Pull what has a system of record. Persist what does not. A folder is not a memory, and most of the knowledge that matters has no system of record at all.

MCP answers how an agent reaches data. The brain answers what the organization knows. Pull facts of record. Persist judgment that has no system of record.

The software meltdown isn't AI killing SaaS. It's Christensen's Law of Conservation of Attractive Profits playing out in real time. Value is migrating from the interface layer to the context layer.

Karpathy described a personal LLM wiki. We have been running the multiplayer version since the start of the year.

What does it actually mean for a document to 'decide' something? The four components that give an invoice or contract genuine agency.

Finance leaders are deploying AI agents like they're adopting teenagers. Curfews, check-ins, no real decisions. Then they wonder why the agents don't deliver.

Everyone agrees on where AI agents are going. The destination is not the differentiator. The route is. Constraint profiles, a task vocabulary, and democratized creation.

Most finance teams deployed AI and got incremental results. The problem isn't adoption. It's that no one specified which kind of AI to adopt. There are four, not one.

We've been thinking about AI in finance backwards. The end state isn't smarter agents processing dumb documents. It's documents that process themselves.

Sixty percent of enterprises require human oversight before AI agents access sensitive data. It feels responsible. It's a placebo with a 24-month shelf life.