The Company Second Brain
The missing layer for enterprise AI.
Himanshu Sharma1. The question nobody can answer
In the Monday leadership meeting, someone asks: “Why did revenue drop last quarter?” The analyst says they will look into it. They know it will take time: pulling the right data, checking how the company defines revenue, working out which period to compare, and tracing the decline back to what actually changed. Three days later they return with a good answer. The meeting has already moved on.
Now give the same question to any LLM. Give it the warehouse, all 800 tables, and permission to query any of them. Four seconds later it returns a chart, three reasons, and a confident explanation. The SQL is valid. The numbers are real. The answer is wrong, and nobody in the room can tell.
It counted one-time setup fees as revenue, though the company keeps those separate from recurring revenue. It logged a customer upgrade as new business, though the company counts upgrades as expansion. And it blamed the drop on fewer new deals, because that was the biggest movement it could see, missing a group of customers who had paused their spending.
The first mistake changed the size of the drop, from 9 percent to a reported 14. The second put it in the wrong bucket. The third gave it the wrong cause. Nothing in the output says a choice was made at all.
The analyst knew the business. The AI knew the data.
Everything that follows is about closing that gap.
2. What the analyst knows that the AI does not
Think about what the analyst carries in their head. They know setup fees are billed once and never counted as recurring revenue, because Finance settled that in a meeting in March 2024 and the board deck has used it ever since. They know an upgrade is expansion, not new business, so it belongs on a different line of the report. They know the status column on subscriptions holds single letters, and that P means paused, not cancelled, so those accounts are still customers even though they bring in nothing this quarter. They know there are two customer tables, left over from a migration nobody finished, and which of the two is real. They know trial and internal accounts never count, though no rule anywhere says so. They know the last time revenue fell this sharply it was a billing outage, not churn, which changes what you look at first.
Not one of these things is in the database. A few are in a Notion page nobody has opened in a year. The rest is in three people, and one of them is leaving.
3. The company’s context has to be mined, not written
The obvious fix is to write it all down. Define what a customer is, which table is the real one, how revenue is calculated, and which accounts never count. Put it in a semantic layer and let the AI read from it.
The goal is right. The method breaks down, for three reasons.
There is too much of it. An enterprise has thousands of tables across many systems, millions of saved queries, and decades of decisions. Writing all of that by hand is a project that never finishes.
People don’t know what they know. The analyst would never think to write down that P means paused. It only comes up when someone gets it wrong.
Written definitions go stale. The business changes, the page doesn’t, and nobody owns keeping it current. That is how the Notion page from the last section ended up unread.
But look again at what the analyst knew. Almost none of it is truly secret. It is spread across the systems the company already uses every day. Nobody has put it together.
The knowledge is already there
Every one of the analyst’s facts leaves a trace in one of four places. These are the four legs of a company second brain.
- Databases hold more than data. Its query history shows how people actually use it. If Finance has joined subscriptions to
accounts_v2four hundred times and to the old table almost never, that is the real customer table. If nearly every revenue query filters out trial and internal accounts, that is a rule, even though no one wrote it down. And if accounts markedPstill have open contracts but no invoices, they are paused, not gone. - Business context is how the business works: its business model, products and pricing plans, customer segments, the metrics leadership tracks, and how each metric is split. It lives in BI dashboards, KPI reports, the product catalog and billing plans. Every revenue dashboard puts upgrades on the expansion line, not new business. That is how this company counts growth.
- Company knowledge is what the company has written down and decided: policies, SOPs, rules, goals and meeting notes. The March 2024 Finance notes explain why setup fees are not recurring revenue.
- Third-party tools such as Slack, Jira, the CRM and the billing system hold what is happening and why. The last sharp drop in revenue has a Jira incident and a Slack thread behind it: a billing outage, not customers leaving.
An agent that reads the company
So instead of asking people to describe the company, we let an agent read it. The context mining agent works through all four legs: the databases, the business context, the company knowledge and the third-party tools. For each thing it learns, it records a finding along with the evidence behind it: which queries, which document, which thread.
It does not copy everything into one large pile of text. A pile of text is what the company already has. It pulls out the specific facts needed to answer questions correctly, and links each one back to where it came from.
How the agent builds a first draft
A draft, not a final answer
The first draft will not be fully right, and it doesn’t need to be. Some of what the agent finds will disagree. The Finance notes say setup fees are not recurring revenue, but dozens of older dashboard queries still count them. The evidence alone cannot settle that. It is a business decision.
So the agent does not guess. It flags the conflict, shows the evidence on both sides, and sends it to the person who owns the answer. A Finance lead can settle it in a minute. Most findings need no one at all. People step in only where the evidence disagrees, the meaning is unclear, or a real business call is needed.
And because the agent keeps reading, the draft does not go stale the way a written page does. When a new rule starts showing up in the queries, it shows up in the draft too.
The old approach asks people to write down the company’s context. The Second Brain turns that around. The agent mines it, and people decide only what the agent cannot.
What the agent builds has a specific shape: four parts, each doing a job the others cannot. That is the next section.
4. What the company second brain is made of
The agent can find thousands of facts. But a pile of facts is not a second brain, and an AI cannot answer a question from a pile. The knowledge needs structure.
That structure has three foundational layers, each built on the one before it, with skills on top:
- Ontology: what things are.
- Context graph: how things connect, and why.
- Semantic layer: how things are counted.
- Skills: how the work gets done.
The first three describe the business. Skills describe what to do with that knowledge. And the order matters: you cannot count revenue until you know which tables hold it, and you cannot find those tables until you know what revenue means here.
Where the company second brain sits
Four sources
Company second brain
Ontology: what things are
The ontology is the company’s agreed model of its business. It names every kind of thing the company deals with: customer, subscription, invoice, upgrade, setup fee. It gives each one a clear definition, and it sets the rules for what can and cannot be true about it. A subscription always belongs to one customer. A trial account is never a customer.
It is not a textbook model. It describes this company, odd parts included. A paused subscription still belongs to a customer. An upgrade is growth from an existing customer, not a new deal.
Without an ontology, every tool and every AI works out what “customer” means on its own, each time it runs, and each one can get a different answer. With it, meaning is decided once and applied everywhere. That is exactly what went wrong in the first section. The AI did not fail at SQL. It failed at meaning.
Context graph: how things connect, and why
The context graph is the company’s memory of how things connect, and why. Each point on it is something the company has: a customer, a table, a metric, a policy, a person, a decision. Each line is a named link between two of them: a customer has subscriptions, revenue is owned by Finance, the setup fee rule came from the March 2024 Finance meeting, the last sharp drop in revenue was caused by a billing outage. Every link also records when it became true and who approved it, so the graph can tell you what was true on any past date.
The ontology says what can be linked. The graph holds the real links, tied to the real data: customers live in accounts_v2 and join to subscriptions on customer_id. An AI can only use a join the graph has proven, so it can’t invent one.
Because the graph holds decisions and history, not just structure, it answers what a table diagram or a catalog cannot: why this number exists, who owns it, and what changed and when.
Semantic layer: how things are counted
The semantic layer is where meaning becomes a number. For every metric the business uses, it holds one approved definition: what the metric means, how it is calculated, and which ways it can be broken down. Recurring revenue, for example, is invoice lines, minus refunds, minus setup fees, without trial or internal accounts. It can be broken down by plan or by region.
A graph is good at connections, but not at adding numbers up the same way every time. That is this layer’s job. The graph works out which things a question is about. The semantic layer decides exactly how to count them, so the same question always gets the same number.
It also makes the brain honest about its limits. If a question needs a metric that has no approved definition, or asks to break a number down in a way that makes no sense, the brain stops and says so instead of guessing. A clear “not defined yet” is far better than a confident wrong number.
Skills: how the work gets done
Skills hold the company’s know-how: the steps to follow for a kind of question, and what is allowed along the way. When revenue falls sharply: check billing first, then paused and cancelled customers, and only then new deals. Before a number goes to the board: get a Finance sign-off.
A skill holds no numbers or definitions of its own. It uses the other three layers, and every AI tool follows the same steps. Without skills, an AI can have every number right and still look in the wrong place first. That is how it blamed the drop on fewer new deals and missed the paused customers.
Governed, and shared by every tool
All four layers follow the same rules:
- Every fact links to its evidence.
- Every definition and skill has an owner and a version history. Change a rule once, and every new answer uses it. Re-run an old answer, and it uses the rules of its time.
- When the data drifts away from a definition, the owner is told.
- Access rules apply before a query runs, so people only see what they are allowed to see.
The brain does not answer questions by itself. AI analysts, dashboards and other tools all read from it, so they give the same answer, and a fix made once reaches all of them.
That is the company second brain: what things mean, how they connect, how they are counted and how the work gets done, in one place every AI tool reads from.
The next section follows one question through it.
5. Inside a real AI analyst query
Back to the Monday meeting. Here is what happens between the question and the answer when an AI analyst uses the company second brain.
1The question
“Why did revenue drop last quarter?”
2Understanding the question
The language model works out what is being asked, and nothing more. The metric is revenue. The period is last quarter, April to June 2026, compared with January to March. The task is to explain the change, not just report it.
3Ontology: what the words mean
- “Revenue” in a question like this means recurring revenue.
- A setup fee is a one-time charge, not recurring revenue.
- An upgrade is expansion, not a new deal.
- A paused customer is still a customer. Trial and internal accounts are not customers at all.
4Context graph: where the data lives, and what happened before
Customers live in accounts_v2, not the old table. Subscriptions join to accounts on account_id, and invoices join to subscriptions. Finance owns the revenue definition. And the last sharp drop in revenue was a billing outage, recorded in Jira.
5Semantic layer: how to count it
Recurring revenue is invoice lines, minus refunds, minus setup fees, without trial or internal accounts. To explain a change, it is broken down by movement: new business, expansion, contraction, churned and paused.
6Skills: where to look first
The revenue-drop skill sets the order: check billing first, then paused and cancelled customers, and only then new deals.
7Semantic compilation
The semantic compiler turns all of this into one SQL query. It uses only what the brain says, so the query is built on the same definitions and relationships every time. In plain English, the query:
- reads customers from
accounts_v2, with their subscriptions and invoices - leaves out setup fees, trial accounts and internal accounts
- compares each customer’s recurring revenue in the two quarters
- groups every change by type: new business, expansion, contraction, churned or paused
8Execution
| Movement | Change in recurring revenue |
|---|---|
| Paused | −445,000 |
| Churned | −110,000 |
| Contraction | −40,000 |
| Expansion (upgrades) | +95,000 |
| New business | +140,000 |
| Total: 4,000,000 → 3,640,000 | −360,000 (−9.0%) |
9The answer
The first thing the AI analyst finds is that the number in the room is wrong. Recurring revenue fell 9 percent, not 14. The 14 percent counts setup fees, which Finance ruled out of recurring revenue in March 2024. Setup fees were unusually high in the first quarter because of a large onboarding wave, so counting them makes the drop look bigger. It shows both numbers and says where the 14 came from.
Then it explains the 9 percent, in the order the skill sets. Billing first: no outage and no rise in failed payments this quarter, so this is not a repeat of last time. Then paused and cancelled customers, and that is where the answer is. Paused customers took out 445,000, more than the whole drop. Without them, revenue would have grown by about 2 percent. Most of those pauses are education customers who pause every summer and usually come back in September. New business was not the problem. It added 140,000, and upgrades added another 95,000, reported as expansion rather than new deals.
And it says what it is not sure about. Fourteen paused accounts have no restart date in the CRM. The AI analyst cannot tell whether they are coming back, so it lists them for the account team instead of guessing.
The answer arrives with the definitions it used, the SQL it ran, the rows that came back and the skill it followed. Anyone can check every step. That is the difference between an answer and a guess that sounds confident.
6. A living brain that learns what it does not know
A system that always produces an answer is a system that is sometimes confidently wrong. So the most important thing the brain holds is a clear record of what it does not know. When a question runs into one of those gaps, the AI analyst stops instead of guessing.
There are four kinds of gap, and each one has a different way out.
Ambiguity. “Who are our top customers?” could mean by recurring revenue, by number of seats, or by lifetime value. The AI analyst does not pick one. It asks a clarifying question before it answers.
Conflicting definitions. This is the setup fee problem from section 3. The Finance notes say setup fees are not recurring revenue, but 38 dashboard queries still count them. The AI analyst does not average them or quietly pick one. It shows both, says who uses each and where each came from, and asks the owner to decide.
Missing data. Cost to serve has only been recorded since March 2025. A margin question about earlier months gets a clear “can’t answer yet”, with that date attached.
Questions it cannot support. “Will revenue recover next quarter?” is a forecast, not a query. The AI analyst says so and offers what it can do instead: show how many paused customers came back in past years, and when.
How the brain keeps learning
Gaps are found in four ways, and each one turns into a proposed change to the brain:
- Clarifying questions, in the moment. If the answer to a clarifying question should apply to everyone, such as what “top customers” means, it becomes a proposed change.
- Feedback, after the answer. Anyone can mark an answer as wrong and say why. A feedback agent checks this against the evidence. If it holds up, it becomes a proposed change. If not, the agent explains why the answer stands.
- Monitoring, all the time. A monitoring agent watches for data drifting away from a definition, schema changes, new tables nobody has described, joins that start to fail, duplicate metrics, and the same question being asked again and again.
- Learning from use, in the background. The brain also watches how people and AI tools actually use it. It notices which words people use for a metric and which definitions they rely on, and it suggests better synonyms and descriptions.
Checked, tested, then sorted by risk
Every proposed change goes through two checks before it can go live. First, a rules check: anything that breaks the rules already in the brain, like a revenue definition that counts trial accounts, is stopped. Second, a test against past questions: the change is replayed on questions people have already asked, to see which answers would change and by how much.
No one approves a change without knowing what it will move.
What passes is sorted by risk. Small, safe changes that can be undone, like a new synonym or a better description, go live on their own, and people check a sample. Anything that changes what a number means or who can see what, like the definition of revenue or an access rule, waits for a person.
When a person is needed, they see what would change, why, the evidence behind it, and what the test showed. They make one decision. The decision is saved, every tool picks it up, and the same gap does not come up again.
The gap is closed, and the brain knows more
This is what a living system really means. Not that the brain refreshes itself on a schedule, but that it gets better every time it is wrong in a place people can see, at a pace people can keep up with. Forty decisions in the first week. Five in the second month. None of them a surprise.
7. What stays when people leave
The first question is expensive. Revenue has to be pinned down, the right tables found, trial and internal accounts left out, and one conflict about setup fees settled by Finance.
The second question about revenue costs almost nothing. The hundredth costs nothing at all. A brain written by hand costs the same for every metric, forever, because someone has to write each one. A mined brain gets cheaper over time, because each decision covers everything built on top of it.
And it is shared. A forecasting agent, a churn alert and the month-end close all read revenue from the same brain, so they agree without anyone having to prompt them carefully.
The part that cannot be copied
AI models improve every few months, and whichever one is best today will not be best next year. That is fine. The model is the part you can replace.
Your ontology, your decisions and your history are not. They were built from your data, your query history and the judgment of your own people. No competitor has them, and no vendor can hand them to you.
We built an AI analyst first, because answering real questions is the fastest way to find out whether the brain is right. It will not be the last thing built on it. Once a company’s context can be read by software, anything that needs to understand the business can use it.
Closing the gap
The goal was never to give an AI more data. Companies already have more data than any model can make sense of. What was missing was the context around it: what things mean, how they connect, how the company counts them, and what to do when the answer is uncertain.
That context used to live in a few people’s heads, and one of them was leaving. Now it lives in the second brain: mined from the company, decided by its own people, and used by every AI tool that needs it. The analyst no longer has to be the only bridge between the business and its AI.
The analyst knew the business. The AI knew the data. With a second brain, the AI knows both.
Stop asking the model to guess how your company works. Give it somewhere to look it up.