Closer Look · About 8 min · 4 chapters
ieso’s NHS service had shown that typed human-to-human therapy could work. That success shaped the AI product: the prevailing belief was that a sufficiently capable chatbot could deliver the whole experience. Yet live product data and research were showing the limits of chat for some therapeutic moments.
Rather than continue to argue for a different approach in the abstract, I started building and testing alternatives. The question became: what should someone understand, do or feel at this moment—and which format would help them most?
Working prototypes made the trade-offs visible. Once teams could see how people responded to voice, structured reflection and richer media, moving beyond the chatbot stopped feeling speculative. The findings informed three principles and a set of reusable patterns that Clinical, Product, Design and Engineering could apply.
I’m a visual learner… the videos were helpful to explain the principles better.
Study participant
[The message bubbles] feel like too much information. If I'm feeling anxious, I think maybe I’d struggle to focus.
Study participant
What We Learned From the Comparison
The varied version was easier to follow and made room for different ways of learning. The research did not remove conversation from the product; it stopped chat from being the automatic answer to every therapeutic moment.
01
Less text-message styling made longer clinical content easier to read.
02
Video and audio made some ideas clearer and more engaging.
03
A session could still feel personal without presenting every line as chat.
04
Different moments needed different formats—not a single default.
People using digital therapy may already be tired, anxious or struggling to concentrate. The principles gave us a practical way to respond: present ideas clearly, make it obvious what to do next and ensure the experience still felt personal.
Clear, frictionless and personal—the principles used to review each digital therapy session
Turning Principles into Working Patterns
For each session, I reviewed the clinical content with the team, mapped the experience and looked for moments where another format could make the therapy easier to understand or take part in. Prototyping those moments revealed patterns we could reuse rather than inventing every session from scratch.
Mapping a therapy session to find where another format could improve understanding or engagement
When Showing Worked Better Than Explaining
Some clinical concepts became long, tiring chatbot explanations. I worked with Clinical and an animator to turn key concepts and therapeutic metaphors into short videos, audio and animation that could explain an idea or guide an exercise more clearly. The narrator was an AI voice generated with ElevenLabs. I worked with Clinical to write the scripts and spent time tuning the narrator’s pronunciation, pacing and emphasis until it sounded calm and natural. Some colleagues were initially uneasy about using an AI voice, but people using the live product responded warmly: they praised the voice, found it calming and often singled the videos out as their favourite parts of the experience.
The Radio of Your Mind—using a visual metaphor to make cognitive defusion easier to grasp. Includes audio.
The videos were really powerful.
User of the live product
I liked the metaphors the best. They explained things so well, in such an interesting way.
User of the live product
I could just sit down and… relax, listening to what he was saying.
User of the live product
Turning Someone’s Own Words Into an Interactive Exercise
Sessions could still provide the personal context without remaining in the chat interface. In one cognitive-defusion exercise, the product reused words someone had shared earlier, then let them stretch, shrink and jumble those words on screen. The interaction turned an abstract therapeutic idea into something people could see and manipulate. It connected the conversation to a focused interactive experience. This pattern was better suited to experiential exercises, where people are asked to notice, feel or try something in real time: the interface could guide the moment and make the result visible, without asking someone to work through another long exchange of messages.
A challenging thought shared earlier became part of a personalised cognitive-defusion exercise
In-the-moment support was designed for moments when someone needed help now—perhaps during a flare-up or in the middle of the night—not for a scheduled therapy session. Research suggested people needed an easier way to explain what was happening, but voice was seen internally as technically complicated and lower priority than the product’s existing challenges. I felt that typing asked too much of someone at exactly the wrong moment. Rather than try to communicate the idea through static screens, I built a working prototype in Claude Code, using the Claude API for the interaction and ElevenLabs for speech. I experimented with different approaches to reduce latency until the conversation felt natural enough to use through AirPods while walking. Experiencing it changed the conversation. The prototype was used in customer demonstrations, where people responded enthusiastically and began asking when voice would be available. It turned voice from an abstract technical burden into an experience that Product, Engineering and customers could understand and evaluate.
In-the-moment support combined conversational support with short exercises for difficult moments
What Voice Couldn’t Solve
Voice made the interaction easier, but it could not rescue an experience offering the wrong kind of help. Two participants arrived wanting practical answers, so a reflective exercise felt evasive. The issue was the fit between their intent and the exercise—not whether they spoke or typed.
The Values exercise was one of several Foundations activities designed to run alongside the main therapy programme. Together, they helped people identify what mattered to them, recognise patterns and turn those insights into practical next steps. Those outputs could then make later support more personal. Rather than turn the exercise into another chatbot exchange, I explored it as an AI-guided activity: one decision at a time, personalised examples when questions felt abstract, earlier answers brought back at the right moment and a visible summary at the end. This was not about using less AI. AI still interpreted free-text answers, suggested values in each person’s own language and briefly shifted into conversation when someone needed help getting unstuck.
The AI-guided Values activity combined structured steps with personalised recommendations and conversational help.
Voice Improved the Conversational Version—But the AI-Guided Activity Still Performed Better in Testing
To test whether typing had held the conversational version back, I added optional voice and tested it again.
Adding voice made the conversational version easier to answer, but the guided worksheet still made the exercise clearer and more useful.
Finding
Voice Reduced Input Friction
Dictation made longer answers easier for some participants.
Comparison
Structure Still Mattered
Speaking did not make the conversational version feel clearer or more useful. The AI-guided activity remained the stronger format.
Learning
Combine Structure With Responsive AI
The stronger approach used visible structure, with AI interpreting, personalising and helping someone get unstuck.
