Human in the loop Experiences
Copilot
Agentic AI
AI Governance
Privacy and AI
Enterprise Platform
Unifying and enhancing the in-house
agentic experience
Merging product silos and rethinking the agentic platform to create a cohesive and AI ready experience for One Trust
Project details
My role
Senior UX Designer
Duration
4 months
Team
Me, Lead Designer
Company
One Trust
01 - why we built this
Users don't think
in product boundaries.
Every copilot agent OneTrust shipped was scoped to a product team's domain, with no coherent relationship, shared language, or unified entry point. But users thought in tasks. When the agent they picked couldn't help, there was nowhere obvious to turn and each dead end quietly eroded trust in AI-assisted work across the entire platform.
Business impact
The cost of fragmentation
28%
AI add-on activation rate (below target)
5x
Support tickets from copilot confusion vs other features
73%
Of users couldn't name the right agent for a given task
User wants to execute a privacy task through Copilot
Old screenshots and user anecdotes from the moderated usability test
02 - User Research
9 sessions. One consistent story.
We ran moderated sessions with compliance officers, privacy analysts, and legal ops teams, OneTrust's primary user segments. Participants completed tasks across five copilot surfaces while thinking aloud. The board below captures synthesis across all sessions.
5
distinct copilots
audited
2.4×
longer to complete tasks vs.
no-AI baseline
4 in 5
hit a dead-end error with
no recovery
03 - Converging the Research findings
What was actually breaking
01
No unified entry and overlapping experiences
Copilot and Agents created overlapping AI capability experiences, along with with different entry points to invoke like sidebar, nav buttons, and inline prompts.
02
Unclear mental Model
Users weren't always sure when to use Copilot vs. an Agent, what each could do, or what would happen after invoking one
03
Trust/control issues:
Users needed more visibility and control, particularly before AI took actions.
04
Multiple dead ends
Generic failures or unsupported requests offered no recovery path. Users were often told what their request was without being helped to complete it, leaving them to abandon the task entirely.
04 - Analytics & Instrumentation
Designing for measurability
Analytics weren't a post-launch afterthought. Every interaction in the agentic flow was instrumented in the design spec, giving data science a clear event taxonomy before engineering began.
Agent entry
Source of copilot invocation tracked (entry point, trigger type)
Step completion
Each agentic step logged with duration and user intervention flag
Error + recovery
Error type, whether user retried, and final task outcome
Task outcome
Overall completion status, time-on-task, and whether the agent or user completed the final step
Role
Where I was responsible
I owned the full design arc, from research planning through to engineering handoff, while collaborating closely with product leads and a dedicated research team at each stage.
Research Planning
Structured and ran all research engagements defining goals per session, writing discussion guides and screeners, aligning the research team on what we needed to learn before each round.
Capability Taxonomy
Worked directly with product leads across five copilot areas to map what each agent could and couldn't do, creating a shared language that fed directly into the routing logic and onboarding design.
Error State Framework
Designed a systematic approach to failure across all agents categorising error types, defining recovery actions per type, and producing a reusable spec that each product team could implement consistently.
Unified Copilot IA
Defined the information architecture for the consolidated experience entry points, navigation patterns, and the mental model we wanted users to build across all agent surfaces.
Agentic Flow Design
Designed the transparency rail for multi-step autonomous actions how steps are surfaced, how users pause or redirect mid-flow, and how completed actions are logged and reviewed.
Concept Validation
Led two rounds of concept testing with enterprise users, synthesising findings into prioritised changes before hi-fi began.
Hi-Fi & Handoff
Produced final high-fidelity designs and the annotated component spec that engineering worked from including interaction details, state coverage, and responsive behaviour.
05-Quick Visual Glance · End State
The unified end state designs.
A single copilot surface - intelligently routed, transparently agentic, and recoverable at every failure point. These screens show the experience from first prompt to completed task.
OneTrust · Copilot
Redesigned the Copilot UI with a clearer AI visual language, using gradients and AI sparkle to create a distinct, recognizable experience. Streamlined the header to reduce clutter and reposition key actions, while surfacing suggestions across all available agents on the Copilot home for direct
discovery and launch.
Agent hand off through prompt. No more manual agent selection or recall of capabilities.
Engage through next steps by providingCTAs. No more abrupt end to the user flow or user intent.
Human in the loop . No more
irreversible actions without users visibility and a chance to intervene
Improved visualization cards and source links to improve readability and faster skimming.
Redesigned AI settings page and left navigation
Redesigned the AI settings page for enterprise users with granular controls to manage copilot, agents, recommendations and auto suggest/scan accross the OneTrust Platform
06 - Impact
Measured at the 60-day mark
Analytics were instrumented across all five legacy copilot surfaces before consolidation to establish baselines. Post-launch measurement ran for 60 days against the same user cohort, with the data science team independently validating all metrics.
Business impact
Thank you for reading!
If you'd like to see a SaaS experience, take a look at my Mindtickle case study where I worked on building a content sync.
rath.desh@gmail.com
9040144121