Building Trust in AI Design

ditto is an AI-powered marketing assistant for small businesses and agencies. I co-founded it in 2023 and was its only designer until I left in 2025. My job was shaping how users experienced the product's core promise: let the AI handle your marketing content, and trust that it'll sound like you.

That second part, trust, turned out to be the central design problem, and it ran through everything I designed: the first thing a new user saw, the onboarding, the editing controls, even the pricing.

The Challenge

AI-generated content is only useful if people actually use it. For marketing professionals, that bar is relatively low. They treat AI output as a draft and edit freely. For small business owners, it's a different proposition entirely. Their brand voice is personal. A post that sounds slightly off feels like a misrepresentation of who they are.

ditto also did most of its work automatically, which is a strange thing to onboard someone into. With so little for the user to actually do, the risk was that they'd feel like spectators, never sure whether the AI understood their brand. The experience had to prove that it did, without turning into a lecture.

Research & Discovery

I interviewed five participants: three marketing professionals and two small business owners. That's a small base, and I treated it accordingly, but the split between the groups was consistent and it held up against everything we saw in testing afterwards.

Marketing professionals were comfortable with AI-generated drafts as a starting point. They evaluated output critically and edited freely.

Small business owners were emotionally invested in their brand voice. They were sceptical that AI could capture their nuance, and needed reassurance before they'd engage.

The small business owners weren't opposed to AI, they were protective of their brand. They wanted to see ditto demonstrate competence before handing anything over to it. That finding shaped the entire design direction: show value first, ask for effort second.

Design Approach

Brand analysis as the trust foundation

An early product decision was having ditto analyse each user's existing marketing materials to learn their brand voice. I made that analysis a visible, deliberate moment rather than a background process: ditto was learning about you, specifically, before producing anything.

Our first usability test was the first signal this was the right call. When our beta user saw ditto's first output, content that actually sounded like her business, she hadn't expected it to be that accurate. One user proves nothing on its own, but the pattern repeated with every business we onboarded afterwards: the moment of recognition, "that sounds like me", was where scepticism broke. The rest of the design was built around earning that moment as early as possible.

Onboarding that shows before it asks

The first version of onboarding asked users to connect accounts and fill in details upfront, and research kept returning the same verdict: small business owners, especially solo ones, have no patience for setup. Every form field that arrived before a result read as a cost. So I rebuilt the flow around one goal, getting the user to their first real campaign, and postponed everything else.

Three decisions carried the redesign:

Background AI processing. While users moved through the steps, ditto analysed their materials behind the scenes, so the tedious part of setup simply never appeared.

The tutorial is the product. Rather than explaining ditto, onboarding walked users through creating a real campaign. The tutorial and the first result were the same thing.

Minimal, contextual UI. Each step showed only what that step needed, so new users were never presented with a wall of options.

Testing pruned it further. We assumed users would want to see everything the AI had generated upfront; usability sessions said a few good, concrete examples beat a wall of content. I cut the volume, led with quality samples, and the design got simpler as a result.

Control without friction

Users impressed by the output still needed to make small, practical edits: specific products, current promotions, details the AI couldn't know. I designed a context preview panel with three pathways: edit directly, regenerate entirely, or publish as-is.

The regenerate button came straight from observed behaviour. Some users found editing to be more effort than starting fresh, and a single click gave them a faster route to something they were happy with.

Pricing that doesn't spend the trust

Subscription tiers were as much a UX problem as a business one. AI processing costs varied a lot by model, so I matched each task to the cheapest model that did it well, then defined tiers around content quantity and type. The free tier gave one high-quality pillar piece per week plus three social posts. Users nearing their limits saw a quiet prompt about the next tier, no countdown banners, nothing that would spend the trust the onboarding had just earned.

For the same reason, ditto used predefined flows instead of an open-ended chatbot. Less flashy, far more dependable, and the output quality stayed consistent.

Designing with the engineer in the room

My co-founder was also our only developer, and every design I made was negotiated against two constraints he owned: what was feasible to build, and what it would cost to run. Those negotiations were frequent and sometimes blunt. Designs I believed in died because the AI processing behind them was too expensive, or because the build was more complex than a two-person company could carry.

It was the most useful discipline of the whole project. I learned to bring cost and feasibility into the design early instead of defending finished mockups, and some of the product's best decisions, the model-to-task matching and the predefined flows among them, came directly out of arguments I initially didn't want to have.

Results

I'm careful about what I claim here, because our evidence base was small and I'd rather it be honest than impressive.

When I left in 2025, ditto had 8 businesses actively using it. Four companies were running their content through ditto, with weeks of content generated and published over a period of months, which is the behaviour the whole trust arc was designed to produce. Our first beta user, the one from that first usability test, was still publishing ditto's output directly to her social media a year on.

Ahead of its seed round, ditto received an independent $10M valuation. The product experience was a part of that story, though I wouldn't claim to know how large a part.

Key Learnings

Trust in AI is built through demonstration, not explanation. No amount of reassuring copy would have done what that first accurate output did, and every step that delayed it, account connections, configuration, explanatory copy, was a moment of doubt the user had to push through.

Emotional and practical trust operate differently. Marketing professionals jumped straight to editing; small business owners needed to see ditto get their brand right before they'd touch the output. Leading with the brand analysis result let ditto prove itself before asking them for anything.

And constraints are collaborators. Designing alongside the person who had to build and pay for every decision made the product simpler, cheaper to run, and easier to trust.