Imagine asking an AI to update an opportunity in Salesforce.

The AI has access to the CRM. It can see the account, the stage, the last meeting, the next step, the amount, and every field your company tracks. Maybe it can also search Google Drive, Slack, email, call transcripts, and your sales playbook.

It still might do the wrong thing.

Why?

Because nobody ever wrote a document called "How we use the Next Step field at this company."

An experienced sales manager may know that whenever Next Step changes, Next Step Date should change too. They may know that "waiting on customer" means something different from "no next step." They may know that a meeting does not automatically justify moving an opportunity forward a stage. They may know which fields matter for the Monday forecast call even though Salesforce itself does not require them.

None of that has to exist in a handbook.

People learn it by working with other people. They watch what experienced teammates do. They get corrected. They see examples. They hear, "No, we do it this way here." Eventually the rule becomes obvious enough that nobody thinks to document it.

Then an AI shows up with perfect access to the company's systems and no idea that the rule exists.

That is a different problem from enterprise search.

Most organizational knowledge is not really knowledge management

When companies talk about giving AI company context, the first instinct is usually to connect the places where information lives.

Google Drive. Slack. Notion. Confluence. Jira. Salesforce. Email. GitHub.

That is useful. An AI should be able to find the contract, the latest product spec, the support ticket, the pricing deck, or the conversation where somebody made an important decision.

But the information required to understand those artifacts and act correctly often lives somewhere else.

Consider a few examples:

  • A finance team knows which overdue customers should get a reminder and which ones should go straight to the account owner.
  • A recruiter knows that a "strong maybe" from one hiring manager means something different from the same phrase from another.
  • An engineer knows that one service can be restarted freely during business hours while another requires coordination with support.
  • A product manager knows that a customer request from one account is strategic while ten similar requests from another segment are not.
  • A CEO knows that a certain metric gets discussed every board meeting even though it is not part of the formal operating plan.

These are not necessarily secrets. They are not necessarily policies either.

They are operating context.

The organization may never have written them down because humans are good at carrying this kind of context socially. AI is not.

Search can retrieve evidence. It cannot guarantee the rule.

Suppose an AI can search every Google Doc and every Slack message in the company.

That is powerful. But now imagine the AI finds three pieces of relevant evidence:

One sales document says discounts above 15 percent require VP approval.

A newer enablement deck says 20 percent.

A Slack thread says enterprise deals follow a different approval path.

Search worked perfectly. The AI found relevant material.

It still does not know what to do.

Maybe the newer deck supersedes the old policy. Maybe the old document is still authoritative. Maybe both rules are correct for different segments. Maybe the Slack thread was a one time exception. Maybe the company genuinely has not resolved the disagreement.

Retrieval gives the AI evidence.

The organization still needs a way to decide what that evidence means.

This is why better retrieval is not enough for consistent AI behavior. Even excellent retrieval can repeatedly hand different AIs the same pile of evidence and ask each one to reconstruct the company's operating rule from scratch.

Sometimes they will agree. Sometimes they will not.

The important unit is the concept, not the document

Humans organize information into documents because documents are convenient for humans.

AI does not ultimately need a 30 page handbook. It needs the handful of rules from that handbook that matter to the task in front of it.

A useful AI context layer should be able to represent something closer to:

For enterprise opportunities, discounts above 20 percent require VP Sales approval.

That concept should carry the things needed to use it safely:

  • where it came from
  • who has authority over it
  • who it applies to
  • where it applies
  • what it supersedes
  • what exceptions exist
  • when it changed
  • whether the organization has actually agreed on it

Now the AI does not need to reread and reinterpret the original handbook every time somebody asks a pricing question.

The document remains useful as evidence and provenance. It just stops being the runtime unit the AI has to reason from.

The hardest context may not have a document at all

Now return to the Salesforce example.

There may be nothing to retrieve.

No policy says that Next Step Date should change whenever Next Step changes. Nobody put it in a wiki. It is simply how experienced people at the company use Salesforce.

There are still ways to learn it.

The rule may appear implicitly in hundreds of past opportunity updates. It may show up in manager emails. It may be visible in Slack corrections. It may surface when a sales leader corrects an AI directly:

Whenever you change the next step, update the next step date too.

That correction is not merely a fix to one CRM record.

It is evidence about how the organization wants this class of work performed.

A useful organizational AI system should be able to learn from both directions:

Explicit teaching. A person corrects the AI or explains how something should work.

Imputed context. The system notices patterns in email, CRM activity, documents, Slack, tickets, prior decisions, or other company systems.

Neither should immediately become universal company truth.

They are evidence.

The next problem is reconciliation.

Reconciliation is what makes AI consistent

Imagine two sales teams use the same Salesforce field differently.

One team's correction says Next Step Date always reflects the next customer interaction. Another team uses it for the next internal milestone.

A naive memory system might save both facts.

A naive retrieval system might return whichever one looks more relevant.

A naive company knowledge system might merge them into one rule and corrupt both teams' workflows.

The right answer may be that both rules are correct.

The system has to determine whether two statements describe the same concept, whether they conflict, who has authority, whether one supersedes the other, or whether they remain valid under different conditions.

That is reconciliation.

And reconciliation is not an administrative cleanup step. It is what makes consistent AI behavior possible.

If ten people use five different AI clients, the organization should not get ten independent interpretations of the same operating rule. The same underlying concept should resolve to the same intended meaning wherever it applies.

That means a company AI context layer needs to understand more than relevance. It needs some model of:

  • canonical meaning
  • provenance
  • authority
  • audience
  • applicability
  • version history
  • conflicts
  • unresolved uncertainty

Without those things, "company context" is still mostly a collection of evidence.

With them, it starts becoming organizational state.

Sometimes the correct answer is: the company has not decided

There is another uncomfortable implication.

Companies genuinely disagree with themselves.

The sales team and finance team may have different assumptions. Two executives may give contradictory guidance. A new process may be rolling out unevenly. A formal policy may say one thing while the operating team does another.

An AI system should not be forced to pretend every disagreement has already been resolved.

Uncertainty should be a first class state.

Sometimes the correct runtime context is:

These two current rules conflict. The organization has not established which one governs this situation.

Depending on the task, the AI can ask a person, route the question to the right owner, or avoid taking an irreversible action.

That is much safer than confidently choosing the document with the highest retrieval score.

This suggests a different way to bootstrap company AI

There is an interesting product implication here.

A company should be able to connect Drive, Slack, email, CRM, tickets, repositories, and other systems and use them to bootstrap AI context.

But the result should not simply be a giant index.

The first useful output may be a map of organizational concepts:

  • these five statements appear to mean the same thing
  • these two documents look like stale versions
  • these rules conflict
  • these two versions appear valid for different teams
  • this person is probably the owner of this concept
  • this operating convention appears repeatedly in actual behavior but nowhere in formal documentation
  • this issue is unresolved and needs a human decision

Then the organization can gradually turn the resolved parts into compact, durable context that AI can use consistently.

In other words:

Infer aggressively. Promote cautiously.

The system should be willing to discover candidate rules from the company's digital exhaust. It should be conservative about declaring those candidates to be company truth.

The difference between finding context and making it usable

Enterprise search is solving an important problem: find the information that matters.

For humans, that is often enough. A person can read three documents, notice which one seems authoritative, remember what their manager said last week, and apply years of tacit organizational judgment.

For AI, that hidden interpretation step is exactly the problem.

The AI needs the context in a form it can reliably apply.

That means the next generation of organizational AI infrastructure cannot stop at indexing what the company has written. It also has to capture what people teach, infer what the organization appears to do, reconcile competing evidence, preserve unresolved disagreements, and resolve the right rule for the task at hand.

The end result should be smaller than the source corpus, not larger.

The source systems contain evidence.

The AI needs the operating model.

From company knowledge to company behavior

For the last twenty years, companies invested heavily in storing information so people could find it later.

AI changes the requirement.

The question is no longer only:

Can the AI find the right information?

It is:

Does the AI know how this company expects the work to be done?

That answer will often depend on information that was never formally documented. It will depend on corrections, exceptions, habits, prior decisions, and the accumulated judgment of the people doing the work.

And when that context conflicts, the system has to reconcile it before AI can behave consistently.

That is the difference between giving AI access to company information and giving AI a durable understanding of how the company actually works.

For HexaHQ, that is the goal: capture and infer the context people use to do their jobs, reconcile it into organizational state, and make the relevant parts available so AI can make the right decision and take the right action wherever the work happens.