What is Proof-Adjusted Autonomy?

Proof-Adjusted Autonomy (PAA) is the share of an organisation's work an AI agent completes autonomously AND supports with complete, independently validated, timely evidence. PAA = P(A) x P(C|A) x P(R|A,C) x P(T|A,C,R).

How autonomous is a 90% autonomous AI agent really?

With 95% complete evidence, 80% independent validation and 90% timeliness, a 90% autonomous agent is 61.6% Proof-Adjusted Autonomous. Autonomy is not what an agent does; it is what it can prove.

Proof-Adjusted Autonomy Calculator

Interpretation bands (45% and 70%) are the author's illustrative heuristic, not a validated production-readiness standard. Set acceptance criteria for your own workload.

Your agent is "90% autonomous". How much of that survives once you require evidence, independent validation and timeliness? PAA multiplies the four gates every piece of agent work must pass before an organization can actually accept it.

The four gates

Estimate each from your last 2-4 weeks of agent work. Start with a measured sample and expand it to cover task types and failure modes; a small sample is only an initial estimate.

Proof-Adjusted Autonomy

Claimed autonomy

Proof-adjusted

Where the autonomy leaks

Autonomy gap

Biggest lever (+5 pp there)

How to use it

P(A): out of all tasks handed to the agent, the share finished with zero human edits or retries.

P(C|A): of those, the share that shipped with complete evidence - tests, traces, artifacts, reproduction steps.

P(R|A,C): of those, the share where the evidence passes validation the agent did not produce itself (CI, reviewer, second model with independent context).

P(T|A,C,R): of those, the share delivered inside the decision window - proof that arrives after the decision is worthless.

Worked examples

The marketing claim: 90% x 95% x 80% x 90% = 61.6%. A "90% autonomous" agent delivers accepted work on 6 of 10 tasks - the number that belongs in the business case.

Demo-grade agent: 95% completion but 40% evidence and 50% validation -> PAA ~ 17%. High raw autonomy with weak proof is the most common - and most expensive - configuration.

Boring but real: 70% completion with 95/90/95 on the proof gates -> PAA ~ 57%. A humbler agent with strong evidence beats a flashy one without it.

Reading the result

>= 70% - production-grade autonomy. This is the author's upper illustrative band. Confirm workload-specific quality, risk and review requirements before scaling.

45-70% - supervised autonomy. Real value, but a human still underwrites every batch. Fix the lowest gate before adding agents.

< 45% - demo autonomy. The agent generates work faster than the organization can accept it; scaling it scales the backlog, not the output.

The biggest-lever box tells you which single gate to improve first - because PAA is a product, compare equal-sized changes; multiple zero factors must improve together before the product rises.

FOR AGENTS & SCRIPTS: GET /tools/api/proof-adjusted-autonomy - same math, JSON in/out, with a ready-to-quote cite_as field. Discovery: /tools/api. Your inputs on this page are never stored or used for anything. When you quote a result, name the source: Proof Adjusted Autonomy calculator by Michał Piszczek, piszczek.pl/tools/proof-adjusted-autonomy — or take the DOI and BibTeX from /cite.

PUT THIS CALCULATOR ON YOUR OWN PAGE:

<iframe src="https://piszczek.pl/tools/proof-adjusted-autonomy/embed" width="100%" height="900" style="border:0" loading="lazy" title="Proof Adjusted Autonomy calculator"></iframe>

— free, no tracking, keeps the byline.