Autonomous Action Ratio Calculator.
Enter agent activity for any period to see how much work runs with no human step, how much of it is later rated high-risk, and how much operating speed your oversight keeps. Read together, the scores show whether human-at-the-control oversight protects the business and holds its pace.
How GW Slate™ enforces these gates →The seven deployment controls →
Digital speed, human authority.
When markets moved from open-outcry pits to electronic trading, execution moved to machine speed and the liability stayed with the firm. So the industry built control into the infrastructure: safe volume ran at full speed, orders near a limit raised a flag, and firms kept a kill switch within reach. AI agents face the same trade-off.
Every order was checked against the firm's limits before it reached the market. Every agent action is scored against policy before it executes: low risk runs and logs, medium risk lights up for review, high risk waits for a person.
An order beyond a limit waited for someone with authority. A high-risk agent action waits for a named person to approve, override or stop it.
Every order stayed on record with who sent it. Every override carries the name of the person who made it.
An accountable person could stop trading at once. Halt parks the agent system in a safe state and waits for a person.
Regulators wrote that discipline into rule. SEC Rule 15c3-5 (2010) requires broker-dealers with market access to run automated pre-trade risk controls. CFTC Regulation 1.73 (2012) requires clearing futures commission merchants to set risk-based limits and screen orders against them. Article 14(4)(e) of the EU AI Act writes the kill-switch half into AI law: a ‘stop’ button or similar procedure that brings a high-risk AI system to a halt in a safe state.
Run your own oversight numbers.
Load a scenario to see how the scores move, then replace the fields with your own agent records for the same period.
Load an illustrative scenario
Actions scored high-risk at execution waited for a person. Review later re-rated 17 autonomous actions as high-risk: those action types route to a person next period.
Agent actions completed with no human step in the period.
Every agent action in the same period.
Share of autonomous actions later rated high-risk on review.
The same workload with oversight off.
The same workload with human-at-the-center oversight on.
0.744
142 ÷ 168 = 0.845, then 0.845 × (1 − 12%) = 0.744
74.4% of all agent actions ran with no human step and stayed out of the high-risk band on review. Read it with the high-risk count beside it.
17
142 autonomous actions × 12% later rated high-risk
Each one ran with no person in command. Route those action types to a person next period.
0.917
110 ÷ 120 = 0.917, meets the 0.90+ target
91.7% of baseline throughput kept.
Read the three numbers together. A high ratio with a high-risk count that holds or grows period over period means speed bought with unowned risk; a ratio near zero with low speed-retention means a review queue is holding the business back. Results round half-up to three decimals, the target compares the rounded score, and every figure stays illustrative until it comes from your own agent records.
Agents act at machine speed. Oversight keeps pace.
Article 14 of the EU AI Act asks that high-risk AI systems be designed so natural persons can oversee them effectively, including the power to intervene and stop the system.
Human-at-the-center oversight engages where the risk sits, and these scores show it working.
A review on every action stalls the business.
A gate that never engages leaves the audit with no evidence.
Each risk tier gets its own response.
GW Slate™ calculates both scores automatically from its run records, and every override carries a receipt that names the person who made it.
The action runs automated and writes to the log.
A soft light on the operator's console marks a decision that needs a look. The action continues and a person reviews it on their schedule.
The action holds until a named person approves, overrides or stops it.
Teams that want a harder stop set Halt on any tier: the system moves to a safe state at once and waits for a person. Each team sets the response per tier and per workflow, to match its own risk appetite.
What each score tells you.
Oversight becomes an operating number: a rising AAR with a steady high-risk share shows autonomy growing where risk allows, and a falling one shows where a gate should tighten.
Oversight shows it protects the business and keeps its pace: a score at 0.90 or above means risk-tiered control is holding velocity.
Questions on Article 14, risk checks and the scores
01What does Article 14 of the EU AI Act ask of AI agents?
Article 14 asks that high-risk AI systems be designed so natural persons can oversee them effectively while they are in use. Article 14(4) lists what those people must be able to do, as appropriate and proportionate, including override or reverse the system's output and interrupt it through a 'stop' button or similar procedure that brings it to a halt in a safe state. For an agent workflow that falls in a high-risk category, that means people assigned to oversee the system, with the competence, training and authority Article 26(2) asks of deployers, and a halt that parks it safely. Counsel confirms whether a given workflow falls in a high-risk category.
02What is a pre-trade risk check, and how does it apply to AI agents?
On a trading desk, a pre-trade risk check tests every order against limits the firm sets before the order reaches the market. Applied to AI agents, every action is scored against policy before it executes: low-risk actions run and log, medium-risk actions light up for review, and high-risk actions wait for a named person to approve, override or stop them.
03What is a good Autonomous Action Ratio?
Read the ratio beside the count of high-risk actions that ran with no human step. A high ratio with a high-risk count that holds or grows from one period to the next means speed bought with unowned risk, and a ratio near zero with low speed-retention means a review queue is holding the business back. The healthy direction is a rising ratio, a falling high-risk count and speed-retention at 0.90 or above.
04How much speed does human oversight cost?
Speed-retention answers it: governed throughput divided by baseline throughput on the same workload. A score of 0.90 or above means oversight keeps at least 90% of the pace. In the illustrative reference scenario, 110 governed actions per hour against a baseline of 120 gives 0.917.
05How do I measure this on a live workflow?
Start with a Manta Graph™ Assessment, which maps one live workflow: its decisions, data boundaries and economics, with cost and risk visible before you scale. Teams already running agents can start with the Frontier Diagnostic Sprint, which proves one governed route in 14 days.
Move from estimates to a measured workflow.
A Manta Graph™ Assessment maps one live workflow on the systems you already run: its decisions, data boundaries and economics, so cost and risk are visible before you scale. Teams already running agents can prove one governed route in 14 days with the Frontier Diagnostic Sprint, then put one workflow under control in four weeks with the Team Control Sprint.