Customers Shouldn’t Have to Say It Twice
Companies have poured money into AI for customer service, but far less thought into how those systems pass customers along. Avaya’s Strategic Head of AI and Machine Learning explains where context gets lost, why more AI agents make it worse and how to start fixing it.
A customer asks a virtual agent about an unexpected bank charge. The virtual agent identifies the transaction but cannot resolve the dispute, so it transfers the customer to an employee. The employee opens the conversation with an all-too-familiar question: “How can I help you today?”
The customer has already explained the problem, and now they have to explain it again.
Enterprises have spent years adding AI to customer communications. Virtual agents answer questions, AI assistants support employees and newer AI agents can carry out tasks across enterprise systems, including making decisions and taking actions within defined permissions. Yet the connections between these capabilities often receive less attention than the capabilities themselves.
That creates a gap between what an individual AI system can accomplish and what the customer experiences. A virtual agent may answer accurately or route a request correctly, while leaving the next person without the information needed to continue.
As enterprises introduce more AI agents, those gaps can multiply. Improving the experience requires a deliberate approach to how context, responsibility and work move between systems and people.
Context is more than a transcript
Customers see one relationship with a company. Whether they move between channels or from AI to an employee, the context from previous interactions should follow them.
In my conversations with enterprises, I increasingly hear the same request: zero repetition. Customers should not have to explain who they are, what they are trying to accomplish, and what has already happened every time they switch to another channel, AI agent, or human employee.
While a transcript can help, it still requires someone to read through a conversation to reconstruct the situation. A handoff should make several things clear, including what the customer wants, which information has been verified, what actions have been completed and what remains unresolved.
For the disputed bank charge in my opening example, that might include the transaction in question, the customer’s explanation, the policy the virtual agent consulted and the reason it escalated the request. The employee should also know whether a credit was discussed, requested or issued, because each requires a different next step.
Context also sometimes extends beyond the current interaction. It can include both what is happening in the conversation and the customer’s history with the company. Combining the immediate details with relevant insights from previous interactions across channels gives AI agents and employees a fuller understanding of the customer and their needs.
More agents make coordination more important
Splitting a request among AI agents sounds straightforward, until one of them runs into a problem. Behind the virtual agent in the disputed charge example, separate AI agents might retrieve account records, check dispute policies or process an approved credit. What happens if the account record is incomplete? Does the AI agent that retrieves it wait for the missing information, or send the request to an employee? If another agent has already issued a credit, how will the employee know? These details need to be worked out before the agents start handling customer requests.
Orchestration gives enterprises a way to manage that work. It determines which agent takes the next step, what information it needs and when an employee should get involved. Deloitte identifies this coordination as essential to getting value from multi-agent systems.
Importantly, not every AI agent needs access to the customer’s entire history. Just as employees have different responsibilities and access permissions, AI agents should receive only the context needed for their specific tasks. Defining those boundaries helps agents work together while limiting unnecessary access to information.
Context must follow the customer
Customers should not have to relay information between departments. When a request needs an employee’s approval or judgment, the employee should receive the relevant records, actions already taken and unresolved questions.
In the disputed charge example, an employee approves a credit after reviewing the transaction. That decision needs to appear in the billing system and the customer’s account history so anyone handling a later inquiry knows what was approved and whether the credit was applied.
Whether customers interact with AI or a person, they should be able to continue their request without having to explain it again.
Start with the places where context breaks
Ask your support team to pull a few cases that took longer than they should have, then work through what happened. You may find that an employee spent half the call looking for information the customer had already provided in chat. That’s a handoff worth fixing, and a useful place to start when deciding how AI should fit into the process.
Once you’ve found the gap, work out what the next system or employee actually needs. In that chat-to-phone handoff, that may include the customer’s explanation and a record of anything the AI agent already tried. They also need access to the account or application to resolve the issue. A summary alone won’t help if the employee still has to chase down the underlying information.
This gets harder when the tools come from different vendors or were built by different teams. Most enterprises already operate in that kind of environment. Before introducing another AI agent, check whether it can use the records you have and leave a usable account of its work for whoever takes over.
Testing should include those transitions. Have someone begin a request in chat, switch to the phone and ask for an employee. See what reaches the employee and what they have to ask again. Test what happens when an action fails, or the customer changes their request midway through. Those are the moments when missing context becomes apparent.
After deployment, check whether customers still call back about the same problem, and ask employees where they spend time reconstructing earlier conversations.
A customer calling about a disputed charge wants to know whether it has been corrected. If your AI can check the account, help an employee reach a decision, and ensure that decision reaches billing, it has helped resolve the request. That’s the outcome to build toward.
ALSO READ: AI, Creativity, and the Risk of “Meh”
