When several AI agents handle one customer, the failures cluster in the handoffs between them. Each agent can work well on its own and the customer can still have a bad experience, because the seams between the agents have no owner. Naming that owner is the work most teams skip.
A modern contact centre rarely runs one AI agent. It runs several: one for billing, one for orders, one for technical questions, plus the routing between them and the handoff to a human. Each is bought, built, or tuned by someone. The connections between them usually are not.
This article is about those connections, why they fail, and who should own them.
The failure lives between the agents
Test each AI agent alone and it can look fine. The billing agent answers billing questions. The orders agent answers orders questions. The trouble starts when a customer's problem crosses both, and they are handed from one agent to the next.
At that handoff, the context can fail to travel. The customer repeats their account number and their problem. The second agent starts cold. Or the routing misjudges the problem and sends the customer to the wrong agent entirely. None of this is a fault inside any single agent. It is a fault in the space between them, and that space is what nobody tested.
Orchestration is the name for the seams
Juan Martin Maglione, writing in the Service Matters newsletter, separates orchestration into distinct layers: the process that governs the workflow, the agents that execute steps, and the model underneath. The handoffs are the process layer, and they are a real job, not a thing that happens automatically because the agents share a platform.
Mark Levy, in Decoding Customer Experience, makes the consequence concrete: he argues you should stop calling it orchestration if the customer has to start over. If the customer is the one carrying their context from one agent to the next, the orchestration exists only in name. The seams were never owned.
Why the seams have no owner
The reason is how the work is divided. The billing agent has an owner. The orders agent has an owner. The platform has a vendor. Each owns a box. The arrows between the boxes belong to no box, so the handoffs, the routing, and the escalation to a human belong to no owner.
When a customer has a bad multi-agent experience, the post-incident review asks each agent's owner, and each can honestly say their agent worked. The failure was real and nobody owns it, because the org chart was drawn around the agents and not around the seams.
Draw the org chart around the seams
The number of AI agents in a contact centre will keep rising, and so will the number of seams between them. A structure that assigns an owner to every agent and no owner to any handoff will keep producing failures that the review cannot pin on anyone.
The fix is to treat the handoffs as a real piece of the system, with a named owner and a clear duty, the same way each agent already has one.