[Introduction]
For an IT support agent, resolving a ticket starts long before the reply. First comes reading the conversation, understanding what has already been tried, checking the affected asset, and deciding what to do next. Repeated across a busy queue, this work becomes a significant part of the job.
During my four-month industry capstone at Atomicwork, I explored how AI could make these moments easier. Agent Assist brings ticket summaries, contextual responses, and structured categorization into the agent’s existing workspace, with the agent retaining control over each decision.
My Role
As a UX design intern, I worked across research and benchmarking, concept exploration, high-fidelity Figma prototypes, accessibility reviews, and developer handoff. I collaborated with my mentor and product managers to refine the designs around agent workflows and product constraints.
[Description]
Understanding the Problem
The brief was to explore how AI-driven Agent Assist could improve IT service management. I approached it through the experience of the person resolving the request: what information do they need, where do they lose time, and which decisions should remain theirs?
1. Rebuilding Context
Ticket history was spread across comments, responses, asset details, and follow-ups. An agent taking over a request had to reconstruct the story before taking action. The opportunity was to make the important context easier to find.
2. Repeating the Same Work
Acknowledgements, progress updates, and requests for more information appeared repeatedly in the workflow. Writing them from scratch added effort and made the tone of responses less consistent across agents.
3. Making Classification Reliable
The existing two-level categorization structure lacked the dependent choices needed for more precise classification. Agents had to check combinations manually, creating friction for both ticket handling and reporting.
Observed workflow friction and the design response to each problem.
Research & Discovery
I studied enterprise AI tools, including AIsera, Moveworks, and Microsoft Copilot, then reviewed how agents worked inside Atomicwork. The external research helped me compare approaches to automation and contextual assistance; the internal workflow review showed where that assistance could be useful in everyday ticket handling.
Key Research Insights
Context needs to be available at the moment an agent opens a request.
Suggested responses are useful when agents can preview, edit, or ignore them.
Clear dependencies can make classification easier to understand.
Assistance should fit the current workspace and reduce context switching.
My Approach
I explored ideas ranging from automatic assignment and close suggestions to form auto-fill and knowledge recommendations. I then focused the design on three recurring moments: understanding a request, responding to it, and organizing it correctly.
This gave the project a connected experience rather than a collection of isolated features. Each intervention had a clear place in the workflow and a clear boundary between the system’s suggestion and the agent’s decision.
The three core moments addressed by Agent Assist.
Ideation & Early Concepts
I began with quick paper and FigJam sketches to explore placement and interaction. For summaries, the question was how to keep context visible. For macros, it was how to make a suggestion discoverable inside the reply editor. For categorization, it was how each selection should shape the next.
I moved into high-fidelity Figma concepts using realistic ticket content so the team could evaluate the ideas within Atomicwork’s existing interface. This made wording, hierarchy, and interaction behavior part of the conversation early in the process.
Final Design & Experience
AI Summaries: Start with the Context
The summary panel brings together the issue, impacted asset, relevant history, and recommended next steps. Its purpose is to give the agent a useful starting point before they work through the full request.
I placed this assistance within the request view and refined the design to keep the panel visible during longer reviews. The summary supports the agent’s understanding; the next action remains their choice.
Original summary-panel concept with annotations explaining the information hierarchy.
Contextual Macros: Make Repetition Easier
The reply editor is where repeated communication happens, so it became the home for contextual macros. Agents can find a response, preview its content, insert it, and adjust the wording before sending.
The design considers request type, workflow stage, and available ticket context. Early suggestions were too easy to overlook, so I strengthened their contrast and hover feedback and simplified the content for quicker scanning.
Original canned-response interface, presented with its preview-and-edit sequence.
Dependent Categories: Guide the Next Choice
I designed a Category → Type → Item structure to make classification more precise. Selecting a category narrows the available types; selecting a type narrows the available items. The interface guides the agent through valid choices instead of leaving them to check unrelated options.
This work extended beyond the dropdown’s appearance. The handoff needed to explain dependencies, alternate paths, and field behavior so the design could support consistent routing and reporting.
An illustrative classification path showing the dependency model; this diagram explains the logic rather than reproducing a product screen.
Testing & Iteration
I used informal peer walkthroughs and focused tasks to evaluate the prototypes: locate the summary, insert a response, and classify a request. Weekly reviews with my mentor and product managers helped connect the findings to feasibility and product priorities.
Summary visibility: keep useful context available while the agent scrolls.
Macro discoverability: use clearer contrast and hover feedback.
Faster scanning: simplify labels and response-card content.
Interaction clarity: add shortcut tooltips and refine spacing and focus order.
Screen recordings helped me identify scroll friction and confusing click paths. These were qualitative feedback cycles that shaped the design, rather than a measured post-launch performance study.
Designing Beyond the Core Features
Working inside the wider product revealed related opportunities. I proposed keyboard shortcuts for common ticket actions and audited Request Creation and Detail views for contrast, descriptive labels, tab order, and visible focus. The shortcut system’s full implementation was outside the internship scope.
I also designed a Power BI configuration experience with dataset-sharing controls and refined the Analytics landing page. The analytics work included space for exports and a future Ask Atom reporting concept, extending the same emphasis on clear choices and understandable information.
Outcome & Developer Handoff
The project concluded with three core feature designs prepared for developer handoff. Deliverables included Figma prototypes, component libraries, user flows, interaction annotations, validation and error states, and documentation of the categorization dependencies and AI suggestions.
The intended benefit was less repeated effort and clearer decisions within the agent’s workflow. The project established the design direction and interaction details; post-launch efficiency gains were not measured in this work.
Key Learnings & Reflections
Trust Is an Interaction Decision
Previewing a response, editing it, or choosing to ignore a suggestion are small interactions with a large role in trust. This project taught me to make the agent’s control visible in the experience.
Workflow Fit Matters
A useful feature needs to meet the agent where the work happens. Thinking about placement, timing, and context helped me evaluate ideas by their practical value to the workflow.
Design the System Behind the Screen
Categorization made the relationship between interface and system logic especially clear. A dropdown affects data quality, reporting, and future workflows. Documenting those relationships became as important as designing the visible states.
Conclusion
Agent Assist brought my focus back to the everyday work of IT support: understanding, responding, and resolving with confidence. It strengthened my ability to connect research observations to interface decisions and carry those decisions through to a clear, usable handoff.
