Most companies build their AI agents the way they build their call centers, setting up one team for support, one for sales, one for order tracking, one for returns, one for product questions.
It’s a sensible way to organize humans, with each person in a lane and measured on it. But when that same logic gets copied into AI, something breaks that customers can feel even if they can’t name it.
Too many companies building AI agents are shipping the org chart directly to the customer, seams and all.
Why AI can hold both temperaments
The instinct to specialize is borrowed from a context where it made sense for reasons that no longer apply. Research has long shown that traits that make someone an excellent support rep, patience, agreeableness, a genuine wish to soothe, can sit uneasily alongside what a strong salesperson needs in the moment, which is a feel for when to press, when to hold ground, and when to make an ask.
Very few humans can hold both temperaments convincingly in the same conversation, so companies split the roles. An AI agent has no fixed personality. It can be patient in one paragraph and assertive in the next, follow a threat of churn with a fix, and follow a fix with a fair upsell, without the tonal whiplash a customer would notice from a human doing the same pivot. That flexibility makes the case for fewer walls.
Specialized agents still earn their place, provided they work behind the scenes. Narrow, well-trained microagents remain the most efficient way to handle deep, specific tasks like inventory lookups, fraud checks, or payment processing. The trouble starts only when each one is allowed to face the customer on its own.
What those agents need is a single orchestration layer in front of them, and that layer has two jobs. It keeps the brand speaking in one voice no matter which system is answering, and it checks what any individual agent says before a customer ever sees it. Behind that front door, the org chart can be as sprawling as the business needs, and the customer never has to see it.
When no one is minding the voice
That second job gets overlooked most often, and two recent failures show what happens when an agent speaks for a brand with nothing standing behind it.
Last year, Cursor’s AI support agent, which signed its emails as “Sam,” told a customer that getting logged out across devices was expected under a one-device-per-subscription policy. The policy didn’t exist, and the logouts came from a technical glitch. Sam’s answer spread across Reddit as though it were an official announcement. Customers canceled publicly on social media, and the company had to correct the record and issue refunds. A layer that checked answers against actual policy would have caught the invention before anyone acted on it.
DPD, the Germany-based delivery company, ran into the same gap from the other direction. A couple of years ago, after a system update, a customer coaxed its chatbot into using profanity, writing a poem trashing the company, and calling DPD the “worst delivery firm in the world.” DPD blamed the update and switched off the AI element. With nothing holding the brand’s voice in place, the agent adopted whichever voice the customer asked for.
Cursor’s agent made up a policy, and DPD’s agent made fun of its own employer. Each was a single agent, and each broke for the same reason. Spread that exposure across five customer-facing agents with five sets of instructions, and the odds of a customer finding a crack go up with every handoff.
Finding hidden revenue
Keeping the voice consistent protects the brand. Giving that voice the full context is where the money starts. A customer asking about a return is often a customer worth re-engaging, whether that means a better-fitting size, an upgrade, or a product that pairs with something they already own. A siloed support agent can’t see any of that, because purchase history and live recommendations belong to someone else’s agent. An orchestrated one can, and the brand gets a sales opportunity inside a conversation it was already paying to have.
Hanssem, a South Korean furniture retailer, shows what becomes possible when one agent can see the whole picture. Back in 2022, the company set out to remake itself as a living tech company, betting that digital transformation and better customer experience would drive its next stage of sales growth. AI agents have since become one of the clearest payoffs of that bet. Conversion rose tenfold when AI agents handled product recommendations in place of human reps because the agent could hold the full catalog, a shopper’s purchase history, and the context of the conversation at once.
Results like that took time to build. In the first week of its proof of concept, Hanssem’s agent resolved fewer than half of inquiries. After weekly rounds of iteration, it was resolving 86%.
Hanssem got there by testing every week, demonstrating that CX leaders can start with a few simple tests of their own. Count how many times a customer has to repeat an order number or re-explain the problem as the conversation moves between agents. Ask the support agent about a customer’s last purchase and see whether it knows. Then try to talk each agent out of the brand voice, the way DPD’s customer did, and note which ones hold. Every repeated question, blank answer, and broken voice marks a seam the customer can already feel.
Customers are never going to complain about “poor agent orchestration”. They were only ever going to feel, correctly, that they were talking to five different brands instead of one.
John S. Kim is CEO of Delight.ai, part of Sendbird Inc.
