Do we need a Company Brain or Nervous System?

AI models are already intelligent enough!

So why is it still so hard to deploy AI inside a company?

Memory and context.

YC's latest request for startups called for a Company Brain and Jack Dorsey has spoken recently about the idea of building a company as an intelligence. Everyone is circling the same question:

How do you give AI enough company context to actually do useful work?

Today, this is my main focus at Flare and the current situation is summed up neatly with this analogy:

Imagine hiring an accounting genius and asking them to do last month's bookkeeping for Client X.

No level of IQ could tell you who the hell Client X is, how they like to see their books, what exceptions were agreed last quarter, or what the CFO was promised in the last email thread.

That's the gap.

But calling the solution a 'Company Brain' feels like it's skipping a step to me.

A brain has sensory inputs, memory, motor outputs and feedback loops. Most companies today have the equivalent of sensory organs scattered everywhere: Slack, Xero, Google Docs, call transcripts, spreadsheets, emails, Whatsapp messages, dashboards.

Before a company can have a brain, it needs a nervous system that connects all these organs together. It needs to sense what is happening, structure it, route it, and check whether the right actions were taken.

Whatever you call it, documentation is becoming a prerequisite for unlocking the true benefits of AI.

However, documentation alone isn't enough.

There's still an AI deployment bottleneck within the enterprise (typically better documented companies). Look at OpenAI's Deployment Company announcement this week which confirms that unless the documentation is executable, it's useless.

YC puts it nicely in their Summer 2026 request for startups:

"A system that pulls knowledge out of all these fragmented sources, structures it, keeps it current, and turns it into an executable skills file for AI."

The challenge is that most company context still lives inside employees’ heads.

Call it intuition, experience, judgement or remembered preferences. It's the stuff that lets someone know how a client likes to work, how a particular process actually runs, when to escalate something, what “good” looks like, and which edge cases matter.

That context needs to get translated into SOPs, workflows and controls that AI can reason over to understand exactly what and how they should be operating inside a specific company.

This is particularly important for services firms.

Services companies are basically bundles of undocumented human judgement where the product is not code, it's thousands of tiny decisions made by humans every day.

For example: in accounting, the hard part is rarely “can the model understand a bank transaction?”. The hard part is “does the model know how this specific client wants revenue recognised or which Slack thread explained the exception?”.

This is what we're currently tackling at Flare where one of our biggest insights so far has been that your context needs controls.

We've started writing SOPs for our most critical processes, not only to standardise human behaviour, but to build a foundation our AI agents can operate on.

Of course, we're using AI to draft the SOPs but the critical part is that every SOP must have a corresponding 'Control' - an automated check that confirms whether the SOP is actually being followed.

For example, if we have an SOP for how a monthly bookkeeping close email should be written, the corresponding Control might be a dashboard that flags emails that do not follow the required structure.

So we're not just documenting how work should be done, we're creating a system that checks whether the work was actually done that way.

An SOP without a control is just a wish. An SOP with a control becomes part of the operating system.

Controls do two things:

1. Improve the quality of our company output, driving customer satisfaction.
2. Make it easier for agents to pickup work, because the agent now has a feedback loop. Tobi from Shopify actually has a fantastic name for this: a fitness function.

This is the part I think most people miss - the “Company Brain” isn't just a chatbot over your company documents.

Becoming AI native means your company must be queryable, executable and self-correcting.

To us at Flare, that means building documentation in the form of SOPs and the associated controls to provide a fitness function for agents to learn against.

The real question is not whether the models are smart enough.

It actually boils down to whether your company is legible enough for intelligence to operate inside it. The companies that nail that will be able to deliver services as if they were software.