The promise your chatbot makes
A chatbot at Air Canada told a customer about a bereavement discount that the airline did not offer, and the tribunal held the airline to it. The company argued that the bot spoke for itself. That argument lost. The promise stood, and the bill followed.
Read that as a pricing decision made by software, in public, without a manager in the loop. Every organisation that puts an agent in front of a customer now has that same exposure on the books. The question for the boardroom is whether your way of working still matches what your technology is allowed to say.
Yesterday: the script and the supervisor
The old model contained risk by containing speed. An agent in a contact centre worked from a script, escalated anything unusual, and had a team lead two metres away. Mistakes happened one conversation at a time. A wrong promise cost you one refund and an apology.
Control lived in the hierarchy. Policy sat in a document, the document sat in training, and training sat in people who could hesitate. Hesitation was the safety mechanism, and it was free.
Today: the agent that decides
An AI agent handles thousands of conversations an hour and treats every one of them as settled. It reads the customer, picks an action, and commits the company. There is no pause, no glance at the desk next door, no colleague who says wait, let me check that.
What changes is the unit of error. A misread policy is no longer one bad call. It is a pattern that runs until someone notices, and noticing takes days.
Software repeats its mistakes at the speed it repeats its successes.
Where it tips
The moment of change is the moment a system speaks with your authority. Recommendation engines suggest, and a human decides. Agents decide. That shift crosses a line that most governance frameworks were written before.
Three things move at once when you cross it:
- The promise becomes binding the second it is made, without review.
- The volume of promises leaves the range a manager can sample by hand.
- The reasoning behind each promise sits in a model, so reconstruction takes work you have not yet organised.
Boards tend to discover this order backwards. The promise lands first, the volume becomes visible second, and the reconstruction starts after the claim arrives.
Guardrails are how you delegate
Set the boundaries first, and you can hand over far more than you would dare to hand over now. A defined discount range, a hard stop on legal commitments, a logged escalation path: each one buys you room to let the agent act alone in everything else.
Treat that as a commercial instrument. The organisations that write their decision rights down are the ones that can deploy agents in pricing, retention and claims while their competitors keep a human in the loop for everything and wonder why the cost curve stays flat.
Trust is on the balance sheet
A customer gives an agent authority over a purchase only when the company behind it is answerable. That is the asset now. Accountability you can demonstrate is what makes delegation acceptable to the person on the other side.
Air Canada paid for a sentence it never approved. Your agent is writing sentences today.
What you do on Monday
- Name the owner of every customer-facing agent, by person, not by department.
- Write down what each agent may commit the company to in money, in time and in rights.
- Log every agent decision with its inputs, and set a weekly sample review with a real reader.
- Put one hard stop in place this week: no agent makes a legally binding promise without a named human approver.
- Publish to your customers what your agents may do. Answerability sells.
Decision rights are the practical part of this, and most teams find the conversation easier with a template in front of them. The AI Decision Rights material maps who may commit the company to what, per agent, per journey step. Bekijk de materialen