Six months to hours
The number is the headline, but the number is not the interesting part. What made it work was a validation programme run with the specialists who sign the document.
A research body produces and maintains the guidelines of its field: the reference documents its decisions are then made against. Keeping them current is a research task in its own right, and it took about six months of expert time to produce an updated guideline.
They had already proved the idea worked. Someone had done it by hand in a desktop AI tool: research a single topic, gather the sources, draft the sections. It was slow, unrepeatable and impossible to audit, but it demonstrated that the hard part was tractable.
From a demo to a system
We rebuilt it as a production agentic loop on the Claude API, running over a retrieval layer of the existing guideline corpus, with the surrounding infrastructure that turns an experiment into an operation: queue-driven jobs, a document store, structured outputs, and a path for a human to intervene at defined points rather than everywhere.
It now produces an updated or new guideline in hours. That is roughly three orders of magnitude, and it is the kind of number that makes people suspicious — reasonably so, because on its own it means very little. A fast pipeline that produces a document the experts will not sign is worth nothing.
The feedback of the experts who validate the document is what made it good. Not the model.
The half that took the longest
The validation programme ran for months. The body's reviewers read outputs, marked what was wrong, and explained why — not as a one-off acceptance test but as a continuous loop that changed the pipeline and the quality metrics behind it.
That is where the real design happened. What counts as a complete section. Where a citation is required and where it is optional. Which claims may never be produced without a source. How a decision tree should be represented so a specialist can check it at a glance. None of that is in a model card; all of it came from the people who put their names on the document.
Why it generalises
Every organisation with a slow, expert, document-heavy process has a version of this problem, and most of them are now being sold the six-months-to-hours part without the other half. The pipeline is a few weeks of work. Establishing what "good enough" means, with the people who are accountable for it, is the project.
The engagement later expanded into the organisation's operational core — a multi-agent orchestrator with human checkpoints across event operations. That expansion is the real proof the first system worked. Nobody hands you their operations because the demo was fast.