CX Leaders Turn Implicit Judgment Into Explicit Decision Rules Before Deploying AI Agents

AI agents in customer service now make decisions on a company’s behalf, from routing a complaint to approving a refund. Each of those decisions rests on a set of rules. In most CX teams, many of those rules have never been written down. They live with experienced employees who know which colleague handles which issue and what to do when a case falls outside the usual pattern. An agent can’t draw on that knowledge unless someone spells it out first. As more companies move agents into live customer conversations, documenting that unwritten judgment is becoming a core part of deployment.

Stéphane Bédère is Global Head of Customer Success and Operations at Ingenico, where he oversees customer success for the company’s Payments Platform as a Service business. He also works as an independent Transformation and Operations Advisor, helping SaaS, payments and fintech companies with operating models, revenue operations and AI-enabled execution. He has led customer success and operations teams in complex B2B and regulated environments across 22 countries, and previously served as Chief Customer Officer at Sidetrade, an AI platform for order-to-cash, and held earlier leadership roles at Anaplan and Jive Software. He developed EOS, or Execution Operating System, a framework for making organizational decisions explicit and connecting them to the processes, systems and governance that carry them out. That work shapes his view of what agents need before they go live, starting with the intent behind every decision they’re asked to make.

"The most important thing with AI today is making things obvious at each step of the delivery process. A lot of things are still implicit in how decisions get made," says Bédère. Intent covers the purpose of a decision, its context, its goal and its limits. Inside a company, people fill gaps in that picture by asking a colleague who has handled the issue before. Agents can’t do that, which changes how teams need to prepare.

Listen for it depends

Finding the unwritten rules starts with a short list. Reviewing every workflow before launch would be an arduous process that delays deployment. A better approach is to pick the processes that matter most to customers and ask the people who run them how each decision actually gets made. In Bédère’s experience, certain answers are a giveaway. When someone says it depends, or that a particular colleague always handles that kind of case, the decision rests on judgment that was never documented. An agent working in that area won’t be able to give customers a useful answer, so the process needs a built-in stopping point there.

Once a team spots one of those decisions, the next step is finding the person who really makes it. That isn’t always who the official process names. Talking with that person shows what they weigh and whether it can be turned into clear rules. If it can, the agent can take on the work a piece at a time. If it can’t, the decision stays with people.

For the decisions that stay with people, the handoff has to be designed before launch. That includes passing along the full conversation history, so the customer doesn’t have to start over with a person. "With an agent, you have to anticipate. You have to provide the context and the right process, and ask the agent to stop and raise its hand," notes Bédère.

AI agents in contact centers now handle more than simple FAQs, including account changes, billing questions and case resolution. That makes ongoing oversight more important. A rule that made sense at launch can go out of date as products, policies and customer needs change, and the agent will keep following the old rule until someone updates it.

Set expectations up front

The same clarity applies to customers. People should know when they’re talking with an AI agent and what it can handle on its own. That expectation now carries legal weight in Europe, where the EU AI Act has required chatbots to disclose that they’re AI since August 2026. Customers are drawing similar lines. In a 2026 Gartner survey of more than 3,500 B2B and B2C customers, half said their service interactions are easier when companies use generative AI. Nearly nine in 10 said companies using it must also offer a way to reach a person.

Being open about timing matters too. Some questions can’t be answered on the spot because they need input from another team, more information or a person’s review. Customers who know that from the start understand why an answer takes longer. "You have to be transparent with your customers. Even a well-designed agent is not able to make every decision, so you have to find a way to include time in the discussion for arbitration, to get information and to share it internally. Your customers won’t always get an answer in a minute," Bédère says.

He calls this setting the rules of the game, and it applies inside the company as much as with customers. The people who pick up a case from an agent need to know what it already decided and why it stopped. When everyone works from the same understanding, a transfer feels like a normal step in the service. It also gives teams a chance to rethink how issues get resolved, with agents and people each handling the parts of a case they’re best suited for.

Treat the agent like a hire

Bédère suggests treating an agent like a new hire. No company hires someone without a clear job for them to do, and an agent needs one too. Launching one just to keep up with competitors leaves the team without a goal to measure against. Without that goal, it’s hard to tell whether the agent is working.

"Be clear with your use case and be humble with the scope," adds Bédère. A new hire also has to work well with the people around them. In a customer-facing organization, that means colleagues in account management, customer success, support and product, plus partners and customers outside the company. An agent needs that range too. If it can only work inside one team’s process, it isn’t ready.

"It’s very important to imagine an agent like an employee. An employee is able to connect with everybody, inside and outside the organization. If your agent is not able to do the same, you have an issue," says Bédère.