Letting AI draft the profile, and the user decide

At a glance

Problem: A complete profile powers everything the platform planned next, but filling it in by hand was long and tedious.
What I did: Designed an AI résumé import where the AI proposes and the user decides, item by item, without ever overwriting data without consent.
Outcome: Approved at every stage and now in development. As the product's first AI feature, it set the patterns future AI features build on.

Context & challenge

A global HR services company was renewing its long-running outplacement platform with AI, starting with My Profile: a hub that would later power job search, résumé generation and reskilling suggestions. All of it depends on how complete a profile is. Importing a résumé with AI was the obvious shortcut. Obvious, not simple:

Sales wanted visible AI, but an AI import has a "so what?" problem: the "aha!" only comes later.
No proofs of concept: we'd design around AI behavior we couldn't test in advance.
A brand-new team: client and studio people working together for the first time, alongside separate dev and AI teams.
AI risks: misread or invented data, cost per request, slow responses.

How might we use AI to fill in a complex profile quickly, without ever making users feel they've lost control of their own story?

My role

External partner from a US design studio and the only UX designer, working inside the client's product team. Over seven months I designed every section of My Profile, plus the AI import that fills several of them at once.

Vision & discovery: collaborated on product strategy and led discovery.
End-to-end design: feature maps, research prototypes, final designs, user stories, handoff and dev support.
Focus: only the people each decision needed, and the simplest viable design first.
Alignment in three stages: concept → full flow → final designs.

Discovery

Not in onboarding. Onboarding was already long, so the AI import became a deliberate action inside My Profile.
Merging is harder than importing. Most profiles aren't empty. Before any screen, I led the team to one shared map of the pipeline, from upload to suggestion, including every conflict with existing data.
A bigger system. Users assumed their coaches could see their profile; they couldn't. Tracing why led to a cross-product AI strategy proposal.

Import flow diagram in three lanes (participant and UX, logic, AI) across six steps, from file upload to suggestion, with example tables comparing imported and existing data. — One shared map, built with the client's PO and product designer, our AI-focused PM, and the dev and AI tech leads. Every technical failure has a UX answer, and every conflict reaches the user as a choice. Click to explore at full size.

Key decisions

1. Nothing enters the profile without approval

Why: the profile is personal, and AI parsing errors were a known risk.

My Skills section marked review pending, with AI-suggested skills that can each be accepted or discarded, and the Accept suggestion tooltip. — Users accept or discard each suggestion, at their own pace. A "review pending" tag shows what's left.

2. Field-by-field choice when data conflicts

Why: conflicts are rarely all-or-nothing, so users keep what's right on both sides.

Work History section marked review pending: one AI suggestion tagged Suggested update with a Review button, and another tagged New entry with accept and discard actions. — A suggestion that would change an existing entry is flagged as a suggested update. Reviewing it opens a field-by-field choice; the entry only updates once every field has a decision. Try the review flow in the prototype below.

▶Interactive prototype · opens in a new tabReview an AI-suggested update, field by field

3. Design for how AI actually behaves

Why: processing time is unpredictable, and every failure needs a way forward.

Three AI import progress steps side by side: Checking your education history, Mapping skills, Generating suggestions, each with a spinner and a Cancel button. — Once the file is validated, each step says what the AI is doing, instead of one progress bar that could lie.

1My draft, with product
2The devs' own list
3Merged, each with a way forward

Try another file8

Size limit exceeded
Format not supported
Password-protected
Empty file
Corrupted or unreadable
No extractable information
Language not supported
No suggestions

Try again2

Processing timeout
Generic failure

Continue, with a warning1

Low confidence

Prevented upfront1

Import limit reached. The button is disabled, with a tooltip explaining why.

I drafted the failure scenarios with the product team, then the devs answered with their own list. Together we merged overlaps and decided how each one is handled, and how the user recovers.

▶Interactive prototype · opens in a new tabAI résumé import, from upload to suggestions

How it was built

Two-week sprints, one section at a time. Manual entry first, then the AI-powered import for each section, reusing the same story pattern and evolving it.

#1Goals
#2Preferences
#3Work History
#4Education History
#5AI import · Education
#6AI import · Work, + conflicts
#7AI import · the file itself
#8Skills, manual + AI import
#9AI-suggested occupations

The pattern, applied: AI-suggested occupations

The next AI feature in My Profile was built on the same foundation: progress states, "review pending" and accept or discard, item by item. What's new is that the AI learns from the user's choices:

Accepted, discarded and manually added occupations all shape the next round of suggestions.
Discarded occupations stay in a list the user can restore, so the AI doesn't suggest them again.

Occupations to consider: three AI-suggested occupations pending review with accept and discard, three accepted occupations, and a Discarded occupations list with restore and delete. — Same review pattern as the AI import, now feeding the AI's next suggestions.

▶Interactive prototype · opens in a new tabGet AI-suggested occupations

Outcome

Approved at every stage, and now in development.
A blueprint for AI across the company: the AI proposes and the user decides, waiting is honest, and conflicts are resolved field by field.
A foundation for the product: complete profiles feed everything planned next.

Reflection

Next time, I'd push harder for an early technical spike with the AI team. It would have de-risked data quality and response times sooner.

Next case · 02Making a complex energy platform workable →