Phased Q3 delivery.
Queryable by any person or agent we authorise. Served to no external LLM. Portable across model providers — if tokens spike or a vendor dies, we move.
A surface is anywhere company information sits — the warehouse, Slack, the support desk, the codebase. Each one has to be wired up before agents can use it. These are the ones that matter, and where each stands.
Every agent logs what it did and the time it saved into one analytics store, at the moment it does the work. The same method applies to every workflow we deploy, so usage, outcomes and cost are queryable at any point — the readout comes from the data, not from anyone's impression of how it's going.
Each shipped step becomes infrastructure for the next, so the value we're getting at each step compounds. Avoiding a big bang release is the clear choice.
An agent can only use a table if it has been granted access and the table says what its columns mean. Without that, the AI guesses — which is why Ben only trusts answers that come from the documented part. Phase 1 grants the access and writes the descriptions for the data that matters most. From then on, an agent checks that every new source lands labelled and accessible, so coverage keeps climbing as the warehouse grows instead of decaying. Two more fixes ride along: one login pattern so agents can run on a schedule in the cloud rather than on laptops, and moving the systems that currently sit on personal machines into company ownership.