Tale Case Study

Enterprise AI | 2024 - Present | Product Designer

Overview

Tale is a privacy-first AI platform deployed on a company's own infrastructure. It combines four things into one system: an AI chat interface, workflow automation, multi-channel conversation management, and human-in-the-loop approvals — all trained on the company's own data and history.

Most AI tools ask enterprises to hand sensitive information to third parties as the price of admission. Tale's bet is that they shouldn't have to.

My role sat between product and design: translating layered AI capability into interactions that felt intuitive — ideally, unremarkable.

Tale folds four different surfaces into one product: AI chat, workflow automation, a multi-channel customer inbox, and human-in-the-loop approvals. Each is non-trivial on its own. Stitched together, they could easily become a system that overwhelms — too many entry points, too many ways to do the same thing, too many places where the AI's role and reliability are unclear.

The users we were designing for were enterprise operators, many of whom started skeptical. They'd seen AI tools overpromise, behave unpredictably in production, or demand they hand sensitive data to third parties.

So the design problem cut two ways: make a layered, multi-surface system feel coherent and intentional, and earn the trust of users who started out wary of the technology underneath it.

Problem

So the design problem cut two ways: make a layered, multi-surface system feel coherent and intentional, and earn the trust of users who started out wary of the technology underneath it.

Constraints

The thing I learned quickly is that enterprise users don't want AI. They want their work done. And many of them are actively skeptical of AI — what we started calling AI aversion.

That created a tension we kept returning to: how do you build an AI product for people who don't trust AI?

The answer became the design philosophy underneath everything else: don't sell the AI, sell the augmentation. Build the product so that AI is unremarkable — a tool that quietly makes someone better at their job rather than a flashy assistant that demands attention.

That principle changed real decisions: which features got AI, which didn't, how confident the AI was allowed to sound, and when it was allowed to act on its own.

System view

Almost every primary action in Tale has two routes. Building a workflow? Chat with the AI, or start from a template. Replying to a customer? Use the AI's suggested draft, or write your own. Saving a prompt? Keep it personal, or share it with the team.

The two-path pattern isn't just user choice. It's trust-building. Users who don't trust the AI yet can use the manual path and watch how the AI's suggestion would have differed. Over time, many of them migrate to the AI path on their own — but only once they've decided to.

Key decisions

We aligned repeatedly on what the architecture made possible, so we'd sit down regularly to restructure pieces of it together. The simplification had to be real, not just visual.

Making the chat-to-automation handoff visible. The first version of the workflow builder was manual-first — users built workflows by hand, then layered AI on top to enhance them. The feedback was consistent: too many steps when users could just describe what they wanted once and make corrections as they went. The mental model was backwards.

The reframe surfaced the real problem: when a user is talking to AI about what they want automated, when does the conversation end and the automation begin? Letting the AI infer would have meant jumping the gun or sitting idle. So the handoff became an explicit, designed moment — the workflow materializes on the right side of the screen as the user talks, visible and inspectable, and they confirm when it's ready. The transition became a deliberate act, not a guess.

Scoping saved prompts at two levels. Personal-only would have meant everyone redoing the same work. Team-only would have meant authorship and governance questions on every save. We shipped both, with personal as the default and team as an intentional promotion — the difference between helping one person and compounding across a team.

Borrowing Gmail's mental model for Conversations. The Conversations module unifies channels — WhatsApp, email, and others — into one inbox with AI-drafted replies from the knowledge base. The temptation was to invent a new mental model. We did the opposite: same layout, same patterns as Gmail, with AI suggestions sitting inside that familiar shell rather than replacing it. The AI never sends autonomously — it drafts, and a human approves before anything leaves.

UI follows from structure; the reverse never quite works.

Impact

A few things landed for me through this work. The hardest problems weren't technical. They were about translating layered systems into interactions that felt simple without lying about the complexity underneath. Customer feedback wasn't a phase at the end of the project. It was a constant pressure on the information architecture. When users told us 'this is too complex,' they were usually right — but the fix wasn't always to remove things. It was often to sequence them differently, or to give the user a quieter path through the same destination. And restraint mattered as much as addition. Just because we were building an AI product didn't mean every surface needed AI — plastering it on features that didn't need it would have undermined trust in the features that did. The strongest moments in the product are the ones where users don't notice the AI is there. They just notice their work is easier.