Paul TinkerProduct designer · Design systems Back to portfolio

AI-native products & design systems

A working log of designing and building digital products with AI

I designed TOBi, Vodafone's chatbot, before today's generative models. Looking at it again: what changes when the AI can say anything, and what doesn't.

  • Vodafone 2017 to 2020, revisited
  • One journey, six steps
  • Three patterns
  • 21 expressions
  1. The questionA customer asks TOBi why their bill is higher than usual.
  2. Shows how sure it isTOBi shows what it's checking, then what it found. It's sure about the £10.00 of calls, and not sure about the £2.50.
  3. Asks you to checkTOBi suggests a plan that covers those calls and shows exactly what would change. Nothing happens until the customer confirms.
  4. ConfirmedThe customer confirms, and the new plan starts with the next bill. TOBi says what it did and emails the details.
  5. Knows its limitsTOBi can't explain the £2.50, and its face says so. It offers a person straight away, and the customer chooses what to pass on.
  6. Hands over with contextAn adviser joins with TOBi's summary, so the customer doesn't have to explain again. The adviser sorts out the £2.50.
1 / 6

I designed TOBi, Vodafone's AI chatbot, between 2017 and 2020, before today's generative models. Looking back at it now, the question is: what changes when the AI can say anything?

Then, the journeys were designed. TOBi was built from a kit of conversation components: messages, quick replies, and cards for plans, bills and call-backs. Each journey was assembled from them in advance. We knew every answer TOBi could give before a customer asked.

You can't design every answer any more. You design what answers are made of.

Now, the system is the design. A generative assistant writes its replies as it goes, so you can't design every answer. You design what answers can be built from. The card kit stops being a set of screens and becomes the boundary: a model that can only compose from approved components can't produce something off-brand or inaccessible. That's where design systems and AI meet.

The principle holds, and matters more. Every TOBi interaction had to show what's happening → what you can do → what happens next. With a generative model, that means three things TOBi rarely needed:

  • Show how sure it is. A confident wrong answer does more damage than a hesitant right one.
  • Let people check. Kavvi's Quick fill reads a pasted email, then asks you to confirm what it found (5 October). The same pattern belongs in any assistant that acts for you.
  • Know when to hand over. The best answer is sometimes a person, offered early, with the conversation passed along.

Personality, still honest. We wanted TOBi to be approachable but clearly a chatbot, and its 21 expressions were part of the system rather than decoration. That line matters even more when the AI sounds human.

If I rebuilt it tomorrow, I'd start with the kit, not the conversations: every card described well enough that a model knows when to use it, and what it must never be used for.

The original 2017 TOBi kitRead the Vodafone case study