AI-Powered Experiences

Pronto (Havas), 2022–2026

Lead Product Designer; working prototypes and comparative testing

Lifted AI feature adoption from 4% to 32% by designing a Review–Refine human-in-the-loop framework for Pronto’s first two production AI features.

Problem
Project managers wrote every task description from scratch, and 72% abandoned their first timeline and went back to a spreadsheet.
Approach
I built auto-commit and Review-Refine as working prototypes, tested them comparatively, and shipped Review-Refine with full opt-out to manual entry.
Result
Review-Refine became the reference pattern for every AI feature that followed, and Timeline Wizard adoption moved from 4% to 32% once the review step was restored.
Timeframe
4 months, for AI Task Helper and AI Timeline Wizard
Team
Product Manager, Project Manager, Business Analyst, Front-End Engineers, Back-End Engineers
The AI Timeline Wizard's Preview & Create step, showing a generated timeline brief to review before adding it
Final design: the AI Timeline Wizard's Preview & Create step, showing a generated timeline brief to review before adding it.

DiscoverStarting From Zero, Twice

Starting From Zero, Twice

Project managers wrote a full description for every task from scratch. On a forty-task project that is hours of writing before any work starts. Timelines were worse: a blank grid with no obvious entry point, and 72% abandoned their first timeline and went back to a spreadsheet. Both surfaced through support tickets routed via the Product Manager.

Pronto task panel with an empty description editor
Task Panel: Description written from scratch every time.
An empty project timeline grid with no records
Timeline Grid: No starting point, 72% abandoned on first attempt.

What They Did Instead

Indirect signals showed where the friction was, not why. I ran user interviews and week-long diary studies following how project managers actually set up tasks and timelines. AI-generated synthetic users were used only to shape early hypotheses ahead of those sessions, as a supplement to real users, never a substitute.

“Every task needs a full description with scope and acceptance criteria.”

“On a 40-task project that is hours of writing before work even starts.”

“My team spends more time writing tasks than doing them.”

“Is there a way to speed this up?”

Diary study extracts.

“We run similar campaigns for different clients and I retype near-identical task descriptions each time.”

“The wording drifts between PMs so nothing stays consistent.”

“Can the system draft these for us?”

Diary study extracts.

“Tried to start a new project timeline and gave up.”

“Blank grid, no idea where to begin with phases, dependencies or who to assign.”

“Went back to a spreadsheet in the end.”

“A starting point or template would make this usable.”

Interview extracts.

Accepting Isn't the Same as Typing

Someone filling a field knows what they typed. Someone accepting AI output has to trust the system got it right, every time, across hundreds of projects and timelines. That is a different risk profile from anything else in the product, and it had to be solved before either feature could save anyone time.

No Precedent
This was Pronto's first AI feature. No pattern existed anywhere in the product for how AI-generated content should behave.
Expectation of Error
Users wanted the time back but assumed the output would sometimes be wrong. Auto-commit meant errors compounding across a project.
Stakes Vary
A task description is easy to fix after the fact. A timeline structures dependencies across budgeting and resourcing, so the same mistake costs far more.

Tested Auto-Commit, Shipped Review-Refine

The PRD mandated auto-commit. I built both pathways as working prototypes and tested them comparatively. The AI workflow won on task-setup comprehension, but users would not accept generated content written into a live task unsighted. The test data changed the PRD: the review step was added in direct response to that finding, with full opt-out to manual entry at any point.

AI Helper prototype going straight from the prompt to applying the generated task
Pathway 1: Prompt-Generate-Apply.
AI Helper prototype with a Task Brief step followed by a Preview & Create review step
Pathway 2: Prompt-Generate-Review-Refine/Apply.

Updated Workflow

1Prompt
The user provides preliminary information via the prompt.
2Generate
The AI generates the output using the information provided.
3Review
User reviews the output and then decides whether to refine or apply.
4Refine/Apply
Output is refined with additional prompts, and then committed.

A Cue, Not a Redesign

There was no precedent for what AI looked like in the product, so the question was how users would recognise it without breaking the visual harmony of the existing UI. A red-purple gradient set at the same intensity as the primary CTA marked every AI entry point. Placement shifted with context so the entry sat where the work already happened.

AI Helper multi-option button open in a task description header
AI Task Helper: Multi-option button in the header section of the task description.
AI Timeline Wizard button in the timeline control panel next to the primary action
AI Timeline Wizard: Single-click button in the control panel, next to the primary CTA.

From Feature to Pattern Set

The reusable part was never the model, it was the interface around it. AI Task Helper runs two steps, AI Timeline Wizard runs three, and the review carries more visual weight where a wrong output structures a whole project. Those states went back into the library as defined components, so the next AI feature started from a pattern instead of inventing one.

AI Helper's Preview & Create step with a proposed task description, Refine text and Apply text actions
Review-Refine applied to the AI Task Helper.
AI Timeline Wizard generating a timeline, with Refine and Add to Timeline disabled
Review-Refine applied to the AI Timeline Wizard. Commit stays disabled through generation.

The Step That Got Cut

Went back to Develop

Under deadline pressure the Preview and Create step was cut for launch and the generated timeline committed straight into live projects. Adoption came in at 4%. An inaccurate timeline structures dependencies across budgeting, resourcing and every downstream decision. The review step went back in, was tested, and adoption moved to 32%.

AI Timeline Wizard at launch, going from Supporting Materials straight to Next with no preview step
Pre-Fix: Generated output committed directly, no Preview & Create step.
AI Timeline Wizard with the Preview & Create step restored
Post-Fix: Review-Refine step added back, adoption moved up to 32%.
“I should have pushed harder for user testing before launch, and I didn't. That's on me. Now, I ensure at least one early-adopter session before any AI feature ships.”

Review-Refine, in Its Final Form

Both AI Task Helper and AI Timeline Wizard shipped as fully styled, on-system components, not one-off AI UI. Here's the finished Review-Refine step for each, styled for production rather than prototyping.

Production AI Helper Preview & Create step over a live Pronto task
Production AI Timeline Wizard Preview & Create step over a live project timeline

Proven Twice, Once the Hard Way

Review-Refine became the reference pattern for every AI feature that followed. When it was skipped under speed pressure, the adoption numbers made the case for why it existed, and the fix validated it. The pattern held across two features with very different stakes.

  • AI Task Helper saved 5 minutes per task.
  • AI Timeline Wizard adoption moved from 4% to 32%.
  • AI Timeline Wizard time-to-create fell by 60%.

5 minutes saved measured via post-launch analytics. Adoption and time-to-create measured pre and post the Review-Refine release.

Learnings

Trust in AI features is not earned by accuracy alone. It is earned by making the user's ability to catch and fix a mistake visible before they commit.

A pattern working in one feature does not mean it transfers to a higher-stakes one without testing. Speed pressure after a successful launch is exactly when testing discipline slips and exactly when it matters most.

Future Planning

Making automation level a system decision rather than a per-feature one. Every AI touchpoint classified by what it costs the user to be wrong, and that classification mapped to a set pattern: direct apply where the output is reversible, Review-Refine where it structures downstream work, an explicit confirmation step before anything irreversible.

Want the full walkthrough?

Next case study

The Domains screen of a reseller-branded Mercury console, with DNS records for one domain

Multi-Brand Platform

VentraIP Australia, 2019–2022

UX/UI Designer; Figma Variables and Modes, usability testing

Built a theming layer that let each web agency reseller apply its own colour and typography across one live, WCAG-guarded console, which reached its first 100 customers in 19 days.

Read the case study: Multi-Brand Platform