Skip to content
CASE 01 / AI ASSISTANT · ENLABEL · 2026

One AI assistant for every product in a labeling ecosystem

  • 2 design rounds
  • AI Tools
  • Component System
01 Context

enLabel makes labeling software for heavily regulated industries. I designed its AI experience twice: first as a one-shot label generator on a marketing site, then as one chat component for three products.

e-Draw: data-driven label design, free and not yet officially launched. IPM: artwork and data management for labels. eIFU: document publication, distribution and compliance, mainly for medical devices.

Prompt builder selections composed as editable pills in the composer
02 Round 1 · A generator on the marketing site

In February–March 2026, I was assigned a one-shot project to add onto the e-Draw marketing site, as enLabel began exploring AI in its products. A visitor describes a label on the site, then opens the result in the editor. I worked through three placements in Figma, prototyped the front end with Claude and handed it to a developer to connect to our LLM.

I didn’t start from zero. Going in, I drew on background knowledge from my past AI work and some light research:

Prior testing

My SkipLegal capstone (4 interviews, 112 surveys, 15 testers) found that trust in AI depends on being able to verify it. Those users weren’t labeling professionals, so I carried this over as a hypothesis.

AI tools and patterns

My daily use of Claude, Claude Code and Lovable, plus NN/g’s chatbot guidelines and Shape of AI, grounded the prompt and generation patterns.

Human-AI studies

Published work on overreliance, prompting, clarifying questions, citations and chat accessibility.

Products and rules

Public docs from design tools and regulated products, plus FDA and EU AI Act requirements. No hands-on product testing.

Interactive demo. Generation is mocked with a timer, so every prompt returns the same sample label.

One prompt, three labels, then sign up

  1. 1 Prompt: describe a label, or start from a suggestion
  2. 2 Generate: a progress bar while three options fill in
  3. 3 Choose: pick the label to continue with
  4. 4 Open in e-Draw: the chosen label opens in e-Draw first, and users only sign up if they want to save or keep editing the AI's output
WHY I mapped two flows: generate first, or verify an email first. I kept revising generate-first, so visitors see a real result before they're asked for anything.

The order also served the business goal. The site needed visitors to try e-Draw and sign up, so I showed value before asking for anything: people see a real label first, and sign up only if they want to keep it.

Then it stopped. Priorities shifted during the build and the generator never launched, which means the sign-up order was never tested.

03 Round 2 · The brief gets bigger

In summer 2026, I returned to the project, and the scope grew: one reusable chat component that could be integrated into all three products. Each product asks something different of AI, so the part they share had to be the interaction shell, with behavior set per product.

PRODUCTAI CANAI CAN'T
e-Draw · first Generate and revise draft designs on a canvas Invent controlled values or symbols, or release a label
IPM · later Summarize changes and report findings at a review step Approve, sign or change a released record
eIFU · later Help employees at highly regulated companies build an on-brand, fully compliant and localized eIFU deployment through chat Replace published content, publish, or decide who receives it
04 Research · What should AI be allowed to do here?

The bigger brief needed a wider evidence base. enLabel had no approved AI requirements and no recorded e-Draw user research, so I built the evidence from what I could reach.

WHERE I LOOKEDWHAT I FOUNDWHAT IT MEANS FOR THE DESIGN
Design tools
Figma, Canva, v0
Chat captures intent, and the canvas is where people inspect and edit. Selection scopes the prompt. Prompt from the selected object, and keep manual tools beside chat
Regulation
21 CFR 820.45, AI Act Art. 50
Device manufacturers must examine labeling before release, and people must be told when they're interacting with AI. My reading: keep AI output as a marked draft until a person releases it
82% → 45.5%

Very experienced radiologists' accuracy with correct vs. incorrect AI suggestions. Sham AI, small subgroup (Dratsch 2023)

9.77%

of cited AI answers were actually checked, and citations raised trust even when random. Low-stakes study (Ding 2025)

17–33%

hallucination rate in commercial legal research AI, lower than GPT-4 but far from zero. Tools have since changed (Magesh 2025)

AI can be leveraged to explore, draft, and check for errors, but a qualified person must have final say in acceptance.
05 Exploration · From e-Draw wireframes to a generic shell

When I picked the project back up, my wireframes were built specifically for e-Draw. I focused on making the interactions very clear, especially scoping exactly where the AI would look and what it would touch.

Early wireframe with a bright AI Assistant button in the e-Draw header and a gray fixed panel over the canvas BEFORE

A header button and a fixed panel

My first placement study put a loud AI button in the app header, opening a fixed block with skill suggestions under the composer.

AI Assistant tab hanging from the ruler, with a chat window over the canvas beside a selected object AFTER

A tab and a movable window

The entry point moved to a quiet tab on the canvas. The window can move, resize and minimize, because people need the rest of e-Draw during a chat. Placeholder content.

Two ways in, each with a clear scope

  1. 1 General chat: the canvas tab opens the window for document-wide requests
  2. 2 Targeted chat: an AI button on the selection toolbar opens chat beside that object, scoping the request to what's selected
  3. 3 Context label: the window's top bar names the data the AI is working from
WHY Comparable tools scope prompts to the selection. In a study of simple edits, direct manipulation was 50% faster than typed prompting, which still suited global tasks (Masson 2024, n=12). So chat carries intent and the canvas stays in view.

Then the canvas came out. The brief then changed to a product-agnostic component, so I designed without the e-Draw canvas. That led me to a generic template: a standalone chat window with no canvas around it, which I filled with dummy data relating to e-Draw, like a shipping label for organic food following FDA 2026.

Context is visible

The top bar names the data source the AI is working from, so users know what it can see in each product.

Structure beats a blank box

Quick prompts and a prompt builder turn domain choices into a complete request.

Never a black box

Progress steps, a stop button and token cost show what the AI is doing and what it spends.
Prompt builder variant with a blue Clear Selection button in the builder and pills in the composer BEFORE

Variant A: press Begin

Dropdown choices took a Begin click to reach the composer. Clear Selection then sat up in the builder, away from the pills it removes.

Prompt builder variant with no button in the builder, and pills with a clear icon in the composer AFTER

Variant B: write straight through

Choices land in the composer as pills right away, and the clear icon sits next to them. The composer is the one place a prompt lives.

06 Key decisions
AI assistant window, empty state with quick prompts and prompt builder

A prompt builder that writes the hard part

  1. 1 Quick prompts: one-tap starts for common jobs (product label, packaging, IFU, compliance check)
  2. 2 Build Your Prompt: document type, category and standard dropdowns
  3. 3 Editable pills: choices land in the composer as a sentence ("Shipping Label for Organic Food following FDA 2026") the user can still change or extend
WHY Non-experts prompt opportunistically and overgeneralize from single results (Zamfirescu-Pereira 2023, n=10). In labeling, a missing standard or category can mean a wrong label, so structured inputs carry those details without taking away free text.

The AI asks instead of guessing

Problem → vague prompts produce confident but wrong output.
Decision → follow-up question cards with numbered answers, an "Other" field and paging (1 of X).
Outcome → missing details get resolved before generation, in a few clicks.

Generation you can see and stop

Problem → long AI tasks feel like a black box.
Decision → step chips (Read guidelines → Check compliance → Generate layout), an always-on Stop and token cost before and after.
Outcome → users know what's happening, what it costs and how to bail out.

The smaller states follow the same logic: the user can see what’s happening and stays in control.

STATEDESIGN RESPONSEBECAUSE…
Voice input Live waveform, confirm/cancel, transcript preview before send users approve what was heard
Mic permission blocked Warning icon with "review microphone permissions" tooltip clear fix, not a silent failure
Different needs per task Model picker in the composer trade speed for depth
Generation fails Plain-language cause, Retry / Edit Prompt, recovery tip recover without starting over
07 The flow

Start from structure

Quick prompts for common jobs, or the prompt builder for precise ones. The top bar shows which data source the AI is working from.

Compose, then refine

Builder choices become editable pills in the composer. Attachments and free text add detail, and the token cost shows before sending.

Answer what’s missing

If the request is ambiguous, the AI asks follow-up questions with numbered options, an "Other" field and keyboard support.

Watch it work

Step chips track progress from reading guidelines to generating the layout. The output streams in, and Stop is always available.

Fail gracefully

Errors say what went wrong in plain language, with Retry, Edit Prompt and a tip for what to try next.

Start from structure
Compose, then refine
Answer what’s missing
Watch it work
Fail gracefully
08 Component kit · Built for reuse

Every piece is product-agnostic. Each product swaps in its own data source, quick prompts and builder fields, and the shell stays the same. Components bind to usage tokens, so products can change signature color and IPM can add a compact density.

Top bar with context label
Top bar · Context
Status chips: done, in progress, queued
Status chip · 3 states
Prompt builder pill, default and remove-hover
Prompt pill · Default / Remove
Quick prompt pill
Quick prompt pill
Model dropdown options
Model picker
Composer states: empty, recording, mic error, transcript, typed, built prompt, model menu
Composer · 9 states
09 Validation & reflection

None of this has been tested with labeling users yet, and that’s the biggest open risk. Here’s the plan:

Discovery

Contextual interviews on how label creators and reviewers work today, and which steps AI should touch.

Usability rounds

Moderated prototype tests of scope, question cards, proposals and history, including side panel vs. window.

Accessibility

Screen reader task tests with NVDA, JAWS and VoiceOver, against WCAG 2.2 AA.
Designing a shell for many products meant deciding what stays fixed and what each product gets to change.

Round 1 taught me a demo isn't a product. Round 2 taught me the hardest part of an AI interface here is deciding what it must never do. Rollout starts with e-Draw, where output is a reviewable draft, then IPM and publisher-side eIFU. I'd keep chat out of the clinician-facing viewer for now, because a wrong answer at the point of care could directly affect care.

My plan moving forward is to build a design system, with components living in Storybook, so devs and I can easily collaborate. More and more of my projects end by delivering a reusable component that can be shared across products in the company, and this kit is one of them.