What you know is what you own.
When every business has access to the same models, the only durable advantage is the way your business actually works: the corrections your reviewers make, the edge cases your team has learned, the judgement calls that nobody wrote down.
That tacit knowledge is the moat. Most of it goes uncaptured.
Your moat as a company is your tacit knowledge. In a world where AI exists, and network effects of AI exist, you need your own hill-climbing machine in which the models are learning.
Every correction is a private eval.
Your people approve extractions, reject attributes, and override models every day. Each of those decisions is a private benchmark: a clear signal of what "good" looks like for your business. Today it's almost always lost.
The reviewer makes the call.
Someone who knows your business sees what the model proposed and adjusts it.
The decision is captured.
The reviewer's change becomes part of the content object's audit trail: who, what, when, and why.
The signal you didn't know you had.
Each correction tells you exactly what the model got wrong, on real work in your context. It is the benchmark.
TODAYWhere the signal goes
Spreadsheets. Email threads. A backlog of tickets. A senior reviewer's head. None of it flows back into how the next document is processed.
TODAYWhat it costs
The same mistakes get corrected every week. New hires re-learn the same edge cases. The model that gets retired didn't know what your best reviewers knew.
Hill-climbing, every day.
The audit captures every correction. The system distils the patterns, proposes new knowledge, and humans confirm it. The next run is smarter on your data, in your operation. There is no staging environment and no pre-deployment phase: the improvement happens inside the work itself.
The work runs
A document moves through the workflow. Humans review where they're needed.
Corrections land in the audit
Every approval, override and edit is captured against the content it was made on.
The system distils patterns
An agent watches the corrections, finds the patterns, drafts the lesson.
It proposes new knowledge
"Vendors of this shape get normalised this way." A small, reviewable change.
A human confirms
The right person sees the proposal in context and accepts, edits, or rejects it.
Knowledge attaches
It joins the workflow's knowledge set, ready for the next document.
The eval is the operation itself.
There is no "eval set" sitting on a wiki, refreshed quarterly. Every document the business processes contributes to the benchmark, and every reviewer is part of the panel. The model improves on the work as it gets done.
"Today's failure cases are informing you to change the benchmark continuously. It's not a static thing."
Satya Nadella · CEO, Microsoft