The week ChatGPT changed every meeting we had
In fifteen days we went from explaining what a language model is to explaining why it can't reach your ERP.
The public demo did a decade of commercial work in a month. It also created impossible expectations. Every leader suddenly wanted an AI assistant, and most imagined it working on their company data without understanding what that actually meant. We spent a week explaining the gap between the demo and production.
The speed of iteration was humbling: nobody had seen a system adapt across conversation turns so fluently. The phrasing, the tone, the logical flow — all of it matched what people had assumed an AI assistant should be. But matching the assistant to a specific company's constraints was a different problem entirely.
The useful conversation with leadership starts by separating what the model knows from what the company knows. The model knows English and a cut of the internet up to September 2021. It doesn't know your sales pipeline, your product, your client names or your internal policies. That distinction matters profoundly.
We had to ask hard questions: what's the return if we connect this to our data? Do we need the model to know our facts, or just to help users find them? Is this a customer-facing tool or internal? Each answer led to a different architecture.
Everything we've built since lives in that gap: connecting the model to your data, with permissions and with verification. The assistant needs to know what it doesn't know and ask a human. And the human needs to see exactly what data went into the answer.