Summary: Every enterprise asks the same question about agent autonomy: how do we know when it is safe? Vendor benchmarks do not answer it, and even “accuracy” begs the question of who defined the right answers. The metric that actually works is humbler: edit distance, how much your own reviewers change the agent’s output before accepting it. This post makes the case for edit distance as the standard, and shows how the Audit, Assist, Automate ladder turns it into a governance mechanism.
The question accuracy cannot answer
Accuracy presumes an answer key. For classification tasks you can build one; for judgment work (a credit memo, a claim determination, a SAR narrative) the answer key is your senior reviewers’ judgment, which is exactly the thing you are trying to automate. Benchmarks measure the model against someone else’s answer key, on someone else’s cases. None of that tells a compliance officer whether this workflow, on these cases, meets this institution’s standard.
Edit distance sidesteps the philosophy. Put the agent’s output in front of the people who own the standard, and measure what they change. Nothing to argue about: either your experts accept the work as-is, or they fix it, and the fixes are countable.
Why the metric is trustworthy
Three properties make edit distance the right backbone for autonomy decisions:
- It is generated by the work itself. No eval harness, no synthetic test set. Every reviewed case emits the signal for free.
- It measures your standard, not a benchmark’s. House style, institutional risk appetite, and the unwritten rules live in reviewer behavior, and edit distance captures all of them.
- It trends. A single week means little; a quarter of near-zero edits on a case category is evidence a regulator can follow. The trend also runs the other direction, flagging drift after policy changes or new document types before it becomes an incident.
The metric has one honest requirement: reviewers must actually review. Rubber-stamping poisons the signal, which is why the review workload has to be small enough to do well. That is an architecture property: systems that route only genuine exceptions to humans keep review real.
The ladder: Audit, Assist, Automate
Edit distance powers a three-stage autonomy ladder that turns “do we trust it” into a sequence of measured gates:
Audit. The agent processes real cases end to end; humans keep making every decision. Output is compared to human decisions case by case. No risk is transferred, and the baseline edit distance is established on live volume. This is how a 60-day deployment spends its sixth and seventh weeks.
Assist. The agent prepares the decision and its evidence; reviewers approve, edit, or override at review gates. Every edit is logged with the case context. Categories where edits approach zero become candidates for the next rung; categories where reviewers keep intervening stay put, and the edits point at exactly which policy needs refining.
Automate. Clean-case categories execute end to end, exceptions still route to humans, and the metric keeps running. Autonomy is never global: it is granted per category, held by evidence, and revocable by the same dashboard that granted it.
The ladder’s power is who it convinces. “The vendor says 99%” persuades nobody in a supervisory meeting. “Our own reviewers changed 0.4% of outputs in this category over 90 days, here is the log” is a defensible position, with a why-trail under every case in the sample.
The compounding loop
Edit distance is also the improvement engine. Each edit is a labeled example of the standard the agent missed; refined plain-English policies absorb it, get backtested against history, and deploy. Edits shrink, autonomy expands, and the loop runs again. On MightyBot this is the progressive autonomy model in production: autonomy raised on evidence, not optimism.
If you are evaluating platforms, ask one question: what measurement justifies moving a workflow from review to automation, and can I see it? A platform without an answer is asking you to decide on faith. The measured alternative exists.
