AI Design: Should the agent speak like your team?

The Importance of Conversation Design for AI Agents

In the world of customer interactions, an AI agent that has not been trained to communicate like your team risks coming across as impersonal, like a simple language model (LLM). This is where conversation design comes into play, an emerging discipline that has become essential in AI-focused support teams. The role of the conversation designer is crucial: they define the tone, structure, level of detail, customer experience, as well as the processes for transfer and escalation.

Without clear guidelines, an AI agent might make inappropriate decisions, such as providing overly detailed responses or adopting a neutral tone in the face of a frustrated customer. These mistakes can lead to measurable costs. Customers faced with poorly structured responses may lose trust, even if the information is accurate, and will prefer to escalate to a human team member for a more satisfactory answer. Moreover, a poorly managed transfer can leave an already irritated customer in the hands of a human representative. Therefore, conversation design is essential to avoid these pitfalls.

The Measurable Impact of Conversation Design

At Fin, the importance of conversation design was demonstrated through an A/B test. Two opening messages were compared: one warm and conversational, the other more formal. The result was clear: the welcoming message increased the CSAT score from 72.8% to 78.4%. This simple adjustment in conversation design, applied from the first contact with the customer, had a significant impact.

Key Areas of Conversation Design

Conversation design covers five main areas, each influencing a different facet of the customer experience:

• Tone and Personality: This includes the agent's voice, level of detail, and degree of formality or friendliness, which can vary depending on the situation.

• Response Structure: The agent must adjust the level of detail to the question asked. They should know when to escalate, how to communicate the transition, and what context to convey.

• Interaction Flow: The progression of a conversation, from question to answer, then to resolution or transfer, must be smooth.

• Response Quality: Responses should be clear, helpful, and aligned with the brand image, even if they are technically correct.

Implementing Conversation Design

Defining the Conversation's Tone

Before refining individual responses, it is crucial to define the agent's voice. A simple paragraph can suffice to establish how the agent should sound. This reference point is essential for maintaining tone consistency.

Different types of conversations require varied registers. For example, a customer having trouble accessing their account needs direct and quick responses, while a customer exploring a new feature might appreciate more context. The voice should remain consistent, but the register must adapt.

Carefully Designing Transfers

The transition from the agent to a human representative is a critical moment. Customers should not have to repeat their issue. The representative must receive the complete conversation history, the context of the problem, what the agent has already done, and the reasons for escalation.

How the agent communicates the transfer is also important. A phrase like “Let me connect you with a team member who can help you with that” is more reassuring than a silent transfer.

Not Overlooking Follow-ups

Follow-ups are just as important as transfers. If a conversation is interrupted, whether with the agent or a representative, it is crucial to reach out to the customer to ensure they received the necessary help. Many teams overlook this aspect, but customers notice.

Knowing When the Agent Should Be Silent

A common mistake in conversation design is over-explaining. The agent, having access to a wealth of information, can easily provide more details than necessary without clear guidelines.

The agent should adjust the level of detail to the customer's request. For example, a person asking how to reset their password does not need three paragraphs of explanation. Conversely, a customer inquiring about a complex integration might need more information. If more information is available, it should be offered rather than provided upfront.

Designing for the Conversation the Customer Is Having

Customers do not follow scripts. They may change direction mid-conversation or ask follow-up questions unrelated to their initial inquiry.

The agent must be able to handle these transitions without forcing the customer to return to a fixed flow. If the agent insists on resolving the original question while the customer has moved on, it can give the impression of speaking to someone who is not listening.

Keeping Instructions Concise

One of the major challenges is not overloading the agent with instructions. Teams tend to add rules every time a new specific case arises. Quickly, the language model ends up with paragraphs of instructions to process before it can respond.

It is essential to know when to stop. If it concerns content or information, it should be integrated into the knowledge base. If it concerns tone or managing specific situations, it should be included in the agent's instructions. For example, “Be direct about pricing” is more effective than a long paragraph explaining the communication philosophy regarding pricing.

In the case of Fin, much of this work is done in Guidance, where conversation design takes shape, helping to define how the agent should sound, how much it should say, and how it should respond in different situations.

Starting Without a Dedicated Hire

Most teams will not hire a dedicated conversation designer from the outset. However, it is crucial that someone is responsible for how the agent communicates, even if it falls under an existing role.

Conversation design often begins within support operations or knowledge management. Someone on the team starts paying attention to how the agent sounds. Over time, as the agent handles more conversations, this becomes a formal responsibility, then a dedicated role.

1. Designate a Responsible Person: Someone must be explicitly responsible for the agent's communication, even if they are not a new hire.

2. Choose a Problematic Conversation Type: Identify conversations where the agent responded correctly but the customer still escalated or left negative feedback. Start with these.

   If you are using Fin, the CX Score can help you identify this information. It shows which topics and types of conversations receive poor scores, as well as the reasons behind these scores, so you can see if the issue relates to response quality, customer effort, or something else.

3. Audit the Agent's Instructions: If they have exceeded a few targeted rules, reduce them. Move content into the knowledge base, and keep instructions focused on behavior.

4. Improve the Worst Transfer: Review a few conversations where the agent escalated to a human. Did the customer have to repeat themselves? Did the support representative have enough context? Redesign this unique transition first.

Small Steps That Accumulate

The impact of each of these improvements accumulates. A warm opening message improved our CSAT, while reducing instructions made responses more precise. Designing a better transfer allowed support representatives to no longer inherit frustrated customers.

None of these changes required new knowledge; they required someone to pay attention to the conversation itself.