Note: The initial design of the chat was given to us. The interactive prototype is what we created.
This case study is protected under an NDA. Confidential and proprietary details have been removed or anonymized to protect client information.
Using agentic AI to keep product information systems up to date for a B2B SaaS company.
the Problem
You build a journey map that perfectly captures how a customer experiences your product. After your product ships, your data shifts, and a new role joins the team. Six months later, you're presenting a map that no longer reflects reality, and the blame lands on the CX strategist who never had a clear signal that anything had drifted in the first place.
Journey maps are living documents, but the tooling around them treats them like static files. The Agentic AI maintenance agent is built to close that gap, proactively watching connected data and surfacing what needs your attention, with the fix already drafted.
the context
the solution
Design Decisions
where do users meet the agent?
Three entry points, designed so the agent meets users wherever they already are, not as a feature they have to remember to open.
Workspace
A badge on each journey map shows how many fresh suggestions are waiting, so review is the first thing you see when you open a workspace.
In-Map Intelligence Agent
A lightbulb in the journey toolbar opens the full suggestion list alongside the map, with the affected stages highlighted in place.
Design Decisions
Evidence is always present.
Every suggestion card shows the supporting quote and a link to the original document. The AI's reasoning is visible by default, not hidden behind a "why?" button.
The user always decides.
The agent suggests, the user decides. Approve or Deny covers the full decision space, and an Approved / Denied history makes every call reversible.
How I know it works
We ran moderated usability testing with six UX designers familiar with journey mapping (used as a substitute for live customers, to avoid setting unintended product expectations). Each session walked participants through three core tasks: finding the agent, acting on a suggestion with evidence, and initiating a walkthrough. This was followed by structured reflection questions.
average task difficulty across all three tasks
usability testers across two navigation variants
design iterations integrated from test feedback
The one insight that validated everything
"I trust the suggestion the moment I can see where it came from."
Synthesized from usability testing, n=6. The problem was never accuracy. It was transparency. Once the evidence quote and source link were accessible, the entire interaction shifted from "is this right?" to "yes, that fits."
Design Decisions
What does the agent look feel like?
The maintenance agent had to feel like a collaborator, not an autopilot. Our agent had to act as the careful, evidence-quoting teammate. We chose a warm accent that signals attention without urgency. The panel sits flush to the side of the map, never blocking the work.
The agent collaborates. It doesn't decide.
Every interaction is tuned to the same idea: the user is in charge, the agent is in service. Options for approve and deny instead of auto-apply and inline evidence enable this.
How does this shape my future work?
Go extremely deep into the problem.
Trust is a UX problem, not a model problem.
Combine concepts before you over-build them.
Match the platform's mental model before introducing your own.
Design for the noticer, not just the owner.
