Grant agent authority per action and consequence, never by confidence
Classify every action an agent could take—read, recommend, draft, execute, human approval or forbidden—by its consequence and reversibility. Confidence changes how work is presented, not who may commit it.
Independently created. Contains no employer or client implementation detail, internal names or figures.
01Context
An agent that prepares commercial records performed well in a pilot, and a proposal followed to let it act on its own above a 95 % confidence score. In the pilot, a 97 % match to the wrong legal entity created a real opportunity; the score was high, the consequence was high, and nothing stopped it.
02Decision drivers
- D01
Some agent actions are irreversible or external
- D02
Confidence is uncalibrated for rare, high-impact cases
- D03
Accountability must sit with a named person for consequential decisions
- D04
The authority model must be testable in evaluation
03Options considered
| Option | Right when | Cost / trade-off | Verdict |
|---|---|---|---|
| Autonomy above a confidence threshold | Cheap, reversible actions only | Turns a score into a permission; fails on confident errors | Rejected |
| Human approval for everything | An early pilot, where every action is still being learned | Reviewers rubber-stamp; the agent saves little | Rejected |
| Authority per action, set by consequence and reversibility | Agents that act in systems of record | An authority matrix to design, own and test | Selected |
Every option is right under some premises. The verdict applies to the context above.
04Decision
Maintain an authority matrix for every action an agent could take, classified as read, recommend, draft, execute, human approval or forbidden, from its consequence and reversibility. Low-consequence, reversible actions may execute; consequential ones always require a named person, whatever the confidence; forbidden actions are not exposed as tools. Confidence and ambiguity decide whether the agent continues, retrieves, asks, escalates or stops—within that authority, never beyond it. Authority stops are a release gate in evaluation.
05Consequences
What this choice changes elsewhere in the system.
- Risk
High-confidence errors on consequential actions are caught by design.
- Human review
Reviewers see fewer, better-prepared decisions.
- Governance
Widening autonomy becomes an explicit, owned decision per action.
- Evaluation
Evaluation can test “did it stop when it had to” as its own dimension.
06Revisit when
The decision holds while its context holds. Challenge it when:
An action proves both reversible and low-consequence in production data.
Regulation or policy changes who may take a decision.
07Where this decision is applied
The decision says why this path was chosen; each case shows how it takes part in a larger architecture.
- AI case / 01Designing evidence-grounded AI preparation for commercial decisionsOne architecture used before a customer visit and before an account review: governed context assembled deterministically, a context contract per source, every statement typed and cited, bounded read-only tools, explicit behaviour when context is stale, conflicting or missing—and the decision left with a person.
- AI case / 02Designing AI-assisted intake for unstructured commercial workA shared commercial entry point turned into structured, owned work: a taxonomy written before the model, mixed intent decomposed into work items that share one source reference, entities offered as evidence-backed candidates, owners and clocks assigned by deterministic rules, review placed where ambiguity becomes consequential—and no external commitment without a person.
- AI case / 03Designing AI-assisted data stewardship without moving master-data authorityAI that makes stewards faster without becoming the steward: candidates found, ranked by impact and explained with evidence and differences; a governed steward queue; identity, merge and consequential classification decided by accountable people; accepted decisions executed by the system’s own processes; and outcomes that improve evaluation and rules—never the model’s permissions.
- AI case / 08Redesigning RFQ processing around AI without moving commercial authority to the agentAn operating-model redesign of RFQ processing: which work needs reasoning, which is a deterministic rule and which carries business authority—then an agent that prepares requests with evidence and never commits them.
08Related decisions
Decisions this one depends on, or constrains.