Capability does not confer authority
AI capability must be separated from the authority an organisation deliberately delegates, including clear escalation boundaries.
An AI agent may be capable of reading a customer record, changing it, contacting the customer, offering a discount and issuing a refund. That does not mean we should allow it to do all five. This distinction will become increasingly important as organisations move from using AI to delegating work to AI. We have spent a lot of time asking what artificial intelligence can do. Managers increasingly need to ask a different question: What is it authorised to do on our behalf? Capability and authority are not the same thing.
A salesperson may be capable of offering a customer a 50 per cent discount. Their employer probably does not give them unlimited authority to do so. The same principle applies to artificial agents. We need to decide what information an agent may access, which actions it may take, what resources it may consume and what commitments it may make. We also need to determine when it must stop. That last point matters particularly when work requires judgement. An agent may handle ordinary cases extremely well while encountering a situation in which the organisation would expect an experienced person to recognise something unusual.
Escalation is therefore not simply an error mechanism. It marks a boundary. On one side are circumstances in which we are prepared to delegate action to artificial agency. On the other are circumstances in which we want situated human judgement. The fact that an AI system becomes more capable does not remove our responsibility for deciding where that boundary lies. Quite the opposite. As capability expands, authority needs to become more deliberate. The management question therefore changes from: What can this agent do? to: What are we prepared to allow this agent to do on our behalf, under what conditions, and when must it stop? That is not primarily a technical question. It is a management decision.