How do I get my contact centre ready for customer AI agents?
The short answer
Start by assuming customer AI is already reaching you, then design for it on purpose. Make your own AI declare that it is AI and what it can do. Give your team a protocol for when a customer's AI calls. Check an agent's delegated authority before acting on it. Publish what AI agents may do with you. And design the hand-off to a person, with context, for when things don't line up.
1. Find where customer AI already reaches you
Customers already use AI to draft complaints, check your replies for the catch and, increasingly, to place the call. Most of it never shows up in CX analytics, so start by asking where it would be invisible today: inbound email, chat and voice.
2. Make your AI declare itself
Your AI should say it is an AI, state its scope and hand off when a request is out of scope. The strongest operations declare this in a machine-readable way, so a customer's AI can read it too. Article 50 of the EU AI Act makes disclosure a legal requirement from August 2026.
3. Give your team a protocol for customer AI
When a customer's AI reaches a human agent, most teams don't know: the AI passes verification and the agent assumes it is the customer. Write a protocol, train on it, and over time instrument your channels to detect agentic input and route it to an AI-aware pathway.
4. Check delegated authority before you act
An AI acting for someone is not the same as that person. Before acting, check the delegation: signed authority, scope and beneficiary. Watch for triadic cases, such as a daughter who set up an agent for her father's account. Setting up an agent is not authority over someone else's account.
5. Publish what AI agents may do with you
If your only AI-facing surface is your website and phone number, incoming agents have to guess. Publish a declaration they can act on, such as a Service Handshake declaration, and test it against real agent traffic.
6. Design the hand-off and the impasse
When an interaction stalls, it should go to a named human pathway with full context and a stated SLA, not loop or drop into a general queue. The best-designed pathways add a breach chase and a backup associate on standby.
Common questions
How do I know where my operation stands?
Take the Dual CX Diagnostic. It is eight questions, takes about two minutes, and gives a scored readiness view for each of the four modes plus a recommended starting point.
Do I need new technology?
Mostly not. The State of Dual CX audit found the gap is not a technology gap: the technology to declare, verify and escalate is off the shelf. It is a design and governance gap.
What does the EU AI Act require of contact centres?
Article 50, in force since 2 August 2026, requires that people are told when they are interacting with an AI, unless it is obvious from context, and that AI-generated content is labelled in a machine-readable way.