Recognize the need
Does the service understand what the customer is trying to accomplish? Look for relevance, context, missing information, correction, and a credible interpretation.
01Contributor · AI interaction
Ava
A mobile assignment exposed a larger AI service question.
Ava explored how customers might interact with an AI banking service through conversation, speech and supporting graphical controls. The work quickly raised a larger question: how do you judge the quality of a conversation that can take many different paths?
Role
Mobile Experience Designer · Contributor
Contribution
Conversation-quality framing and evaluation
Stage
Concept and evaluation
Context · Early AI banking concept
A different kind of interaction.
Unlike a predefined interaction or chatbot flow, the conversation could develop in many directions. That made a basic design question much harder to answer: what makes one of those conversations good?
I focused on making that quality examinable: what does the service need to recognize about the customer’s situation, and what does it need to be capable of resolving?
Inquiry · Defining conversation quality
Criteria, not an ideal script.
Existing conversation-design guidance offered useful principles, but Ava could not be judged against a single scripted flow. Each response could change what happened next.
To understand what quality should mean, I looked across conversational-design principles, the behavior of language models, and human service conversations in branches and customer support.
Across them, one distinction kept returning: understanding what someone needs is different from being able to help.
I used that distinction as the basis for an evaluation model. Rather than asking whether a conversation followed an intended flow, the criteria examined whether it remained relevant, developed enough understanding, and moved toward a useful outcome.
Resolve the need
Can the service fulfil the request or make its remaining boundary clear? Look for sufficient information, available workflow, transparency, and closure.
Design practice · Capability shown through the work
What the evaluation revealed.
Applied with researchers, the criteria helped distinguish problems that could be improved through conversational behaviour from problems where the service lacked the knowledge, workflow or capability required to help.
Conversation issue
The service misunderstood or lost context.
- → conversational intervention
Knowledge issue
The response lacked information required to resolve the need.
- → service/content dependency
Capability issue
The service understood the request but could not perform the required action.
- → underlying service dependency
Reflection · Later lineage
Conversation revealed the service system.
Ava changed the scale at which the work was understood. A response could be polished while the service remained unable to support the customer’s progress.
The lasting question was therefore not only how an AI should speak, but what it must know, what it must be able to do, and where its responsibility must end.
“The quality of an AI conversation is inseparable from the quality of the service participating through it.”
Original working material and research outputs remain internal. Public examples reconstruct the supported project logic.