AI for business is broken. Not because the AI isn't smart enough — it's brilliant. It's broken because the AI doesn't know your business. And every approach so far has hit the same wall.
The wall is this: using AI for real business work means re-explaining your business every time. We're a counter-crafting firm. Our top product is X. Our pipeline looks like Y. Our compliance constraints are Z. Rey is the decision-maker. The Anders deal stalled because of price, not fit. Our voice in customer emails sounds like this, not like that. You type the preamble. You attach the PDFs. You re-orient the model. Then you ask the question.
This isn't a UX problem. It's the bottleneck.
The smartest AI on Earth, dropped into a business that hasn't organized its context for AI, produces useless work. Generic emails in the wrong voice. Proposals priced for the wrong market. Recommendations that ignore the relationship you've spent five years building with that supplier. The model is operating in a vacuum where context is supposed to be.
Context is the bottleneck. Not intelligence.
Everyone has tried the wrong fix.
The market has thrown four kinds of solutions at this, and all four fail for the same reason.
Knowledge bases (Notion, Confluence, internal wikis). Human-readable, AI-illegible. Pages are walls of text the AI has to re-read and re-interpret every time. No state. No workflow. No skills. No model of the business — just a model of some documents.
RAG / vector search. Retrieval is not memory. RAG finds the doc that contains the answer. It doesn't know whether the doc is still current, whether the policy got overridden last quarter, whether Rey approved the exception in Slack. Snapshots, not state.
Custom GPTs and Claude Projects. A folder of files plus a system prompt. Static. Single-user. No integrations. No persistence across conversations. A bookmark folder for a chat. Stale by Tuesday.
Vertical AI agents (sales, support, legal). Each owns one job. None owns the brain. The moment work crosses functional boundaries — and almost all real business work does — the agent hands off to you.
What's missing across all four is the actual primitive: a live, structured, audited brain of your business that any AI can act on. Not pages. Not a search index. Not a chat with attached PDFs. Not a vertical bot. The brain.
The protocol layer is being handed to us.
Here's the part most people haven't noticed yet: the frontier labs are publicly handing this layer to third parties.
Anthropic shipped MCP — the Model Context Protocol — explicitly to let external systems be the context layer for Claude. Bring your own context. We'll consume it. OpenAI shipped the Apps SDK with the same intent. Google with Extensions. Microsoft with Copilot connectors. Every major AI vendor is sprinting to make it easier for someone else to be the brain — because they want every business's context flowing into their model, but they don't want to be the company that organizes it.
That's a strategic choice, and a permanent one. Building the brain requires multi-tenant state machines, ETL from fifty third-party tools, policy engines, audit logs, version history, conflict resolution, vertical-specific opinionation for every industry. That's infrastructure work, not model work. Different company shape. Different P&L. Anthropic isn't going to wake up tomorrow and decide to compete with Salesforce, Notion, and Snowflake on top of being a frontier lab. They built the protocol so they wouldn't have to.
An entire layer of the stack with a protocol already in place, an obvious demand signal, and no one named in the slot yet.
What a workspace brain actually is.
A workspace brain is the layer between your business's raw operational reality and any AI that wants to act on it. It has four properties no current product combines:
- It's structured for AI, not for humans. Shaped like a system: inputs, state, processes, outputs, feedback, control. The AI doesn't have to re-derive how your business works every time it's asked to help. The structure is the help.
- It's alive. Continuously ingesting from the tools your business already runs on. Gmail. HubSpot. Notion. GitHub. Stripe. OneDrive. The state of your pipeline today, not the state when someone last updated a doc.
- It's multi-agent-ready. Cruma's own agent runs on it. So does Claude, if you plug it in. So does ChatGPT. So does the custom agent your engineering team will build next year. One brain. Every AI.
- It's auditable. Every action every agent takes on your business is logged in a single trace. When the model gets it wrong — and models will keep getting things wrong — you can see what happened, override it, and the brain learns.
That bundle of properties is a new category. The name we've planted on it is: the workspace brain.
→ Mechanical deep dive: how Cruma assembles context vs how Claude does
The smarter AI gets, the more useful your Cruma gets.
Here's the part that makes this a real bet, not just a feature.
The smarter models get, the more valuable your workspace brain becomes. Frontier models are getting better at consuming external structured context — that's literally what MCP and Apps SDK are optimizing for. So every time Claude or ChatGPT improves, the brain you own grows more valuable, because better models do more useful work on the same structured input.
This is the opposite of how most AI tools age. A chatbot built on GPT-4 gets less valuable when GPT-5 ships — users want to talk directly to the new model. A workspace brain gets more valuable when GPT-5 ships — the same brain now runs better workflows.
You don't compete with the model. The model makes you more valuable.
That's the rare property of infrastructure: it compounds with its consumers' progress instead of being commoditized by it.
What Cruma is doing.
Cruma is the workspace brain. Concretely:
We pre-build Spaces — structured systems shaped to how a particular kind of business actually runs. A consultancy doesn't get a blank page; it gets a Space pre-shaped around proposals, follow-ups, pricing memory, won/lost patterns. A manufacturer gets one shaped around suppliers, pipeline, R&D, logistics. The Space is the unit of sale.
We ship a cockpit — the web app where you configure your Space, audit what your AI did this week, and override what you don't like. The cockpit is the trust artifact. Not the workspace.
We ship a notch — an ambient surface in your menu bar. Glanceable status. One-keystroke approval. Push notifications from the Space. This is where Cruma lives in your day-to-day.
And we expose the brain on an open protocol. Plug Claude in. Plug ChatGPT in. Plug your own agent in. The brain is yours. The agents are interchangeable.
The bet underneath all of it: when the agent wars settle, the company that organized the world's business context will be infrastructure for every AI in the category. Whoever named that category will own it.
The workspace brain.
Cruma builds it.