What we measured
212 runs of recurring work from nine design-partner teams between 1 and 26 September 2026: investor updates, month-end reconciliations, churn follow-ups, contract reviews, metric digests and lead triage.
How a run was scored
Each output was compared with the version a person on that team signed off by hand for the same week. A run counts as correct only if it needed no edits to the facts; tone edits were allowed and are counted separately.
Who we compared against
A single general-purpose agent given the same tools and keys, a scripted RPA flow where the team already had one, and each team’s own first week doing the task manually.
What this does not cover
Tasks that need judgement calls nobody wrote down, tools without an API, and anything we have not run nine times. Teams chose which workflows to include, so the sample leans toward work that was already well defined.
Reproduce it
The scoring rubric and anonymised run ledgers are available on request to anyone evaluating Bottega. Write to evals@bottega.ai.