The interesting question about AI agents is no longer whether they can complete a task without a human. It is how an organisation decides they are allowed to. McKinsey’s Rich Isenberg framed it well this month: design for trust first and speed second, and make teams convince you they have earned autonomy rather than granting it because the technology is capable of it.
Safe, fair and fixable
Isenberg’s test for anything customer-facing is blunt: when something goes wrong, people do not care that it was AI. They care that it was safe, fair, and fixable. Those three words are a better acceptance standard than most AI policies we read, because each one is checkable before deployment.
Safe — what is the worst action this system can take unsupervised, and can it be bounded? Fair — would the outcome survive being explained to the person it affected? Fixable — when it goes wrong, how quickly is it detected, who can reverse it, and is there a record of what happened?
Four decisions to make before go-live
- Decision rights. Which specific actions may the agent take alone, which require approval, and which are prohibited outright? Write the list. Ambiguity here is how scope quietly expands.
- Accountability. A named person owns each class of action the system takes. “The system decided” is not an answer a regulator, an auditor or a complainant will accept.
- Escalation. What triggers a handover to a human, how fast, and to whom? Define it by condition — low confidence, unusual value, a vulnerable service user — rather than leaving it to the agent’s judgement.
- Controls and monitoring. What is logged, who reviews it, and how often? Autonomy without monitoring is not delegation; it is abdication.
A ladder, not a switch
In practice we find it helps to treat autonomy as four rungs rather than an on/off decision, and to move a system up only when the monitoring from the rung below justifies it:
- Suggest. The agent proposes; a person does the work. Useful for learning where it is reliable.
- Draft for approval. The agent produces the output; a person checks and releases it. Most organisations should live here longer than they expect.
- Act with notice. The agent acts within defined bounds and tells someone, with a window to reverse it.
- Act autonomously within bounds. Reserved for actions that are low-impact, easily reversed and well-evidenced by monitoring.
The virtue of the ladder is that it turns an argument about principle into a question of evidence. A team asking for more autonomy has to show what the logs from the current rung demonstrate.
Why this matters more in public services
A retailer whose agent misprices an order refunds a customer. A council whose agent mishandles a homelessness application, a benefit claim or a safeguarding referral is dealing with a statutory duty and a person with nowhere else to go. The asymmetry means public bodies and charities should climb the ladder more slowly than the commercial case studies suggest — and should be able to show why each step was justified. That record is also, incidentally, what makes an AI deployment defensible at audit.
Go deeper
- Trust in the age of agents (McKinsey Podcast) — Rich Isenberg on decision rights, escalation paths and earned autonomy
We help organisations set proportionate governance around AI systems — decision rights, escalation and the evidence trail that makes a deployment defensible. If you are about to give a system more independence, talk to our advisory team.