← Back to work

Case Study · Copley · 2024–2025

Designing AI onboarding for people who don't trust AI.

Copley is an AI-heavy marketing platform built for DTC brands. The challenge wasn't teaching users how to use the product — it was getting them to believe it was worth trusting in the first place.

Onboarding flow + FTUE agent

Role

Product Designer

Timeline

2025–2026

Team

Angela Lin · Mike Torra (CTO) · David Henriquez (CEO)

Tools

Figma

Copley's core product is powerful — it analyzes store data, surfaces high-impact creative opportunities, and uses an AI agent to take users from brief to production-ready ad variants in minutes. But we had a problem: the people who needed it most were the least likely to trust it.

DTC marketers are skeptical of AI tools by default. They've seen bad output, they worry about losing creative control, and they don't want to feel like they're being replaced. The existing onboarding put users through a form-heavy sign-up before showing them a single pixel of value — and dropped them into a blank dashboard with no guidance.

My job was to design an experience that built trust before it asked for it.

The problem

Three fears standing between users and value.

Before designing anything, I needed to understand what was actually holding AI-hesitant users back. Through user interviews and behavioral analysis, three distinct fears emerged — and they required completely different design responses.

FEAR 01

'AI will replace me'

Not a fear of the tool — a fear of becoming irrelevant. Marketers who've spent years developing creative instincts don't want a system that makes their judgment unnecessary.

FEAR 02

'I don't trust the output'

They've seen AI generate bad creative. They need to see the reasoning, not just the result — transparency about how a recommendation was reached matters as much as the recommendation itself.

FEAR 03

'This is too complex'

Copley has real depth — briefs, variants, traits, combinations. But complexity shown too early reads as risk. The wrong first impression and they're gone before they've seen what it can do.

Users don't need to be convinced AI is good. They need to feel in control while it's working.

— Design principle guiding the entire flow

The approach

Value before commitment. Show before tell.

The original onboarding asked users to commit — name, email, company info, integrations — before showing them a single thing Copley could do. For an AI-hesitant user, that's asking for trust they haven't earned yet.

I flipped the logic entirely. The new flow shows value first, earns trust through transparency, then asks for commitment. By the time a user reaches the sign-up form, they've already seen a real insight from their own data. Sign-up becomes confirmation, not a blind leap.

Original vs Redesigned Flow

Flow map — original 7-step form-first vs new 5-step value-first flow

Key design decisions

Three decisions that changed the outcome.

Decision 01 — The Insight Card

Instead of landing on an empty dashboard, users arrive at a ranked list of opportunities already found in their data. The first thing they see after connecting is proof that Copley has been working.

The copy pattern was intentional: name the pattern, show the specific number, explain the implication, offer an action. “Close-up product shots are your best-performing creative — and you only have 2” is not a generic recommendation. It's a finding, rooted in their account.

Insights screen — HIGH IMPACT + MEDIUM IMPACT cards with 'Build this Brief' CTA

WHAT CHANGED

The insight card is the landing page

Specific to their account, ranked by estimated revenue impact. This is the first thing they see after connecting their data — proof that Copley has been working since they signed up.

WHY IT MATTERS

No empty state. Ever.

An empty dashboard is the worst first impression for an AI-hesitant user. It puts the burden back on them. The insight card removes that entirely — the action is obvious, the value is visible.

Decision 02 — The FTUE Brief Agent

The existing brief agent was built for power users — it opened cold and asked a series of configuration questions before generating anything. For a first-timer, that's a blank-page problem disguised as a feature.

I designed a parallel FTUE agent with a completely different interaction model: show → confirm → generate. Copley opens pre-seeded with the user's specific insight. The agent proposes a creative direction — the user approves or reacts. Four exchanges, all chip-based. No typing required, no prior knowledge needed.

FTUE agent — seeded entry, chip-based exchanges, Copley proposes the brief

POWER AGENT (EXISTING)

Interview then generate

Opens cold. 8 configuration questions before a single ad is produced. Designed for power users who know what they want — not a first-timer's entry point.

FTUE AGENT (NEW)

Show then confirm

Opens seeded with the user's specific insight. Copley writes the creative direction and proposes it. The user approves or edits. 4 exchanges, chip-based, no typing required.

Decision 03 — The Dismissal Feedback Loop

AI-generated ads won't always be perfect. Pretending otherwise erodes trust faster than bad output does. The question was: how do you turn a bad ad into a trust moment?

When a user dismisses a variant, a small message appears in the agent panel: “Not quite right — I'll use that to refine the next one.” The replacement appears in the same slot. The first time it happens, the agent explains it. After that, it's silent. The system learns visibly, in real time, from the user's own judgment.

Variants screen — Suggested · Saved · Deleted feedback system

Creative direction

Teaching the agent to speak like a strategist.

One of the more unexpected parts of this project was realizing that the copy the agent produced mattered as much as the UI surrounding it. If the agent sounds like a chatbot, users disengage. If it sounds like a senior creative strategist who's looked at their account, they lean in.

I designed a trait-based insight system where Copley surfaces specific creative opportunities rooted in actual performance data. The copy pattern: name the pattern, show the number, explain the implication, offer an action.

Close-up product shots are your best performing creative — and you only have 2. Ads where the product fills more than 60% of the frame averaged a 2.4× higher click-through rate than your lifestyle and flat-lay creative.

— Copley insight card, High Impact recommendation

Language Audit

I ran a full language audit across the product. Words like “AI,” “automate,” and “machine learning” consistently triggered skepticism. Replacing them with outcome-oriented verbs — “surfaces,” “flags,” “highlights” — reduced perceived threat without changing what the product actually does.

BEFORE

Alienating language

'Our AI will automate your experiments' · 'Machine learning-powered insights' · 'Gross Merchandise Value' · 'Tell us about your company'

AFTER

Trust-building language

'Copley surfaces which tests to run next' · 'Finds patterns your team would miss' · 'Roughly how large is your store?' · 'Help us show you the right results'

Future states

Where this goes next.

Trait-based creative intelligence

Copley learns each brand's top-performing creative patterns over time and surfaces them proactively — not just at onboarding but throughout the brief agent flow.

Graduated trust system

As users build history with the product, the FTUE agent retires and the power agent takes over — surfacing only the configuration options each user has previously engaged with.

Dismissal feedback at scale

Aggregate dismissal patterns across accounts to improve generation quality for entire creative categories — model-level learning from real creative feedback.

Designing for AI-hesitant users forced me to think about trust as a design material — something you build deliberately through sequencing, language, and what you choose to show or hide at each moment. The instinct when working on an AI product is to explain the AI. I learned that the opposite works better: let users experience the output first, and trust follows from seeing it work.

The dismissal feedback loop is technically a simple regeneration trigger. But positioned as “your judgment is improving the system,” it transforms the same interaction from cleanup into collaboration. That reframe changed everything about how the feature feels — without changing what it does.

Skills & Tools
AI workflow designConversation designUX strategyCopy systemsOnboarding designFigma