The 2026 Assembly Index: what we found

This year’s Index draws on anonymised run data from 210 organisations. The headline is the silent correction rate: across all agent categories, roughly a third of completed tasks are amended by a person before use.

The detail

None of those corrections were reported as failures. In most cases there was no feedback mechanism at all — the person simply edited the output and moved on.

The rate varies enormously by task type. Summarisation sits near 22%; anything touching numbers or policy sits above 40%.