Skip to content

The Dock, Accenture · 2025 · 5 min read

AI-Driven Assessments

RoleLead Interaction Designer
Team2 designers, 3 engineers
Duration10 weeks
DeliverableConversational survey platform
Outcome518 completions across 5 campaigns — now a repeatable platform used across multiple teams

How do you show AI thinking to someone who didn't sign up for AI, and still have them trust it?

Problem statement

The brief came from a business development need: assess a company's AI readiness in a legal context before any engagement has started. The problem was that every company starts somewhere completely different. Some still run their contract management lifecycle with physical filing cabinets. Others have a full CLM system. Some are already agentic. Each of them needs a completely different conversation. The survey was what made that conversation possible.

My role

I designed the full experience end to end: the question architecture and adaptive branching logic, the conversational onboarding, the topic-based progress sidebar, and the chapter summary pattern — the moment that made the AI's reasoning reviewable rather than opaque. There was no existing template for a conversational survey that worked this way. The interaction model was built from first principles.

Process

The LLM decides when an intent is satisfied and generates the next question from what the person has already said — every session looks different. The hard design question wasn't how to build a survey. It was how to make that reasoning visible rather than mysterious. I explored three approaches. First, an artefact that updated live after every question: too complex, too distracting. Then separate artefacts after each block: still too much. We landed on chapter summaries, the AI synthesising what it had learned before moving on. That rhythm worked. It turned the AI's logic into something reviewable, something you could look back at and understand.

For progress, I designed a sidebar that showed topics rather than time. We had no idea how long any given session would take — the LLM's routing depended entirely on what each person said. Topics appeared at low opacity and sharpened as you approached them. If an answer fed back into an earlier topic, that topic would be re-emphasised. The sidebar showed the shape of the conversation, not a countdown.

The survey itself scrolled question by question, open and wide. More conversation than chat interface. The onboarding was detailed on purpose, but it took a revision to get there. The first version treated it like any other survey intro — no mention that an AI was actively routing based on what you said. People would answer without knowing their session was being shaped in real time. That's a different trust problem than normal survey design. The final onboarding made the agentic nature explicit before the first question.

I ran two concept testing sessions with the legal team before building out the full flow. The conversational format was a real risk — legal contexts tend to expect formal, structured input, not an adaptive conversation. The sessions confirmed it worked, which wasn't a given.

Outcome

What started as a business development tool became a repeatable platform, picked up by different teams for different purposes without rebuilding it each time. 518 completions across five campaigns — before this, BD discovery relied on manual calls: variable in quality, time-intensive, hard to repeat across teams.

Your Campaign is Ready to Launch — completion screen focused on forward momentum, not compliance confirmation
Your Campaign is Ready to Launch — completion screen focused on forward momentum, not compliance confirmation

Dynamic vs. Static — the core proposition that framed the entire design direction
Dynamic vs. Static — the core proposition that framed the entire design direction

The more you answer, the smarter it gets — the value proposition screen that set expectations before the first question
The more you answer, the smarter it gets — the value proposition screen that set expectations before the first question

Let's get to know you and your organisation — conversational onboarding that establishes trust before asking anything
Let's get to know you and your organisation — conversational onboarding that establishes trust before asking anything

Tell us a bit about yourself — context-gathering framed as orientation, not interrogation
Tell us a bit about yourself — context-gathering framed as orientation, not interrogation

Which contract-related performance metrics do you track? — multi-select that feeds the adaptive branching logic
Which contract-related performance metrics do you track? — multi-select that feeds the adaptive branching logic

Voice input on an open-ended drivers question — removing the typing barrier for complex, contextual answers
Voice input on an open-ended drivers question — removing the typing barrier for complex, contextual answers

Chapter Completed — the AI's synthesis of everything shared, reviewable before continuing
Chapter Completed — the AI's synthesis of everything shared, reviewable before continuing

Dynamic vs. Static — the founding constraint: every company starts somewhere different, so every session had to look different too

Dynamic vs. Static — the founding constraint: every company starts somewhere different, so every session had to look different too

Conversational onboarding — the adaptive questioning only works if the first question doesn't catch someone off guard; trust had to come first

Conversational onboarding — the adaptive questioning only works if the first question doesn't catch someone off guard; trust had to come first

Adaptive branching — multi-select as the minimum signal the LLM needed to route the next question; single-select would have cut off the branching too early

Adaptive branching — multi-select as the minimum signal the LLM needed to route the next question; single-select would have cut off the branching too early

Chapter Completed — the third approach tried; a live updating artefact was too distracting, post-block summaries still too much; reviewing before continuing was the rhythm that worked

Chapter Completed — the third approach tried; a live updating artefact was too distracting, post-block summaries still too much; reviewing before continuing was the rhythm that worked