Action Risk Classification — Patterns

Score the dimensions separately, combine them by policy

Problem. A single number collapses properties that behave differently. An action affecting twelve thousand records reversibly and an action affecting one record permanently can land on the same value, and every consumer downstream then treats them identically despite needing opposite responses.

Approach. Emit a vector: reversibility, reach, monetary magnitude, external visibility, time to recover. Keep the components in the record, and let each consumer apply its own rule over them — an approval policy that triggers on irreversibility regardless of size, an audit policy that triggers on reach regardless of reversibility. A combined headline value may still be computed for reporting, but nothing enforces on it.

Trade-off. Multi-dimensional thresholds are harder to explain, harder to visualise, and produce rule sets that interact in ways nobody fully predicts. Teams also want a single number for dashboards and executive reporting, and the combined value quietly becomes the real one unless the separation is actively defended.

Anchor every level to a worked example

Problem. Levels described in adjectives are assigned inconsistently by different people and drift over time. Nobody can say what separates moderate from significant, so assignments settle towards the middle, and the scale stops discriminating between anything.

Approach. Define each level by concrete, organisation-specific reference points: a named action at that level, an estimate of what undoing it costs in hours or currency, and who finds out. New actions are placed by comparison against those references rather than by judgement against an adjective. Review the references periodically, because what counted as significant at one customer count is routine at ten times that.

Trade-off. The references are themselves estimates, usually produced in a workshop by people guessing, and they lend the scale a credibility the underlying numbers do not support. They are also specific to one organisation at one size, so the scale does not transfer and has to be rebuilt whenever the business changes shape.

Accumulate within the run

Problem. Valuation is naturally per call, and consequence is not. A plan composed entirely of low-valued steps can produce an outcome nobody would have authorised, and each individual evaluation is correct.

Approach. Maintain a running total on the run across the dimensions that genuinely add — records touched, money moved, external messages sent — and value each proposed call against the projected total rather than against its own size. Define what resets the accumulator: a new session, a new objective, an explicit human confirmation. The run's accumulated value becomes a first-class property that consumers can threshold on, separately from any single call.

Trade-off. Some dimensions do not add meaningfully — irreversibility is not cumulative — so the accumulator covers only part of the vector and invites misuse of the rest. A long-running agent also accumulates until it exceeds every threshold and becomes unable to act, which means the reset rule carries more weight than it looks like it should.

Treat the scale as a model and back-test it

Problem. Levels are assigned once, at registration, by people reasoning about what might happen. Afterwards, actual incidents, reversals, and complaints arrive, and nothing connects them back to the valuations that were given. The scale is never wrong because it is never checked.

Approach. Record the value assigned to every action alongside what subsequently happened: reversals, support contacts, incidents, manual corrections. Periodically compare, and move levels where the evidence disagrees with the assignment. Re-value a sample of historical actions under the revised scale to see what would have changed, and version the scale so consumers can migrate deliberately rather than being surprised.

Trade-off. The outcome signal is sparse and heavily biased: the actions that caused incidents are disproportionately the ones nobody gated, and the ones that were gated have no outcome to learn from. Recalibration also moves every threshold's meaning at once, which makes the change expensive to coordinate across the consumers that were deliberately kept independent.