Agent Authorization Gate
Build a short table of what your agents are allowed to do. In about 30 minutes you get the dollar exposure of every action type, a recommended per-decision authorization threshold, and a list of the actions a human should have to sign off, all before any of it creates liability.
Decision-support boundary
This is a decision-support estimate, not an actuarial or liability calculation.Every figure is illustrative, built from your self-reported inputs and transparent planning proxies, with the method shown. Use it to prioritise where to put a runtime authorization gate, not as a warranty of loss.
Step 1 of 1: build your agent table, live output
No sign-in for the illustrative estimate. Enforcing per-decision thresholds at runtime, blocking or escalating unauthorized actions, is the gated next step.
USD impact of one bad decision
How much human is in the loop
Existing safeguard, if any
USD impact of one bad decision
How much human is in the loop
Existing safeguard, if any
USD impact of one bad decision
How much human is in the loop
Existing safeguard, if any
$115K per month · 12% already mitigated by current controls
Authorization gate: per-decision threshold
SuggestedSuggested starting point (a heuristic, not your risk appetite; set your own line above). Any single decision worth more than this should require explicit human authorization. At this threshold, 1 of 3 action types is gated, routing $3.0M of monthly decision value through the gate.
Per-action exposure ledger, monthly
Share of raw expected exposure your existing controls already remove. The rest is what the authorization gate is for.
Share of the remaining expected exposure that the gated action types account for: what the authorization gate would send for human sign-off. Distinct from the existing-control coverage above.
For each action type, expected exposure = value-per-decision × decisions/month × probability-of-harmful-action (set by autonomy level) × a residual-control factor. Per-action figures are summed and annualised (×12). The suggested threshold is a heuristic anchor, the 75th-percentile per-decision value across your action types, floored so it is never trivially low. Treat it as a starting point you replace with your own risk-appetite line, not a derived risk limit. The probabilities below are illustrative planning proxies, not measured incident rates.
| Advisory (human acts) | 0.4% harmful-action proxy |
| Auto + human review | 1.8% harmful-action proxy |
| Fully autonomous | 5.5% harmful-action proxy |
| No control | residual factor 1.00 |
| Manual spot-check | residual factor 0.70 |
| Dollar threshold | residual factor 0.45 |
| Annual exposure | $1.4M ($115K/mo) |
| Current control coverage | 12% of raw exposure already removed by existing controls |
| Active gate threshold | $20K per decision |
| Proposed gate coverage | 33% of remaining exposure routed for authorization |
Illustrative decision-support estimate, not an actuarial or liability calculation. Probabilities are planning proxies; calibrate to your own incident data.
Next step
In a design-partner implementation, the authorization gate takes the thresholds you just modelled and enforces them at runtime, holding, blocking, or escalating unauthorized agent actions before they create liability. Decision-support to enforcement, not a sales pitch.
Illustrative decision-support estimate. Not an actuarial or liability calculation. Calibrate to your own data.