Your customers already told you what the AI should handle first

A Stacker piece that ran in local papers across the country this week opens with a big number. Nearly nine in ten companies expect AI agents to be talking to their customers by the end of 2026, according to Sinch’s survey of 2,527 senior decision makers.

The more useful line is buried further down. When Sinch asked consumers which tasks they would trust an AI agent with, overall confidence was high, but it fell the closer the task got to money and account security. Where’s my order? Fine. Change the card on file, issue the refund, reset access to the account? Much less fine.

Most coverage reads that as a warning. We think it’s a plan. Your customers have ranked your support topics for you, from “go ahead” to “I want a person.” That ranking is the order you should hand them over in.

One switch for the whole desk ignores the ranking

The common way to launch an AI support agent is all at once. Connect it to the help desk, give it the knowledge base, set a confidence threshold, turn it on. Every topic goes live on the same day with the same level of freedom.

That treats a shipping question and a refund request as the same kind of work. Customers don’t. And the gap shows up in the data. Verint’s 2026 consumer research found 61 percent of customers say they prefer a human agent, up five points in a year. But 69 percent of those same people said they’d happily use automation if it fully resolved their issue. Among 18 to 34 year olds, 93 percent.

So people are fine with AI. What they don’t want is an AI that fumbles, and they especially don’t want one fumbling with their money. When one launch covers every topic, the refund conversations set the reputation for the shipping ones.

What the money end looks like when it goes wrong

A recent whiteboard video on rogue AI agents, built on an IBM breakdown, walks through a scenario every support lead should picture. A customer asks when their order will arrive. The agent answers by approving a full refund plus a $50 credit. The replay shows why: it pulled the VIP policy for a regular customer and followed it.

The video’s answer is observability. Record every step so you can trace the refund back to the bad retrieval. That’s worth having, and you should demand it from any vendor. But notice the order of events. The trace explains the refund after the money is gone.

For the topics at the bottom of your customers’ trust ranking, you want the refund to be impossible before anyone needs an explanation.

Read the ranking as a rollout order

Here’s how we’d turn the trust gradient into a practical sequence for a small support team.

Start where trust is already high. Order status, hours, shipping times, return windows, how-to questions. These are the questions customers are happy to have a machine answer, and they’re usually a big share of the inbox. Let the agent practice on them against your real history first, then go live on those topics only.

Keep account and money topics in practice longer. Refunds, billing changes, cancellations, and anything touching account access stay in draft mode. The agent writes the reply, a person reads it and sends it. You watch how often you’d have sent it unchanged.

Put the hard limits in code, outside the conversation. When a money topic does go live, the model shouldn’t be the thing deciding what’s allowed. The AI proposes the refund. Plain software checks whether this customer is verified, whether the order is theirs, and whether this kind of action needs a person’s approval. If the checks fail, it waits for a person. A persuasive customer or a bad retrieval can’t talk its way past an if statement.

Graduate topics on evidence. Move a topic from practice to live when its drafts have held up on real conversations. Move the next one when it has earned it too. If one topic starts going sideways, pull that topic back and leave the rest running.

That last point matters more than it sounds. The same Sinch data found most companies running AI agents have had to roll one back at least once. If every topic is tied to one switch, one bad refund conversation takes down the order status answers your customers liked.

Where we come in

This is how we built Celeste. It learns from your own support history, practices privately on real tickets before it answers anyone, and goes live one topic at a time as you decide it has earned that. Write actions like refunds can be set to wait for your approval, and customer data is scoped to the verified customer asking. Those checks run in code the agent can’t argue with. Here’s how the graduation works, step by step.

Your customers already told you where they trust an AI and where they don’t. The simplest thing you can do is listen, and start at the top of their list.