Principles for using conversational and deterministic UI
One of the big challenges with building agentic AI-enabled services is discerning how and when to use conversation together with deterministic UI to create intuitive, genuinely useful user experiences.
In AI Studio, we’ve been exploring when and how to use each approach respectively, and how they interact. What’s the difference between conversational and deterministic UI? Traditionally, AI-enabled chat systems have communicated with users in ordinary everyday language and conversational UI.
Users conversing with a chat system can ask a question or make a statement like they are talking to a person. The AI system interprets their intent and steers the conversation accordingly.
Agentic AI – that is, AI that can go one step further and take actions on behalf of its users – is able to call deterministic functions within a system. This means that more than just chatting, it can – with its user’s permission – perform tasks and has capabilities.
To do this, AI agents often require specific information so that they can achieve a given task with higher confidence, and we can use deterministic UI to help collect this information.
Unlike conversational UI, deterministic UI uses more traditional UI elements like radio buttons, checkboxes, buttons and text inputs. They provide explicit choices and clear affordances to let users define or modify specific datatypes within a larger system:
- “Click X, and Y will happen”
- “Enter A, and the system will validate it against set rules and act on it”
Principle 1: Use conversational UI to exchange general information
Conversational UI is a really efficient way of gauging a user’s broad intent. This is usually helpful at the beginning of a conversation.
For example, a person meeting with a financial advisor for the first time would likely start with an informal conversation about their overall objectives. Then, as the meeting progresses, they may move on to filling in a form about their income and assets, before signing an agreement that allows the financial advisor to make investments on their behalf.
In the same way, a conversation with an AI agent would likely begin with an informal discussion, before possibly progressing to action.
In the early stages of a conversation, a user might say something like “I’ve hurt my leg, I can’t work at the moment and I’m worried about money”. The AI agent would most likely use conversation to interpret their objective, and provide generalised responses like “you might be eligible for Universal Credit”.
As the conversation progresses, the AI agent could begin to narrow down its user’s goals.
An important feature of traditional conversational UI is that it only allows an agent to talk about a service’s capabilities but, usually, it needs specific data in order to act upon them correctly.
With the user’s permission the agent may move towards actionable output, using deterministic UI to collect the discrete pieces of data that it needs to better serve the user.
Principle 2: Use deterministic UI to exchange specific information
Deterministic UI performs 2 important roles:
- it signifies to the user that they are moving from discussing and planning into doing
- it captures data in a way that it can be validated, saved, modified or acted upon
It moves the conversation from informal to formal, and allows users to provide accurate information in exchange for accurate responses and enables the agent to accurately perform functions on the user’s behalf.
Deterministic UI is especially important when an AI agent needs to ask for specific data points during a conversation.
For example, the agent could ask “do you already have a fit note (medical certificate) saying you’re unable to work?” and provide the options “Yes” and “No” as radio buttons.
When using deterministic UI to capture data, it’s important to clearly show:
- how the data is being used
- that data is being captured
- where it can be accessed, reviewed and changed
We are experimenting with the best way to do this. In one prototype, we’ve designed an approach where a user can easily access the data and modify it in their account page.
Get in touch
If you’re working on AI-enabled chat interfaces and you’d like to stay up to date with the experiments we’re doing, you can follow along with our progress on this site, or get in touch with us at govuk-ai@dsit.gov.uk.