# The Distillery in Your Inbox

By Wes Pomeroy. Published May 29, 2026. https://wespomeroy.com/writing/ai-company-knowledge-base/

Every company is sitting on a fortune it can't spend.

It's in the email threads where a decision actually got made. The chat channel where someone worked out why the launch slipped. The wiki page nobody's opened in a year that holds the one diagram explaining the whole architecture. The meeting notes, the retros, the things your best person knows and has never written down.

Call it knowledge equity. Most of it sits dead - locked in inboxes, buried in scrollback, dying quietly in the heads of people one resignation away from taking it with them. For twenty years we've treated that pile as exhaust.

That's about to flip. The value of your corpus is going up sharply - not for everyone, but for the people and companies who can distill it. Because something just arrived that takes all that raw material and turns it into something a machine can actually use. A distillery for institutional knowledge.

## It's not the AI. It's what the AI gets to read.

Strip the marketing away and the thing at the center is a *knowledge graph*: a database built to store the relationships between things, not just the things. A map of your company's work - every ticket, doc, thread, file, plus who owns what and what depends on what - stitched into one web where every piece knows what it's connected to.

It inherits your existing permissions, so the AI never gets a god's-eye view. For a bank or a hospital, that's the difference between a working knowledge system and a lawsuit. It indexes most of your data but reads the touchiest stuff (email, chat, etc.) live and stores nothing. And it settles what your systems never agreed on: the Sarah who closed the deal in the CRM is the same Sarah who wrote the spec in the wiki.

None of this is new. Knowledge graphs have been around for more than a decade. Microsoft, Facebook, LinkedIn, Palantir - each built their own version, each feeding their own products. The data went in. You couldn't use those things on your company's data, and AI wasn't even in the picture.

Two things are changing at once. The new builders are opening the graph - outside agents can query it directly, instead of being locked behind whatever app the vendor ships. And the graph itself is being built for an AI to read, not as a private backend for one company's product. Either change alone is incremental. Having both at once is the big deal if you know what to do with it.

## "Tell me how we could have improved Q3 launch performance by 10%"

Here's what it buys you. Today, to pull together everything known about a project, you - or your AI - go tool by tool. Search email. Search the drive. Search chat. You get back differently-shaped piles of results and have to put in work and burn tokens to piece together what actually happened.

With the graph, you ask once. Tell me how we could have improved Q3 launch performance by 10%. What comes back to your LLM is already connected - the numbers, the people, the documents, the decisions, the threads, with the lines between them already drawn. The LLM didn't have to reconstruct how it all fit. The map already knew.

Ask once, get the whole connected picture. That's what every failed AI pilot of the last two years was reaching for. It's what finally makes the corpus usable.

## Who's building it

One example is [Atlassian](https://www.linkedin.com/company/atlassian/). I came across their Teamwork Graph through a [recent LinkedIn post](https://www.linkedin.com/feed/update/urn:li:activity:7465107664881217536/) by [Sanchan Saxena](https://www.linkedin.com/in/sanchans), one of their SVPs. At their conference this month, they opened the graph that had quietly been running their own AI so any AI agent, not just theirs, can read it. Their CEO put the bet in a line: intelligence is the engine, context is the fuel. With AI models getting cheaper by the month, Atlassian is betting the platform that wins is the knowledge graph itself - the software that turns your scattered work into something an AI can use. Your data stays yours; they sell what makes it usable.

They're not alone. The independent making the same bet is [Glean](https://www.linkedin.com/company/gleanwork), founded by a [former Google engineering leader](https://www.linkedin.com/in/jain-arvind/). No other product line, no apps to push, no foundation model to favor - they sell the knowledge graph and nothing else, and they sell it as an open platform. Over a hundred connectors into Microsoft, Google, Slack, Salesforce, Jira, Confluence. Fifteen-plus large language models across the major clouds, so the model is your choice, not theirs. Doing about $200 million a year, doubled in nine months. Customers like Samsung, Reddit, Databricks. The pitch is what Atlassian gestured at and Glean built the whole company around: be the neutral ground that connects everything, because most companies use everything.

The giants have their own versions - Microsoft Graph wired into Office, Google's Workspace Intelligence into Workspace, Salesforce's Data 360 into the CRM. Each reads beautifully across its own products and stiffly across everyone else's. The bet they're making is that you'll consolidate into their stack to get the AI working. The open bet is that you won't, because nobody does.

## This only gets you halfway

This is where the excitement stops short of the actual problem.

A graph that can show you everything you're allowed to see is not the same as a system that knows what it ought to show you. You've handed the AI *can*. You've said nothing about *should*.

A concrete version. You have access to a chat message where a colleague vents about their manager. It names a project. You ask your AI to pull together everything about that project. The graph does what you told it: the rant comes back, lands in the summary you forward to your team. Permission said you could see it. Nothing in the system knew you shouldn't have used it.

When I scraped my own corpus to get an AI working across my own projects, this is the wall I hit - and I hit it by hand. No graph waiting to map it for me. Getting the data in was the technical half: scripting, parsing, hand-wrangling formats until everything was queryable. Anyone with patience could do it.

What took years of being a manager and a practitioner was the second half: teaching the thing judgment. Don't treat a vent as a decision. Don't pull a draft as final. Leave the personal stuff alone even when you can read it. Strip the medical, the legal, the HR. Every rule was me deciding what should pass through, not what could. None of it came from permissions. All of it came from knowing my context.

The graph gives you *can*. It does not give you *should*. And the should is where you earn it.

## The layer nobody's building

All that filtering - what's relevant, what's appropriate, what's authoritative, what to leave alone - has to happen somewhere. After the graph hands back what you can see, and before the AI acts on it. That somewhere is a layer that mostly doesn't exist yet as a product. Call it the *judgment layer*.

The whole industry is racing to own the two layers on either side. The knowledge graph underneath - Atlassian, Glean, the giants. The AI reasoning on top - the model companies. The middle layer is an open field. That's where the value concentrates. Graphs are becoming infrastructure; every company will have one. Models are a commodity, cheaper every month. The judgment - what matters and what doesn't, what's appropriate and what isn't - is the part that's different for every company and team. You can't buy it off a shelf.

It can't be a single product, either. Judgment doesn't generalize across a company. To a marketer, the decision happened when the team aligned in a thread on Tuesday. To a lawyer, nothing was decided until it was documented, reviewed, and signed. Neither is wrong. They're running on different rules, and no single filter serves both. So the judgment layer is really a stack of them. Marketing's judgment. Legal's. Engineering's. The graph captures what happened. Judgment captures what it meant. The mapping between them isn't data - it's expertise.

## Your new job: write down what you know

If that's right, it changes what a senior person is for.

For most of knowledge-work history, you showed your value by doing the work - the campaign, the deal, the call. The output was the proof. With a knowledge graph underneath and a judgment layer on top, the most valuable thing you ship isn't the campaign anymore. It's the encoded judgment that lets juniors and agents produce it as if you'd run it yourself. What you ship becomes the frame the decisions get made inside.

Anyone senior has already made a version of this move. Early on you write the thing yourself - the release, the email, the post. Later the job becomes shaping the voice and the rules others write inside of, so the work comes out right without your hands on it. The judgment layer is that same shift, made explicit and handed to machines. Your expertise stops sitting in your head and starts as a file your team's agents can load.

I'm building one of these at work right now. The project side decides what to trust - by rules I wrote. A strategy deck and a go/no-go memo are signal; a status update is noise. The people side grounds an agent in firsthand observations I've written down about each of my reports - how each one works, what motivates them, who their allies are, what they need watching for. The calls are mine. The system executes them. Years of pattern recognition, finally outside my head.

That's the part you can't hand off. Not to an agent run by someone without the tradecraft. Let alone to one running on its own. Writing judgment down looks like writing rules, which makes it look like something you could spec out and assign. It isn't. The judgment is what you've earned by doing the work. The writing-down is part of the work. Once it's on paper a machine can run it. The earning stays with you.

For managers, the assignment is more direct. The institutional knowledge your best people carry is, today, a liability the morning they resign. The judgment layer is the first real chance to capture it - not as a wiki page nobody reads, but as something the team's tools actually run on. The job is getting that expertise out of heads and into a form that compounds.

## What it looks like when it works

Two quick pictures.

A marketing team at a big company. Their knowledge graph connects everything - briefs, brand guidelines, past campaign reviews, the Slack threads where creative got argued out, the approval emails. On top of that, the team writes down what counts. A casual comment in chat is brainstorming, not a decision. The brand guidelines doc is the source of truth; a passing comment in a meeting doesn't change that. A plan isn't approved until it's logged in the project tracker. A junior with an agent asks for the launch brief in the team's voice, consistent with the last three launches. The agent pulls the real history from the graph, runs it through the team's rules, and produces something that sounds like the team. The senior marketer didn't write the brief. They wrote the rules the brief gets written inside.

Second picture, regulated. A compliance team. The graph reaches everything, but the team's judgment is strict. A promise made over chat isn't binding. A draft document isn't policy until it's been officially signed off. Customer data has extra rules. An agent putting together a response to a regulator pulls only what the team has marked official and ignores the rest, even though it can see everything. What makes that safe to deploy isn't the graph. It's the judgment sitting on top.

Same pattern in both cases. The graph supplies access; the team's judgment, written down, decides what gets used. Companies that get good at writing it down - per team, per function - are the ones whose AI produces work the team can stand behind.

## What's yours

The value doesn't flow to whoever holds the most data. Everyone has data. It doesn't flow to whoever rents the smartest model - that's a monthly invoice, available to anyone. It flows to whoever can take the raw material of their own work and refine it into judgment a machine can run on.

The knowledge graph is shared infrastructure. The judgment is yours. The work of the next few years, for anyone who's spent a career getting good at something, is to take that expertise out of your head and write it down where your tools can reach it.

That's the distillery. Everything you've made - every doc, every thread, every decision - is what goes in. The graph is what does the work. The person deciding what to keep, what to set aside, what good looks like, and how that keeps changing - that's still you.