The Decision Boundary.
Every individual action was permitted.
A consumer business introduced an AI-enabled remediation agent to handle straightforward customer refunds. The design looked conservative. The agent could approve and execute refunds below £5,000 per customer. Anything above that went to a human.
The control was clear enough to survive a committee meeting: low-value cases automated, high-value cases escalated. A human remained accountable for the service. Exceptions were reviewed. Logs were retained.
Then an upstream data issue incorrectly classified a large group of customers as eligible for remediation.
Over three hours, the agent processed 64 refunds totalling £241,000. No single refund exceeded £5,000. No permission was breached. No exception rule fired.
The agent stayed inside the limit. The organisation moved outside the intended authority.
| Time | What happened | Decision Infrastructure consequence |
|---|---|---|
| 09:00 | Agent begins normal remediation processing. | Authority appears bounded by the £5,000 per-customer limit. |
| 09:17 | An upstream data feed changes eligibility status for a customer cohort. | The evidence feeding the decision changes. No reopening trigger exists. |
| 09:26 | Refund volume rises sharply. | There is no cumulative value or velocity boundary. |
| 10:05 | The agent has processed materially more cases than the normal hourly range. | No challenge mechanism pauses execution because each transaction remains “within limit”. |
| 11:38 | 64 refunds have been issued. Aggregate value reaches £241,000. | Technical permissions remain valid while practical authority exceeds intent. |
| 12:10 | A human spots the pattern during routine monitoring and suspends the workflow. | The stop right exists, but only after the consequence has accumulated. |
The boundary was defined around a transaction, not around exposure.
Below £5,000, the agent may act.
The limit controlled the size of one action.
Routine, low-risk cases may be automated.
The intent depended on normal volume, familiar patterns and trusted data.
Unlimited compliant actions inside the threshold.
Aggregate value, velocity and novelty were outside the boundary design.
This is the difference between a permission and a decision boundary. A permission can be individually correct and systemically incomplete.
| Component | What the case reveals |
|---|---|
| Authority | The agent had practical authority to approve and execute any number of individually eligible refunds below the threshold. |
| Evidence | Eligibility evidence changed upstream. The decision process trusted the new status without a materiality check on the change. |
| Boundaries | The boundary controlled amount per customer, but not cumulative exposure, rate of action, unusual concentration or data novelty. |
| Traceability | Logs could show every refund. They did not by themselves explain why the aggregate pattern remained permitted. |
| Challenge | A human could stop the process, but monitoring cadence meant the right was exercised after exposure accumulated. |
| Reopening | No trigger required reconsideration when refund volume, data distribution or cohort size changed materially. |
Three designs were available. Only one treated the boundary as a system.
Lower the transaction limit.
Easy to implement. Weak answer. The agent could still execute a large number of smaller actions and accumulate the same exposure.
Add a daily aggregate cap.
Better. It constrains total value, but can still miss unusual velocity, new patterns or concentrated cohorts.
Bound amount, volume, velocity and novelty.
Use multiple conditions. Pause or escalate when any material dimension moves outside the approved operating envelope.
The organisation chose Option C. The key change was conceptual: the decision boundary stopped being a number attached to one transaction and became a set of conditions attached to the decision system.
Before delegating a consequential action, test more than the amount.
| Boundary dimension | Question |
|---|---|
| Amount | How large may one action be before additional authority is required? |
| Volume | How many actions may occur before aggregate exposure becomes material? |
| Velocity | How quickly may actions accumulate before the process pauses? |
| Novelty | Which new pattern, cohort, tool, geography or data condition requires escalation? |
| Confidence | At what level of uncertainty does autonomy end and human judgement begin? |
| Consequence | Which actions are too difficult to reverse to allow autonomous execution? |
Not every process needs all six. Every consequential autonomous process needs an explicit answer to which ones matter.
- Are our AI limits defined per action, or do they also constrain aggregate exposure?
- What change in volume or velocity automatically pauses autonomous execution?
- Which upstream evidence changes can materially alter the agent's behaviour without reopening approval?
- Who can stop the workflow in real time, and how quickly would they know to do so?
- Which consequential actions are currently permitted because each is small, even though the aggregate can become large?
Good Decision Infrastructure does not merely ask whether each autonomous action was permitted. It asks whether the organisation intended to delegate the pattern of decisions the system can produce.
Next: Regulatory Radar, 27 October 2026. Regulation Is Redrawing the Decision Boundary.