Adrian DunkleyAI Risk Management

Adrian Dunkley’s proposed method

FAIR-AI puts usage and oversight into the loss estimate

Adapt Factor Analysis of Information Risk to measure how AI failures turn into losses at operating scale.

A model’s error rate can tell you where to look. A loss estimate also needs to reflect how often the system is used, how people respond, whether a reviewer catches the failure and what the consequence costs. FAIR-AI is my proposed adaptation of Factor Analysis of Information Risk (FAIR), which decomposes risk into “Loss Event Frequency” and “Loss Magnitude”. FAIR Institute, Factor Analysis of Information Risk Standard v3.0

Four changes for AI risk

MechanismWhat it representsEvidence to request
Two trigger channelsOrdinary system failures and deliberate interference follow separate paths.Evaluation results, incident history and authorised security findings.
Exposure multiplierA per-use rate scales with the number of relevant interactions.Production counts for the defined workflow and time horizon.
Autonomy coefficientThe chance of a consequential failure escaping effective human interception.Review logs, intercepted cases and evidence of missed cases.
Common-cause shockA shared model, release or dependency can raise loss across multiple uses together.Model versions, release exposure and concurrent incident records.

Treat non-malicious errors and deliberate interference separately. Define the population for each path, and check whether they could count the same harmful outcome. When using conditional rates, name the stage each rate follows. State whose losses are counted and record harms that money alone cannot represent. FAIR Institute, Factor Analysis of Information Risk Standard v3.0 NIST, Generative Artificial Intelligence Profile

Work one chatbot path

Use a constructed example in Jamaican dollars over one year. Take 120,000 eligible interactions and a 0.2% material error rate. Assume a customer relies on 25% of material errors and 20% of relied-upon errors escape effective interception. In this example each escaped, relied-upon error produces one direct-loss event. The expected frequency is 120,000 × 0.002 × 0.25 × 0.20, or 12 loss events per year. At a mean J$25,000 per event, the expected direct loss is J$300,000 per year.

If tested review changes the escape rate to 10% with other inputs fixed, the estimate becomes six events and J$150,000. A control costing J$90,000 per year leaves J$60,000 of expected net improvement under those assumptions. Keep uncertainty, delays and harms visible when discussing that decision.

Check what happens in a shared failure

Model a common cause, such as a release that raises errors across several channels. Compare the annual loss quantiles with and without the shared failure, using the same horizon and currency. I call the ratio at a matched percentile the Correlated Loss Amplification Ratio. A ratio of 1.5 at the 90th percentile would mean the correlated scenario’s 90th-percentile loss is 50% higher, provided the comparison loss is above zero. Record the dependence assumptions and check how they change the result.

Copy this prompt

Using the approved evidence and assumptions below, scope one AI loss scenario. Separate ordinary failures from deliberate interference. Define the perspective, currency, horizon, annual exposure, conditional event rates, interception and loss per event. Show the base calculation and units. Identify data gaps and a shared failure to test. Keep observed inputs separate from assumptions. Compare controls by expected direct loss, cost and residual harms.

Use it for the decision at hand

This article explains a proposed method through a practice example. Use verified operating data and analytical review before making a budget or control decision. Start with the National Agentic Risk Matrix lesson if your team still has to choose which failure path to examine.

Sources and references