AI-assisted automation

Which work gets automated?

Automated work falls under three headings: reading and summarising a large number of official sources, classifying incoming content or enquiries, and producing a first draft of a text. What they share is that the result is used only after a person checks it.

• Scanning: reading official sources and competitor pages on a schedule, extracting what changed.
• Classification: separating an incoming enquiry by subject and urgency.
• Drafting: producing the first version of a text, never the final one.
• Comparison: extracting the difference between two documents.

All four are things a person can do and which grow expensive through repetition. What automation buys is not a new capability but the same work done far more frequently.

How is output verified?

Verification runs on one rule: every numerical claim that reaches publication is traced back to its source and the source is read. A reference the model supplies is opened and checked even when it looks right, because naming a nonexistent source is known behaviour.

Every page on this site carries the same arrangement: a Sources block at the foot holding the claim, the publisher and the date it was checked. Where a figure is carried across, its date is carried with it, because an undated figure goes wrong quietly.

What is the cost measured against?

Cost has two parts: the one-off build and the model usage that accrues while the automation runs. The second follows the volume of text processed and every call is logged, because an unmeasured automation never announces that it costs more than it saves.

That is why every automation built here carries a budget ceiling and a log. When the ceiling is reached the work stops; when it is not, what surfaces is an invoice months later.