inside the system that already holds the work.
ClickUp (Brain)

Common misconceptions

People assume

It is a better search box.

In fact

It is a permissions engine with a search box attached, and that ordering explains the price. Assembling one answer from many documents means resolving every one of their access rules for the person asking, on every query, as those rules change underneath. The retrieval is the easy half.

People assume

It will fix our knowledge problem.

In fact

It makes the existing state findable, which is wonderful when the material is good and merciless when it is not. Contradictory and superseded documents become confident, well-referenced answers rather than pages nobody happened to open, so a messy estate gets louder rather than quieter. Several tools now treat cleaning up as part of the job for exactly this reason.

People assume

Any team would benefit.

In fact

The shape assumes scale, and below it the effort outweighs the gain. Connectors, permission modelling, rollout and governance are built for organisations large enough to have lost track of their own material, and a team whose work sits mostly in one workspace is usually better served by the search already in it.

Telling them apart

Enterprise search vs Knowledge base

Enterprise search

Reaches across systems nobody curated together.

Knowledge base

One curated place, deliberately maintained.

One finds what exists wherever it landed; the other is a decision about what should exist. Organisations that buy the first often discover they needed the second as well.

Questions

How does it avoid showing people things they should not see?: By recording each document's existing access rules when it indexes and applying them per person at the moment of asking. Nobody gains access they did not already have, which is the whole design constraint, and keeping those rules current as they change in the source systems is the part that makes this genuinely hard engineering rather than a search feature.
What decides whether the answers are any good?: The state of your material, far more than the quality of the product. A tidy estate produces answers that feel like magic and a contradictory one produces well-referenced confusion, because the tool reports what it finds rather than judging it. Budgeting for curation alongside the licence is the difference between a successful rollout and a disappointing one.
Do we need this if our work lives in one workspace?: Usually not, and that is worth establishing before a procurement starts. The value comes from spanning systems that were never chosen together, so if most material already sits in one place, that platform's own search plus its AI layer covers a large share of the need at a fraction of the effort.
How is this different from a general assistant that can search?: Scope and permissions, rather than the interaction. An assistant searching the open web answers from what is published; this answers from what your organisation holds, and must respect who is allowed to see each part. The pattern feels identical to use, which is exactly why the difference gets missed in demonstrations.

Key takeaways

It indexes in advance, so the copy inherits every document's access rules.
Permissions resolve per person at answer time; that is the expensive half.
Answer quality is inherited from the estate, not produced by the tool.
Curation is now part of the product because no real corpus is tidy.
Below enterprise scale the workspace's own search usually covers the need.

Tools that use this

Glean
Indexes the estate and answers with references, permission-aware.
Guru
A curated layer whose agents maintain the material as sources change.
ClickUp (Brain)
Search across connected apps from inside the system holding the work.