How to give AI company context and build a company brain

Ask an AI assistant for Q3 revenue and you’ll get a confident answer. Whether it’s the right one depends on whether it knows how your company counts revenue.

To give an AI assistant context about your company, write down what it can’t guess: your definitions, policies, and how work gets done. Then connect it to the tools where work happens, and put both somewhere your whole team works from.

The shared part is the easiest one to skip. If the context lives in one person’s setup, only that person gets better answers. If it lives where the whole team works, every answer draws on what the company knows. That shared, always-current context is what people mean by a company brain.

What is a company brain?

A company brain is a shared, AI-readable record of what your company knows: the definitions, decisions, policies, and workflows that usually live in people's heads, Slack threads, and old docs. Every employee and AI agent works from the same context, and it grows as people teach it and build on each other's work.

The idea isn't new. Organizational memory and institutional knowledge have always existed, but AI can only use them once they're written down and connected. A model learns the internet in training, but it only learns your company from what you give it.

A company brain overlaps with two tools you probably already have, but it does a different job:

Why AI gets your company wrong without context

Out of the box, an AI model knows the internet, not your business. It doesn't know you report recognized revenue, that your fiscal year starts in February, or that a parent account and its three subsidiaries count as one customer.

This is the tribal knowledge AI gets wrong most often:

• Metric definitions: booked vs. recognized revenue, logo vs. dollar churn, and what counts as an "active" customer
• Calendar: fiscal vs. calendar quarters and close deadlines
• Thresholds: who approves what, and above which amount
• Entity and regional rules: per diems, holidays, and policies that differ by country
• Internal names: acronyms, product code names, and team names that mean something else in public

The cost shows up as time. In Atlassian's State of Teams research, leaders and teams said they waste 25% of their time searching for answers. And most of what people teach AI stays with them: in a Microsoft survey of 2,003 U.K. employees, 71% said they'd used unapproved consumer AI tools at work. That's why getting AI ready starts with context and control, not a new model.

How to give AI context your whole team can use

There are two ways to do it. You can build the system yourself: write a shared context doc, connect your tools to an assistant that supports team projects or connectors, lock down permissions, and name owners to keep it current. Or you can use an AI workspace built for teams, like Glass, where shared context, connected tools, and reusable workflows live in one place everyone works from. Either way, the steps below are the same. What changes is how much of the setup and upkeep sits with your team.

1. Write a starter company context doc

Start with one document, one to two pages long, that answers what a smart new hire would ask in their first week. Write it for a reader with no history at your company, because that's exactly what an AI model is.

A starter doc covers seven things:

• Company snapshot: what you sell, who you sell to, headcount, entities, and when your fiscal year starts
• Glossary: internal acronyms, product names, and team names
• Metric definitions: the formula, source system, and owner for each number you report, like ARR, gross margin, or churn
• Customer segments: how you define small business, mid-market, and enterprise
• Policies: expense limits, approval thresholds, and data handling rules
• How work gets done: who approves what, your close calendar, and where the truth lives (for example, "pipeline lives in the CRM, revenue lives in the ERP")
• Tone and format: how you write customer emails, board updates, and internal docs

Put a date on anything that changes, like pricing or headcount, so you can tell when it's stale.

2. Choose where your context lives

Each place you can put context trades off reach, freshness, and setup effort. Here's how the five common options compare.

Fine-tuning is the method people reach for first and need least: OpenAI's accuracy guide maps missing knowledge to retrieval and inconsistent behavior to fine-tuning. The Model Context Protocol (MCP) is an open standard for connecting AI apps to data and tools, and Claude, for example, can connect to any service that supports it.

Most companies end up with two layers: a curated doc for definitions and rules, and connectors for live records.

3. Connect the tools where work happens

Static docs cover the rules, but the answers people need day to day sit in Slack threads, call notes, tickets, and shared drives. Connecting those tools lets AI answer from current records instead of last quarter's export.

Start with the two or three tools where decisions get made. A sales team gets the most from call notes, a support team from tickets, and a finance team from the docs that hold its close checklist and policies. Add more once people trust the answers.

Pick tools that show their sources. ChatGPT's company knowledge, for example, can link to the source behind an answer, so people can check it before they act on it.

4. Set permissions before you connect anything

Connectors usually inherit the access people already have. Claude, for example, "inherits each person's permissions from the connected service," according to its connector docs. That protects sensitive data, but it also means any file shared with "anyone with the link" becomes findable through AI.

Before you connect a tool:

• Audit broadly shared folders and channels, especially finance, HR, and customer data
• Decide who can turn on connectors (on Claude's Team and Enterprise plans, an Owner enables them, and each person still signs in separately)
• Separate read access from write access for any AI agent that can take actions
• Check each vendor's data terms to confirm your plan doesn't train on your data

5. Test with known-answer questions

Write 10 to 20 questions you already know the answers to, and run them every time you change the context. Mix easy lookups with the questions AI usually gets wrong:

• What was our gross margin last quarter, and how do we calculate it?
• Does a customer with three subsidiaries count as one customer or four?
• Who approves a $15,000 software purchase?
• When does our fiscal year start?

Tell the AI to cite where each answer came from and to say when something isn't in its context. A model that says "I don't know" is more useful than one that guesses.

6. Name owners and keep context current

Context goes stale the day a policy changes. Give every section an owner (finance owns metric definitions, people ops owns policies) and a review date.

When the AI gets something wrong, log it and sort it into one of two buckets: a missing fact, which you add, or a stale fact, which you update. A shared log turns every wrong answer into a fix for everyone, not a private correction in one chat.

We saw this firsthand building Glass at Ramp. Setups rarely break on day one. They break later, when someone builds a solid first version and nobody goes back to update it.

7. Turn good work into shared workflows

The most valuable context isn't written in a doc. People create it while they work: the prompt that turns a messy customer call into a clean account plan, or the workflow that catches budget variances before the board deck.

Save that work as a shared workflow in a project your team uses, so the next person starts with both a proven process and the context behind it, not a blank page. That's the shift from personal context to a company brain: knowledge stops disappearing when someone closes a tab, goes on vacation, or leaves.

At Ramp, one person's shared slide format meant people making slides for the first time could start from it instead of working the format out on their own.

Personal AI context vs. a company brain

Most companies sit somewhere across these three stages. The goal is to move your most-used context to the right.

How we built ours at Ramp

We ran into this at Ramp. Our most AI-forward employees had built great setups for themselves, and almost none of that work was seen twice. So a small team built a place where everyone got that setup. It became Glass, the AI workspace where knowledge compounds. Instead of starting with an empty chat, everyone starts from what colleagues have already figured out and chosen to share.

Here's what you get with Glass:

• Your company's knowledge: connect the tools your team already uses
• Shared workflows: the reusable processes from step 7, saved alongside your team’s context in shared projects.
• Multiplayer: bring a teammate into a session, so a support agent and an engineer can finish a ticket together
• Any model: choose from multiple providers, so when a better one ships next week, you just use it

Feature requests from customer calls used to be copied into Salesforce by hand. Then an account manager saved a prompt to Glass that pulls them straight from the calls. Now that prompt captures them for the whole team, another prompt prioritizes them, and a third emails customers when what they asked for ships. It started with one person saving one prompt.

You can now request early access to Glass and give your team's best work a place where everyone can build on it.