Written by Joe Milton
This week I used AI to help me rethink this article. It challenged my framing and caught a date error. I kept some suggestions and set others aside. Either way, my name is on what I publish, so I’m responsible for every word.
You can delegate work to AI. You cannot delegate accountability for it.
Some AI tools generate an answer and wait for us to act. An AI agent can go further: it can use tools and take steps toward a goal on our behalf. It is not something separate from AI. It is an AI system given instructions, access, and some authority to act.
That authority can be useful. An agent might reopen a service request, issue a refund, or route a problem to the right team. But the more authority we give it, the more clearly we must decide what it is trying to accomplish, where its authority ends, and who answers for the result.
In 2024, a Canadian tribunal held an airline liable after its chatbot wrongly told a traveler they could apply for a bereavement fare after their trip. The traveler relied on that advice when booking. The airline argued that it could not be held liable for information provided by its chatbot. The tribunal disagreed. The chatbot was part of the airline’s website, and the airline remained responsible for what it communicated.
That chatbot gave information. An agent can also act. Picture one with permission to close customer-service cases. If the rule it follows is flawed, it can repeat the same mistake across an entire queue before anyone reads the weekly report.
Success Can Look Like a Closed Case
Imagine opening that report on Friday. The agent closed more cases, and customers received faster replies. Then a customer writes back on a closed case: “I still need help.”
If they have to start over and explain everything again, whose time did we actually save?
From the team’s perspective, the case is closed. From the customer’s perspective, the problem is still theirs to solve. Their time matters, even when the report in front of me records only ours.
That is why the most important decisions come before the agent closes its first case. The team needs to agree on what the agent is for, what it may do, and who stands behind its work.
The Four A’s of AI Accountability
Before giving an AI agent permission to act, work through four questions with your team.
Aim. What are we asking it to accomplish?
Closing a case and solving a problem are different goals. If resolution is the goal, look at how often customers return with the same issue, not only how many cases the agent closes.Authority. What may it decide on its own?
Let it reopen an unresolved case. Require a person to decide whether to deny a disputed refund. Before granting permission, ask: Can we undo the decision? Who bears the cost if we are wrong? Can the customer reach a human representative who can help? In a 2026 Gartner survey of 3,566 customers, 87% said access to a human was essential when companies use generative AI for customer service.Assurance. How will we know whether it is working?
Each week, sample cases the agent marked closed and ask some of those customers whether their problem was resolved. If their answers differ from the dashboard, find out why.Answerability. Who owns the result?
Name the leader who will act on what the team learns. Give that person the authority to change what the agent may do and to make sure the customer receives help.
These questions can show me where to look. They cannot make me want to hear what the numbers miss. That requires something deeper than a governance framework.
What Makes Accountability Hard to Own?
When my sense of worth depends on how I perform, every metric becomes a mirror. I am tempted to favor the number that makes both my team and my leadership look good. If I chose that number, it can be especially hard to listen to the customer who reveals what it missed.
In Matthew 20, the mother of James and John asks Jesus to give her sons places of honor in his kingdom. The other disciples are indignant. I recognize something of myself in their concern for recognition and position. A favorable number can begin to matter because of what I believe it says about me.
Jesus points to a different use of authority. He came not to be served, but to serve and to give his life as a ransom for many (Matthew 20:20–28).
Before I ask what serving someone will cost me, I remember what Christ has given for me. He bore my sins. Because my standing before God rests on Christ, I do not need a favorable report to justify me. I am free to face what the numbers miss, admit when a decision was wrong, and help put it right.
That freedom matters before the mistake, too. It lets me choose a measure that serves the customer, even when another number would make my team look better.
At your next review, put the customer’s “I still need help” beside the number of cases you closed. Before giving an AI agent more authority, ask:
If we hit our target and this person still needs help, who will answer them?
References
Moffatt v. Air Canada, 2024 BCCRT 149, British Columbia Civil Resolution Tribunal, Feb. 14, 2024.
Gartner, “Gartner Survey Finds 87% of Customers Say Companies Using GenAI for Customer Service Must Provide Access to a Human Agent,” Aug. 4, 2026.
Matthew 20:20–28.
