Case Studies
01 · Enterprise

Decidr — Decision Intelligence Platform

Turning fragmented commercial data into confident, agentic-AI decisions — one conversation at a time.

Role
Lead UX Designer
Team
In-house UX lead directing an agency pod of 3–4, + product, agentic & data eng, change & adoption
Duration
~12 months
Platform
Web app
Tools
Figma · Figma Make · Claude + VS Code · Miro · Jira
Decidr — whitelabeled decision-intelligence dashboard, chat output and alerts

Overview

An agentic AI decision-intelligence platform built for a global enterprise's commercial teams — brand leads, country presidents, therapy-area leads and analytics functions. Instead of switching between dashboards, users ask questions in natural language and get unified, explainable insights across performance, forecasting, ROI, targeting, segmentation and competitive intelligence.

It mattered because leadership had committed to a step-change in how commercial decisions get made — moving from lagging, manual analysis to predictive, prescriptive, simulation-driven agility. My job was to make an ambitious agentic system feel trustworthy and usable enough that senior business users would actually act on its outputs.

Product name and screens are whitelabeled for confidentiality.

The problem

For the business: decisions were slowed by fragmented tools, inconsistent KPI definitions across HQ, regions and field, and heavy reliance on PowerPoint and Excel to stitch a story together. Lagging indicators meant early signals were missed and opportunities left on the table.

For the user: discovery research surfaced one consistent pain — bouncing between multiple reports with delayed, manual insight-to-action. Five interviews across three personas pointed to the same friction: disconnected tools, invisible subnational data, one-size-fits-all views, and no trust signals on AI outputs. I structured the design around three productised personas (Brand, Country, Above-Country), segmented by scope of decision-making responsibility rather than job title.

“Too much data, not enough insight. I want to know the ‘why,’ not just the ‘what.’”— Discovery interview, commercial lead

My role & constraints

What I owned — end-to-end UX from 0 through MVP and into UAT (ongoing): the dashboard, chat/agent output, prompt enhancer, view-reasoning, sources, confidence score, deep dive, onboarding questionnaire, alerts management, workspaces and settings. I facilitated usability testing sprint by sprint, directed the agency pod's daily output, signed off screens before dev hand-off, and ran POC exploration for high-risk areas.

How I worked across functions — I co-led a two-day design-thinking workshop with business stakeholders, ran weekly UX planning and Figma handoff with engineering to validate agentic behaviours, and ran usability testing across sprints with brand-country, analytics, omnichannel and field-effectiveness personas.

Constraints worth naming
Agentic transparency & trust — every AI output needed traceable sources, view-reasoning and a confidence score. With no fallback confirmation step and no ability to act on external systems, every output had to earn trust entirely on the strength of its own explainability. No black-box outputs.
Evolving requirements — key business questions and scope were still being defined in parallel with design, so early design work had to move on informed direction rather than a fixed spec.
Data fragmentation across markets — inconsistent KPI definitions and data granularity across HQ, regions and field had to be reconciled into one coherent, trustworthy view.
Directing an external agency pod — maintaining consistency and quality across a 3–4 person agency design pod's daily output, rather than controlling every screen directly.
Access control & compliance — role/brand/country-scoped data (RLS/RBAC) plus a digital readiness review for compliant deployment.
Accessibility & multilingual — designed for a global user base across 16+ markets.
Design-system alignment — components were designed in alignment with the existing design system to ensure brand consistency; where new patterns didn't yet exist, I designed them from scratch and added them to the system.

Process

Research through handoff — with AI used to put testable things in front of people early, not as a synthesis shortcut.

01
Research · wks 0–6

5 semi-structured discovery interviews across the three personas; transcripts synthesised with Claude and internal LLM tooling to cluster pain points, sentiment and unmet needs.

✦ Claude + Copilot accelerated interview synthesis and sentiment clustering
02
Synthesis & ideation · wks 6–10

Two-day design-thinking workshop: current-state journeys, How-Might-We framing, use-case templates, future-state journeys, MVP prioritisation and brand identity. Workshop outputs were also synthesised with Claude to sharpen themes before moving into design.

03
Early validation POC · post-workshop

Built a functional front-end POC in VS Code based on workshop insights — before any official design screens existed — to validate direction with the client, sense-check the concept, and drive requirements-gathering and user-story conversations.

✦ Rapid functional POC (VS Code) used to shape requirements before design started
04
Design · in sprints

Design executed by the agency pod, with me overseeing direction and quality. In parallel, I ran feasibility conversations with engineering and supported the agency with rapid concepts (VS Code / Figma Make) to explore direction and pressure-test requirements ahead of their Figma work.

05
Validation · ongoing through design

Regular interviews and usability testing with end users throughout the design sprints (not just periodic checkpoints), with findings fed back into the agency's work each sprint.

06
Handoff & build review

Dedicated design-dev handoff in Figma, sprint reviews with business and dev for approval, then dev pickup — followed by reviewing dev implementation against designs to confirm visual and interaction fidelity.

Key decisions & trade-offs

01 — A rapid AI-built POC to de-risk workspaces — and being willing to defer it

Workspaces was the most complex feature on the roadmap and static mocks weren't surfacing the hard questions. I built a functional front-end POC in Figma Make, referencing the existing Figma designs so it would be directly usable if we moved forward, so business could feel the switch pattern and catch-up feed while dev pressure-tested feasibility against the same artefact. The result: workspaces was deferred on feasibility grounds — reached in days rather than after producing full screens. But it wasn't wasted effort: the POC helped define the requirements we'd need when we circled back to workspaces later, so it went into the backlog scoped and ready rather than shelved with nothing to show for it. The trade-off was a few days of exploration up front; the win was avoiding sunk design effort on something that couldn't ship yet, while still landing on solid requirements for next time.

02 — Having AI suggest alert settings, not just let users configure them manually

Manually setting up alerts meant users working through a large number of inputs per KPI card — a heavy lift just to get baseline monitoring in place. I proposed having the agent suggest alert settings automatically, based on what it already knew about the user and the conversation context, so a sensible alert was set by default with no manual configuration required. Users could still review and adjust anything they didn't agree with, but the starting point was AI-suggested rather than user-built from scratch. The trade-off was more design work to define good defaults and make the suggested/adjustable states clear, but it removed a genuinely heavy manual task and kept the interaction consistent with the platform's core pattern: agent proposes, user confirms.

03 — Treating trust as a designed feature, not a disclaimer

The hardest UX problem wasn't showing insights — it was making senior users comfortable acting on AI outputs. I designed trust into the output itself: every answer carried traceable sources with links back to the underlying data, an expandable "view reasoning" trail, and a confidence score. The trade-off was visual and interaction complexity on every output — but usability testing confirmed these signals were the specific reason business users were willing to move from looking at the intelligence to acting on it.

Outcome & impact

6 90+
brand-market combinations, from pilot to target scale within 5 months
16+ markets
targeted for rollout, designed to support multiple languages
On time
MVP shipped to UAT on schedule, with all core features integrated into one experience
Success criteria co-defined with UX — targets include ≥70% decision-support satisfaction, ≥60% weekly active use, and ≥15% reduction in analysis time.
Stakeholders explicitly credited the visibility and trust features — view-reasoning, sources, confidence score — as the reason business users were willing to act on AI outputs.
Programme footprint spanned 100+ cross-functional working sessions, engaging stakeholders across product, engineering and business throughout the MVP-to-UAT journey.
The KPI card, alert and workspaces patterns fed back into the shared library and are being reused across other agentic platforms in the portfolio.
Reflection
“If I did it again, I'd push harder for aligned problem framing before design work started — some rework came from ambiguous scope handed to us before the key business questions were settled. I'd also lean on rapid AI-built POCs earlier and more deliberately: the workspaces POC proved that a few days of functional prototyping could resolve a debate static mocks couldn't, and I'd want that as a first move on ambiguous features, not a fallback. The biggest lesson from this project: for agentic AI products, trust is a UX feature, not a technical afterthought — sources, reasoning and confidence weren't nice-to-haves, they were what actually let senior users move from watching the intelligence to acting on it.”
All case studies
Next case study
Digital Companion App