Conversational Agents Need Service Design, Not Just RAG

A customer service conversational agent is not only a voice interface connected to an LLM.

In a real service environment, the agent sits inside an operating system made of authentication rules, human operators, digital channels, escalation paths, back-office constraints, customer expectations and risk controls.

This is why some voice AI experiences feel surprisingly weak even when the underlying technology is advanced.

The issue is rarely just whether the model can produce a plausible answer. The deeper issue is whether the organization has designed the conversational agent as part of the service architecture.

The Voice Experience Is Part Of The Product

In a phone call, interaction quality matters immediately. The user notices latency. The user notices whether the agent can be interrupted. The user notices awkward waiting sounds, unnatural transitions and long pauses after a problem has been explained.

Barge-in is a good example. In voice systems, barge-in is the ability for the user to interrupt the agent while it is speaking. Without it, the conversation feels less like a service interaction and more like waiting for a script to finish.

This is not a superficial user-interface detail. It changes the trust relationship. If the agent cannot be interrupted, cannot recover quickly and cannot respond with natural turn-taking, the customer starts to feel that the system is optimizing for containment rather than resolution.

RAG Is Not A Service Design Strategy

Many customer support agents rely on retrieval augmented generation, or RAG. In simple terms, the AI system retrieves relevant support documents and uses them as context for the answer.

RAG is useful, but it is not enough. A knowledge base can help the agent explain policies, procedures and options. It cannot, by itself, decide the right channel, authenticate the customer, send a secure link, escalate correctly or recognize when the documented procedure does not match the real operational constraint.

If the retrieved content is poorly structured, outdated, incomplete or written for internal staff rather than customer interaction, the agent may confidently guide the user through the wrong path.

That is not only an AI problem. It is a knowledge governance problem.

Voice Is The Wrong Channel For Some Instructions

One common mistake is asking the conversational agent to verbally guide a customer through a complex digital journey: open this section, find this menu, click this option, then follow this path.

That may sound reasonable from inside the call-center script, but it is often poor service design.

If the task must be completed in internet banking, a portal or a secure web area, the better design may be omnichannel orchestration: the agent should send an email, SMS, app notification or secure message containing the correct authenticated link or deep link.

The voice channel should clarify intent, explain constraints and coordinate the next action. It should not force the customer to memorize navigation instructions that the system could provide in a more precise digital form.

Tool Calling Is Where The Agent Becomes Operational

A conversational agent becomes more useful when it can call tools, not only generate answers.

Tool calling means that the AI system can trigger controlled actions through approved software interfaces: create a support case, send a secure email, generate a callback, retrieve the status of a request, route the user to the correct authenticated page, or escalate to a human operator with context.

This is where conversational AI moves from explanation to operational capability.

But tool calling also requires governance. The organization has to decide which actions the agent can perform, under which authentication level, with which audit trail, and when human control is mandatory.

Escalation Is Not A Failure

In customer service, escalation to a human operator is often treated as a fallback. For conversational AI, that mindset is dangerous.

Escalation is part of the service design. If the agent gives instructions that contradict a previous human operator, refuses to transfer the call when the user asks, or insists that it can solve a problem it cannot actually solve, the system destroys trust.

A well-designed agent should know when it is useful and when it is not. It should preserve context, route the customer correctly and make the human handoff feel like continuity, not a restart.

This is especially important in regulated or high-trust sectors such as banking, healthcare, insurance, public services and enterprise support.

What Organizations Should Design Before Scale

Before deploying a customer service conversational agent at scale, organizations should design six layers together.

1. Interaction Quality.
   Latency, turn-taking, barge-in, repair behavior and the acoustic experience determine whether the user feels heard or trapped.
2. Knowledge Governance.
   RAG documents need ownership, versioning, testing and service language. Internal procedure documents are not automatically good customer-facing context.
3. Tool Calling And Channel Orchestration.
   The agent should trigger the right operational action through the right channel instead of forcing every instruction through voice.
4. Authentication And Risk Boundaries.
   The system needs clear rules for what can be discussed, shown, sent or executed before and after customer authentication.
5. Human Handoff.
   Escalation should be intentional, context-preserving and easy to trigger when the customer request exceeds the agent's authority or certainty.
6. Feedback Loops.
   Failed journeys, wrong instructions and repeated escalations should update the knowledge base, service process and training program.

Conversational AI Is An Adoption Problem

The best conversational agents are not built only by improving the model. They are built by connecting AI engineering with service design, process ownership, compliance, data governance and organizational adoption.

This is also why internal capability matters. Customer service teams, compliance teams, digital product teams and technical teams need a shared language for what the agent is allowed to do, what it should know, how it should escalate and how the service should improve over time.

In my own work on Sophia, a conversational agent intended to evolve toward a personal assistant, these design questions are not secondary. Low-friction voice interaction, interruption, tool use and operational context are part of the standard, not optional polish.

A conversational agent should not merely sound intelligent. It should help the organization deliver the right action through the right channel with the right level of trust.

Read more ideas on AI, agentic systems and adoption.