Governed Context: What Four Teams Learned Building a Company Brain in 2026
•
18
MIN READ
AI Summary by Fellow
Most teams that try to build a company brain hit the same wall. The context they need is spread across Slack DMs, personal docs, five versions of last year’s positioning and hundreds of recorded meetings nobody can search. Point an AI agent at all of that and you get confident answers built on stale, conflicting or sensitive information.
On November 19, 2026, Fellow hosted a live panel, “Governed Context: Building a Company Brain You Can Trust,” with four people building one right now. In a poll at the start of the session, 52% of attendees said they were exploring the idea of a company brain, 28% were piloting one and 20% were already building one (a small, self-selected webinar audience, but a useful snapshot of where teams are).
Each panelist described a different architecture. They still landed on the same conclusion: culture and data quality come before tooling, and governance has to be designed in from day one. This recap walks through what they built, what they would do differently and how to build a company brain your security and compliance teams can sign off on.
Who was on the panel?
The conversation was moderated by Manuela Bárcenas, Head of Marketing at Fellow.
Panelist | Role | What they’re building |
|---|---|---|
Joe Stych | Senior Lead, Agentic GTM, Zapier | Marketing OS, a shared brain for go-to-market teams, plus a revenue operating system for sales, success and support |
Shane Pahl | Director of Business Design, FinStrat Management | Per-client brains and a central company brain for a firm providing fractional CFO-led finance and accounting to investor-backed startups |
Arturo Delgado | Chief Business Officer, Arkkosoft | A governed knowledge base fed by an AI center of excellence at a software engineering firm that builds AI agents |
Samuel Cormier-Iijima | CTO, Fellow | Fellow’s internal company brain and agent and the governance layer inside Fellow’s product |
What is a company brain?
A company brain is a shared, governed layer of company context that both people and AI agents can query and act on. It combines the documents and decisions a company treats as true, the live signals coming out of meetings and business systems, and rules about what agents are allowed to do with all of it.
Joe described Zapier’s version as “a collection of context that we keep fresh,” plus “a collection of tools that the agents can actually reach for and then go do work in your apps,” all sitting under a governed layer. In practice, it has three parts:
Canonical context. The foundational documents a team treats as the source of truth, such as positioning, company strategy and playbooks for going after certain deals.
A signal layer. Real-time inputs, which at Zapier include “meeting notes from Fellow or actions that happen in HubSpot” and product metrics. Agents use these signals to “suggest updates to the canonical context and then kind of improve over time.”
A governed action layer. Controls that make sure people and agents use the correct context and “only using the sorts of actions that you would want them to.”
Samuel added that a company brain is both the knowledge and the agent you ask: Fellow’s internal brain is something employees talk to directly. Better models have also simplified the build. Instead of ingesting everything into one retrieval system, agents can now reach into connected tools themselves. If your tools are already connected to your own systems, he said, “To me, that is a company brain already.”
Why are teams building a company brain now?
The panelists started for business reasons, not because AI was the goal. Three motivations came up.
Accountability for what gets said in client meetings. “People say now AI first. I don’t believe in it,” Arturo said. “It’s maybe data first, and not even then. It’s business first, right?” At Arkkosoft, decisions and action items from client calls were falling through the cracks because nobody owned them. If AI is going to execute on those commitments, he asked, “who owns it, right? And who’s gonna be liable for it?”
Moving fast without adding risk. FinStrat competes with AI-native accounting services while handling data tied to fundraising rounds and exits, so compliance comes first. Shane described the challenge as figuring out “how do we not lose all this really important valuable qualitative information that we’re collecting constantly on all of these video calls” and getting it in front of CFOs compliantly. A side benefit: new CFOs and employees can query past meeting intelligence to understand a client’s gaps, opportunities and sentiment instead of waiting on a handoff.
Pooling the best agents in one place. At Zapier, people were building their own agents and workflows without sharing what worked. The shared brain changed that: “whoever writes the best copywriting agent or the best outreach agent, we’re going to have that kind of in a central place and all working from that same context,” Joe said. He also sees it as preparation for agents that sit in the org chart. “Making that stuff kind of visible to them and not keeping it in your own personal brain is just going to set you up much better.”
Lesson 1: Why does a company brain start with culture, not tooling?
An agent can only learn from work that is written down in a shared place. If decisions live in private messages, no tool will find them.
“A lot of this is cultural, right?” Joe said. “If you’re not writing things down or you’re not working together outside of like Slack DMs or individual emails, then the agent’s just not gonna be able to pick up on it and learn from it.” Zapier has had “default to transparency” as a value for its whole history, which gave it a head start. His advice for everyone else: “I would start to do the cultural shift part now to just say, hey, we should be writing things down.”
Fellow made a similar call with its internal company brain. “The stance that we took with [our internal company brain], and it kind of fits in with how we are a culture as a company, is that everything would be public, or public within our company,” Samuel said. Weekly town halls and dev weekly meetings feed into it. Sharing broadly by default does not mean sharing everything, though: “You need to have ways to govern this information and provide just the right context to the right people.”
FinStrat got there by changing how it plans. Every new tool download triggers a compliance review, which showed Shane how many people were pulling in different directions. So the company retreat opened with one question: “What is our goal as a company?” Now promising work moves from weekly departmental show-and-tell Fridays to a monthly all-hands, then into cross-functional scoping, training and change management.
Lesson 2: What should you put in a company brain first?
Start with the handful of documents a new hire would need on day one, consolidate every version into one, and assign an owner to each. Joe framed the test as: “If I was hiring a new marketer, what are the like 6 to 10 things I would want to hand them on day 1?”
Zapier used to refresh its positioning at two big launches a year, so “there would just be a ton of these documents just floating around that we never retired, we never updated.” Here is how the team seeded its brain:
Pick the core documents. Zapier's Marketing OS started with five, including the ideal customer profile, positioning and company strategy.
Gather every version. Collect the docs and research people had written on each topic.
Consolidate with an agent. Use an agent to answer “what do we think are the things that we truly believe about our positioning?” and find the freshest version.
Get leadership to ratify it. Present the synthesis to the leadership team as the new foundation.
Retire the old docs. Delete them or point them to the new one.
Assign owners to keep it fresh. “From here on out, this is our garden that we need to keep weeded and fresh and growing things from.”
There is a counterweight. Joe’s first mistake was “being too strict about the content that comes into the brain.” Models are now much better at judging what is recent and relevant. “Don’t be too precious about what makes it in. You’ll start to see if it makes mistakes, right? So you can kind of retire things that don’t work.”
The two pieces of advice fit together: keep the canonical layer small and curated, and let the signal layer connect broadly, pruning based on the mistakes you actually see.
Lesson 3: Who should own the data in a company brain?
Every piece of context needs a named owner who keeps it current and decides who can see it. Without owners, a company brain fills up with stale data and nobody can explain where an answer came from.
At Zapier, each canonical document has an owner who updates it when positioning changes and approves new information before it goes in. Arturo takes the idea further: “It’s almost to the data point has to have an owner. And that owner might say who’s gonna see it and who’s not.” Picture a dashboard of 20 data items showing who owns each one and who has access. Once that exists, he said, layering an AI copilot on top is easy because every answer can be traced back to its origin.
Arkkosoft also runs an AI center of excellence. It doesn’t dictate what people build (“you cannot cut innovation”), but each initiative gets checked for licensing, ownership, data lineage and data stewardship before it becomes part of the company brain. “You don’t want shadow AI as well, right? Because it’s very dangerous.”
Arturo’s other rule is to measure before you build. If the data is unstructured, stale or unreviewed, “it might be trash in, trash out,” and you end up making bad decisions in front of a client. Get processes and data ownership mature first, record a baseline, then check whether the company brain actually improves on it.
Lesson 4: Why do AI agents need guardrails before they get autonomy?
Agents act faster than any person can supervise, so permissions have to be set before an agent acts, not reviewed after.
“The reality with agents that we’re seeing is that they work at superhuman speeds,” Samuel said. “The timescales aren’t human timescales anymore.” If an agent is emailing or changing records, you need standardized sources of truth so agents aren’t stepping on each other, “because otherwise, by the time that they’ve done the action, it’s already too late.” Once strict controls are in place, “then you can let them loose and have the confidence and the peace of mind that they won’t go run amok.”
Zapier handles this at two levels. Context is governed by document owners. Actions are governed per tool, per account and per action. With HubSpot, for example, IT can let agents read account and deal data while blocking anything that would change or delete it. As Joe put it: “You can read about the accounts here, you can look at the deal data, you are not allowed to change it or delete it.”
The same principle applies to meeting data. With the Fellow MCP Server, a workspace admin has to turn on MCP connections before anyone can use them, then switches each tool on or off, for example allowing read tools like searching meetings or getting action items while keeping agenda-editing tools off. Assistants connected through MCP only see data the user already has access to in Fellow, and every change to those settings is recorded in the workspace audit log.
Lesson 5: What security questions should you ask before connecting a tool to your company brain?
Ask whether the vendor trains on your data, how long it keeps it, who can see it, and whether you can trace every answer back to its source. Between them, the panelists covered a full vendor checklist:
Question to ask | Why it matters |
|---|---|
Does the vendor or its AI providers train models on our data? | Arturo wanted a self-contained environment “where the data is not going to be used to train the world, so to speak.” |
Can we trace an answer back to its origin? | Traceability is what makes a company brain auditable. |
How granular are access controls? | According to one of the panelists, "many decisions are made that not everyone needs to know about." |
What retention settings exist? | Retention decides how long sensitive content stays exposed. |
What authentication and authorization features does it have? | Agents inherit whatever access the connection allows. |
Can we set policies on what flows into external systems? | Context leaving approved boundaries is the core risk. |
Which certifications does it hold, and is there a trust center? | SOC 2, HIPAA and ISO AI certifications simplify approval decisions. |
Does it work with our compliance monitoring? | One of the panelists mentioned Vanta to flag unauthorized connections and data leaving approved systems. |
Has a third party reviewed the risk and contract terms? | Regular outside audits catch regulatory exposure you would miss. |
Shane’s advice is to pair tools “that have connectors, and they have a trust center that you can go review their compliance documents.” He also brings in third-party auditors regularly to review new initiatives. “It’s not the fun part of building a company brain, right?” he said. But it keeps the brain healthy, and “you can build fast and you can build compliantly. And I don’t think that they’re necessarily in opposition to each other.”
On self-hosted versus cloud-hosted brains, Samuel was candid: “I don’t think there’s a right answer there.” Running your own makes sense if you have the engineering team to maintain it and set up evals. If you use a vendor, check SOC 2 and the other safeguards you’d expect from any external provider.
Lesson 6: Why does a company brain need to forget?
A company brain has to delete what the source system deleted, or your retention and redaction promises stop being true. This came from an audience question, and Samuel’s answer was one of the most important of the session: “Forgetting is as important as remembering in a lot of cases.”
He described the ideal as bitemporal or even tritemporal memory, where the brain records “not just this is a fact, but when did you know about this? When was it valid? When did you forget about it?” In plain terms, every fact carries three timestamps: when the brain learned it, when it was true and when it was removed.
The governance risk is copies. If a transcript is deleted or redacted in the original system, “which with Fellow, we do that for a lot of customers,” the deletion has to reach the brain too. “You need to have that filter back into the brain and remove that as well, because otherwise you’re not getting those same guarantees about what context is available.”
In practice, map every place meeting content gets copied (knowledge bases, vector stores, agent memories, exports), have agents query context at the source instead of keeping their own copies where you can, and test deletion end to end by redacting a test transcript and confirming it disappears downstream.
On the source side, Fellow lets admins set separate retention periods for recordings, transcripts and AI notes, including zero-day retention, which deletes recordings or transcripts once processing is done while keeping the AI notes. Transcript redaction permanently removes selected sections of a transcript and its recording. Both are Enterprise plan features.
Lesson 7: How do regulated firms keep client context separate?
Use separate brains for separate obligations. FinStrat runs a per-client brain for each client plus a central brain for company performance and objectives. Each client brain is limited to that client’s data, under signed agreements that spell out what FinStrat retains and how it can be used.
For sensitive and in-person conversations, the work happens after the meeting too. “If someone says, can we please like remove this from the transcripts and recording, we go and do that, no problem,” Shane said. FinStrat also uses document management tools that flag possible PII a person may have missed in a transcript so an admin can review it.
The same structure applies to law firms, private equity teams, hedge funds and investment banks, where client walls and information barriers matter as much as speed.
Why do company brain pilots stall, and how do you get one to production?
Pilots stall when cost, latency and data quality only show up at scale. Arturo pointed to research finding that about 88% of AI pilots never reach production. (The figure comes from IDC research with Lenovo published in 2025.) A pilot “runs well, you know, on static data, only with 5 users,” then real data volumes arrive. “Oh, cost. Nobody talked about cost.”
His fix is a staged path with a cost and effectiveness baseline at every gate:
Explore. Test whether the idea is feasible at all.
Prototype. Build a pilot or MVP against a measured baseline.
Operationalize. Add real data and users, and watch token cost and latency.
Scale. Expand only once cost and effectiveness hold up.
Samuel would also design for cost and flexibility. Fellow’s internal company brain started as one monolithic system, but “people are going to bring their own tools. They want to use different agents.” He would make it modular, track what each task costs, and route simpler work to cheaper models. Joe’s addition: build with internal design partners rather than in isolation, and add a feedback mechanism to everything, because “agents are super good at just combing over that feedback and then improving based on it.”
Where do meetings fit in a company brain?
Meetings are the signal layer: the place where decisions, commitments and customer context first show up, and where some of a company’s most sensitive information is said out loud. Three of the four panelists named Fellow meeting notes as part of their company brain.
Fellow is the most secure AI meeting assistant and notetaker for enterprise organizations and regulated industries. It turns every meeting into one shared, searchable, governed record that people and agents can use. Here is how it maps to the panel’s lessons:
For teams building a shared company brain, the distinction between a personal and collaborative AI meeting note taker matters: meeting context needs to be available to the right people, not siloed with one person.
Capture the signal wherever meetings happen. Record with the Note Taker in Zoom, Google Meet and Microsoft Teams, use botless recording from the desktop app on any platform, or record in-person meetings with the mobile app. Every meeting produces a transcript, AI meeting notes and action items, so commitments have owners from the start.
Make it queryable. Ask Fellow searches your meeting history in natural language, and the MCP Server and integrations bring meeting context into the assistants and systems your company brain already uses.
Control who sees what. By default, only internal meeting attendees can access a recording, transcript and summary. Recordings live in a permission-controlled recording library, and password-protected recap links keep external sharing in check.
Govern capture and retention. Configurable data retention, zero-day retention, transcript redaction and consent capture for auto-recorded meetings are available on the Enterprise plan. The compliance portal gives compliance teams one place to search, review, flag and audit recorded meetings.
Pass security review. Fellow is SOC 2 Type II certified, HIPAA compliant and GDPR compliant, and never uses your meeting data to train AI models.
As Manuela put it before the Q&A: give your brain the right context with the right boundaries, “and therefore you’ll get the context without the liability.”
A company brain checklist from the panel
Use this as a starting sequence:
Make writing things down in shared spaces the default before buying any tooling.
Record a baseline for the decisions or workflows the brain should improve.
Pick five to ten canonical documents a new hire would need on day one.
Consolidate every floating version with an agent, ratify the result with leadership and retire the old docs.
Assign a named owner to every document and data source.
Connect signal sources like meetings, CRM activity and product metrics, then prune based on observed mistakes.
Set per-tool, per-account and per-action permissions before agents act on their own.
Vet every vendor for training policy, retention, access controls, certifications and traceability.
Make sure deletions and redactions in source systems reach every copy in the brain.
Move from explore to prototype to operationalize to scale, checking cost and effectiveness at each step.
Conclusion
The four teams on this panel built very different company brains, but the order of operations was the same: work in the open, curate a small canonical core, name owners, lock down agent actions and make sure the brain can forget as reliably as it remembers. Tooling came last, and governance came first. If meetings are going to be your richest signal layer, start with a secure AI meeting notetaker that gives you the context without the liability.
Frequently asked questions
Record, transcribe and summarize every meeting with the only AI meeting assistant built with privacy and security in mind.






