Designing trust into an AI platform

The AI worked. The business around it didn't.

Alana AI automated up to 98% of customer replies across social media, inboxes, and email. The technology was real, and the clients were names like Spotify, Diageo, and UNICEF. But three years in, retention was slipping and upsell was stalling, for reasons that were all about the experience, not the model.

Clients did not believe an AI could run customer conversations unsupervised, so they read every automated reply by hand, hunting for mistakes. Anything the AI could not handle got routed out to a third-party tool, so the product felt both expensive and fragmented. And the whole thing was priced and built for enterprise only. Ahead of a $5M round, investors wanted the opposite: an SMB-ready, self-service product. I led the design of it.

Four problems, one root cause: people could not see or steer the AI

• Distrust - Clients manually reviewed every automated reply, which erased the efficiency the AI was supposed to deliver.
• A third-party crutch - Tickets the AI could not solve were handled in an external tool, adding cost and breaking the experience in two.
• The wrong price shape - The pricing model sold isolated skills, which clients disliked and which locked out anyone smaller than enterprise.
• No front door - Onboarding happened in live sessions with our team, so acquiring a client was slow and expensive and SMBs were out of reach.

Trust comes from control. Show people what the AI did, let them judge and correct it, and they will let it do more.

Exposing the AI instead of hiding it. The automation was the product, and it was also the problem. The better the AI got, the more it unsettled the people relying on it, so they policed every reply by hand and the value evaporated. The tempting move was to make the AI feel more finished and hide its work. I bet the opposite: build a QA surface where users could easily follow what the AI was doing and grade its answers. For the client, that was the control they had been missing, the reason they no longer needed to check every reply themselves. For the company, every grade was structured feedback that fed straight back to the engineers retraining the models. The same surface answered both sides at once, and control and trust stopped competing and started reinforcing each other.

Refusing the familiar-versus-different tradeoff. The founders wanted a look and feel radically unlike Zendesk, so the product would never be mistaken for one more ticketing tool. Sales wanted the opposite: the more familiar it felt, the easier it was to sell. Rather than pick a side, I went looking for a third option. It took several rounds of sketching and dead ends before the Insight card emerged as the synthesis. Instead of a flat queue of tickets, the platform grouped messages into cards by type, then let teams filter them by relevance and urgency against their own goals, so anyone could see what to handle next. It gave the human team an intelligence layer even over the messages the AI could not answer, and it read as something genuinely new rather than a reskinned help desk. Familiar enough to sell, different enough to matter.

One principle, applied across the platform

• Designed the Insight card - The signature move. Messages grouped into cards by type, filterable by relevance and urgency against each team's goals, so the human team always knew what to handle next. It carried an intelligence layer over even the messages the AI could not answer, and it made the product read as new rather than as another help desk.
• Made the AI accountable - A QA surface where clients could follow what the AI was doing and grade its answers. For the client, the control they wanted, so they no longer had to check every reply. For the company, structured feedback to retrain the models.
• Killed the third-party dependency - A native ticketing flow, so clients stopped paying for and context-switching into an external tool for anything the AI could not handle.
• Opened the market - Self-service onboarding so an SMB could try the product with no sales call, plus a shift from selling individual skills to automation levels, so smaller clients could start low and grow.
• Let evidence pick the interface - I tested three mental models (email, feed, forum) with stakeholders and users, abandoned the familiar email frame because it could not surface AI insights, and rebuilt the flow to feel like a messaging app after testing showed that was what people expected. A second round of testing confirmed the redesign before it shipped.

I ran an unmoderated remote test with five participants, each working through six real tasks in a high-fidelity Figma prototype while I tracked how long each one took and whether they finished. The first round was blunt. People did not understand the Insight cards or what they pointed to, and they struggled to find and join conversations, telling me they expected something closer to the messaging apps they already used every day.

So I took those findings apart, turned each into a hypothesis, and designed against them: clearer card titles and tags, objective headers that surfaced the right information up front, and a conversation flow that behaved like the messaging apps people expected. Then I ran the same six tasks again with a fresh set of participants and measured the same things, so the before-and-after held up honestly.

Shipped, trusted, and measurably faster

The redesigned platform shipped as a self-service B2B SaaS product. In a second round of usability testing, participants completed every task, and did it in far less time, with report generation markedly faster. In year one the business saw double-digit revenue growth, a sharp drop in customer acquisition cost, and a solid lift in upsell, alongside lower operational cost and a validated path into SMB and international markets.

Three things carry over to any AI product I work on:

• Trust in AI is a design problem, not only a model problem. People adopt an AI they can see, judge, and correct, and they reject one they cannot.
• The human check-in that reassures users can double as the loop that improves the model. Designed well, oversight is an asset, not an apology.
• Repackaging the value, through automation levels and self-service, opened a whole market that better features alone never would have.

And three things I would do differently:

• Validate information architecture before visual refinement. I invested in high-fidelity directions before fully validating what users expected from queues, cards, and conversations. I would lock the structure first.
• Measure trust, not only automation. Automation rate turned out to be an incomplete measure of success. I would define measures for intervention, review behavior, perceived control, and confidence, and track those alongside it.
• Design the human handoff as a core experience. Human intervention deserved to be a primary workflow from the start, not something treated as overflow from the AI.

This is the work I keep coming back to, because the hardest part of an AI product is rarely the AI. It is giving people enough visibility and control to let it do its job.