Ashmi Roy
Back to all work End-to-End UX · AI Product · Confidential

Agentic Maintenance Platform

Redesigning journey map maintenance for a customer experience platform, with an agentic AI layer that keeps content fresh without taking control away from the people who own it.

Jump to solution

Client and proprietary details have been withheld under NDA. This case study describes process, decisions, and outcomes at a high level only.

RoleUX Researcher and Designer, Purdue UXD Experience Studio
TimelineJan 2026 to May 2026
Team9-person Purdue UX team
ToolsFigma, Usability Testing, Prototyping
Context

Journey maps go stale. Nobody notices.

Journey maps are living documents that need to evolve as customer experiences change, products update, and new data emerges. But on this platform, there was no system to alert users when their content became outdated, personas aged out of date, insights no longer matched current metrics, touchpoints drifted from reality.

Users relied on manual review and indirect signals to catch stale content, which led to inefficiency, disconnected data, and eroding trust in the maps themselves.

Problem statement

How might we help journey map owners identify and address outdated content, so their maps stay accurate, trustworthy, and actionable over time?

Who are the users

The client is a customer journey management platform that helps organizations understand their customers by building, analyzing, and maintaining journey maps. The platform already supported creation, analytics, and AI-assisted content generation, but maintaining those journeys over time was a gap.

CX professionalsCustomer experience strategists and journey owners
Product managersUsing journey maps to align product decisions
Service designersMapping end-to-end service experiences
Enterprise teamsOrganizations managing multiple journey maps

What we worked within

  • We could interview current platform users, but not test new concepts with them, so usability tests ran with 6 UX designers familiar with journey mapping.
  • Limited time and access to some competitor platforms narrowed the competitive analysis.
  • Client and proprietary details stay under NDA.
Research & discovery
01

Conducted desk research, competitive analysis, and a heuristic evaluation against Nielsen's 10 usability heuristics on a competing platform, grounding our direction in real gaps rather than assumptions.

02

Led 3 of 6 semi-structured interviews with real platform users, surfacing unmet needs, most notably that users had no way to know when their content had gone out of date.

03

Ran a Crazy 8s sketching session, moving from individual ideation to group whiteboarding, to pressure-test concepts before narrowing down what to bring to sponsors for selection.

04

Translated research into a high-fidelity interactive prototype across multiple rounds of iteration, from early concept sketches to a working agentic AI experience.

05

Helped synthesize findings across 6 usability tests, then presented design decisions and rationale directly to the client's CEO and VP of Product.

Process artifacts

click on an image to expand
Heuristic analysis matrix scoring a competing platform against Nielsen's 10 usability heuristics
Step 01, scoring a competing platform against Nielsen's 10 heuristics by severity, with a recommendation for each.
Whiteboard photos from a Crazy 8s sketching session
Step 03, individual sketches on the whiteboard before converging as a group on what to bring to sponsors.
Annotated slide from the final sponsor presentation with callouts on verification timestamps and evidence-based rationale
Step 05, walking sponsors through the annotated final design, verification timestamps and evidence-backed rationale as the two decisions we called out.
Design decisions

An agentic AI maintenance agent

The final design introduces two complementary AI capabilities. Together they reduce manual maintenance, improve visibility into outdated content, and keep users in control at every step.

Flowchart of the full agent suggestion flow, from entry points like workspace and journey map notifications, through the suggestion and intelligence agent interfaces, to reviewing evidence, walkthroughs, and approve, deny, or skip actions
The full flow

Users can enter the experience through several entry points, an agent notification on the workspace, an agent notification on a specific journey map, or the chat interface directly. From there they can review a suggestion in detail, explore the supporting evidence, ask AI for more context, or step through a guided walkthrough before deciding to approve, deny, or skip.

To reduce the risk of unintended decisions, approved and denied suggestions stay accessible afterward, so users can revisit their supporting context and undo a decision if needed.

Proactive · Agentic

Proactive monitoring agent

Always-on detection
Continuously monitors journey maps for signs of stale content before it compounds
Evidence-backed suggestions
Every flag comes with a clear rationale, not a black-box recommendation
User control
Approve, deny, or snooze, the agent assists but never overwrites
Reactive · Prompt AI

Reactive intelligence agent

User-initiated
Fully reactive, only activates when someone chooses to ask
Ask, don't dig
Get direct answers about the journey instead of hunting through data
Linked to context
Conversation stays grounded in the specific map element being discussed

Why the agent works this way

Each pattern traces back to our interviews or usability tests.

Is this still a current pain point?Interviews, a question stakeholders keep asking
Proactive stale-content alertsFlags on the workspace and map, with timestamps.
Users still wanted to verify AI output themselves.Interviews
Approve, deny, skip, or undoThe agent never edits on its own.
Testers checked the evidence behind every suggestion.Usability tests
Evidence on every cardOne click from the flag.
Testers mixed up the storytelling AI and the maintenance AI.Usability tests
One panel for the agentCombined with the chat assistant.
Asking about a suggestion took too many steps.Usability tests
A shorter flowFewer actions, plus a guided walkthrough.
Outcomes & evidence
6User interviews across 5 companies. I led 3 of them.
6Usability tests on the high-fidelity prototype.
3 / 10Average task difficulty reported by testers, where 10 is hardest.
4Design iterations made from testing insights before handoff.

What changed

  • Testers found the prototype simple to use, and valued being able to check the evidence behind each suggestion.
  • Design decisions and rationale were presented to the client's CEO and VP of Product, then refined with the sponsors' design guidelines for handoff.

AI & Technical Approach

Designed both the proactive monitoring agent and the reactive chat interface in Figma, then wrote the prompt structures and evidence-based suggestion logic that made the AI's reasoning visible instead of opaque. The goal throughout was an agent that assists without ever quietly taking control away from the user.

Figma Prompt Engineering Agentic AI Design
Reflection & next steps

This was my first time working on an agentic AI product, so there was a lot of context research and understanding I had to build before I could even start designing. Working across a 9-person team also pushed me to communicate design decisions more clearly, since not everyone was looking at the problem the same way I was.

I also learned that trust in AI is built in the details of the interface. Watching testers slow down on vaguer suggestion cards and lean on the evidence behind each one showed me that transparency has to be something people can see and check.

Recommended next

  • Revisit the walkthrough, since its overlap with manual suggestions still needs testing.
  • Measure real-world impact with the platform's own users a few months after launch.
  • Add onboarding that separates the maintenance agent from the platform's other AI features.
Back to all work Confidential · NDA Project