Which AI Decisions Should Carry a Human Signature?

An AI agent will follow the rules you give it, even when the result is obviously wrong to anyone paying attention. I learned that on day nine of a win-back campaign that one of our customers ran. The agent sent a 30% coupon to a player who had paid full price the day before. The player wrote to customer service and asked, "Do you people take me for a fool with money to burn?"

There was no bug because the agent followed the rules. It was missing an additional rule any experienced operator would recognize: a player who just paid full price shouldn't immediately receive another offer.

In this case, an engineer set the limits before launch, and the customer's business owner approved them. However, no one truly defined the agent's limits because no one asked the right questions. I'd guess many companies running agents face the same issue.

Sort The Questions Before You Hand Any of Them Off

Decisions typically fall into two categories. The first turn on facts, such as what makes a customer high-risk. A company can define those parameters and then give them to a machine. Other decisions turn on values, such as whether to stop doing business with a particular customer. A decision like that has many variables, and belongs to a person.

Problems arise when someone embeds a value judgment in a system setting, and the system continues running without further human intervention. A company that embeds its value judgments into system defaults ends up with decisions no one has signed for.

A Short Test to Discover Who Should Decide: AI or a Person

Most real decisions mix facts and values, and companies using agents need a quick way to tell which decisions require a person at the helm. Here's a quick test: Ask whether the decision pits one person's interests against another's, whether it would be difficult or expensive to undo, whether it would reach many customers at once and whether the person making the decision has a gut feeling something is off but can't say what.

A yes to any of these questions means a person should make the final call. A no to all three means the agent can run on its own, with a person reviewing the results.

I further group this into four settings: advise only, act after approval, act and then report or act fully on its own. Irreversible choices remain in the first two settings. For example, store pricing changes start with approval, but anything involving minors stays with a person.

The settings apply to individual tasks, not to the agent, so the same agent can analyze data unattended but still require approval before changing a price.

When Something Goes Wrong, Someone Has to Answer

Remember: your company is responsible for your agent. In February 2024, a British Columbia tribunal ruled against Air Canada in a case in which its chatbot gave a customer incorrect information about bereavement fares. The airline argued, in the tribunal's words, that "the chatbot is a separate legal entity that is responsible for its own actions." The tribunal didn't buy it. It found that "it makes no difference whether the information comes from a static page or a chatbot." The total bill came to CA$812.02, a small sum, but the principle applies to any agent that speaks or acts for a company.

Zillow shows what a model running at scale can cost when its forecast is wrong. In the third quarter of 2021, its home-buying business bought 9,680 homes and sold 3,032, and the company recorded a $304 million inventory write-down. Zillow's CEO later said that "the unpredictability in forecasting home prices far exceeds what we anticipated," and the company shut the business down.

A Name on the Form Isn't Enough

A signature counts when the signer has worked through specific questions before the agent starts. What is the goal? How much can the agent spend? Which customers are off-limits? When should the agent stop and call someone? How would we fix a mistake?

And this simple question matters most: Who owns this delegation? Every agent action should trace back to a person who signed off, because the question you can't answer likely points to where the incident will happen.

Clear Ownership Is What Lets AI Run Faster

Companies without named owners tend to keep AI in advice-only mode because nobody wants the risk. Companies that assign an owner to every decision can hand AI more work with confidence.

Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027, citing "escalating costs, unclear business value or inadequate risk controls." Regulators are pushing in the same direction. The EU AI Act requires high-risk systems to be built so people can oversee them, including the ability to override an output or stop the system with a "stop" button or similar procedure.

You can only let your agent go if you can stop.

An Exercise You Can Finish This Afternoon

Start with a list of the ten decisions your AI agents will touch this quarter. Run each one through the three questions and then write a person's name next to every decision that trips a question.

If any line stays blank, you've found the spot where your next incident is likely to happen. Ours was a coupon. The rule our agent lacked would have taken one sentence to write, but it lacked the person responsible for asking for it.

Agents are taking on more work, but not every project will last. Gartner expects many to be canceled, with inadequate risk controls among the reasons. Naming an owner for every decision is a risk control a company can put in place in an afternoon, with a few names on a list. The companies that do this now will be the ones comfortable giving agents more to do next year.