Fragmented copy-paste across tools
Developers switch between PRDs, test plans, and trackers to move the same details.
Menu
Test planning is a hidden bottleneck in quality assurance: repetitive, manual, and disconnected from developer flow. I designed the Progressive Assistance Model (PAM), a contextual AI framework that helps developers generate, review, and refine test cases without losing control of the workflow. Modeled workflow analysis showed a 55–75% reduction in planning effort by moving from a 6-step manual process to a 4-step AI-assisted flow.
All visuals and product details are fully anonymized and recreated to respect confidentiality while accurately reflecting the challenges, design process, and impact of my work.
Developers spent meaningful planning time searching for context, re-entering duplicate information, and reconstructing test logic across disconnected tools.
The experience broke cognitive flow, introduced avoidable errors, and made it harder to trust the planning process.
“I copy test steps from past cases every time.” — Developer pain point
The manual workflow created three recurring issues:
From stakeholder conversations and workflow analysis, I identified where manual planning created the most friction. I translated each pain point into a design goal that kept AI assistance contextual, reviewable, and under developer control.
Fragmented copy-paste across tools
Developers switch between PRDs, test plans, and trackers to move the same details.
“Most planning time is spent reformatting and pasting from other docs.”
AI in the flow of work
Surface AI help directly in context. No extra tabs.
Review bottlenecks hide issues
One-by-one reviews slow releases and let problems slip through.
“Reviewing 100+ cases manually is unrealistic. We skim and miss things.”
Faster reviews
Batch actions and previews make large reviews quick and manageable.
Speed vs. oversight trade-off
Teams want speed but also need visibility and trust in results.
“Automation helps, but I need to see what changed and why.”
Keep developers in control
All suggestions stay editable, reviewable, and rejectable. AI proposes, humans decide.
I mapped potential AI entry points across the test creation workflow, evaluating each option through discoverability, interruption cost, and developer focus.
The goal was to make assistance easy to access without letting AI compete with the primary task.
Universal access for power users anywhere in the experience.
Expandable guidance and deeper exploration without blocking the workspace.
Context-aware help near acceptance criteria, aligned to intent.
In early prototypes, assistance appeared through modals and a global chat panel. What seemed flexible in theory proved distracting in practice, breaking user flow as attention fragmented across multiple panels.
I shifted to inline assist bubbles within the editor, keeping guidance visible only when relevant and preserving uninterrupted focus.
Test case generation took place inside the chat panel, allowing developers to edit and approve results while keeping the workspace in view.
Assistance builds trust when it aligns with attention, not when it competes with it.
I created the Progressive Assistance Model (PAM), a scalable UX framework for deciding when, where, and how AI should appear inside developer workflows. Instead of treating AI as a separate destination, PAM keeps assistance close to the user’s current intent.
After defining PAM, I refined the interaction model so AI felt like contextual assistance rather than a separate command surface.
Each refinement addressed the same core issues: fragmented workflows, slow reviews, and low trust in automation.
The goal was to make AI suggestions feel visible, editable, and reversible.
The inline icon appears only when a developer hovers over acceptance criteria. A quiet cue that says “I am here if you need me.”
Suggestions appear only after the developer asks for them. The user keeps control and the system does not interrupt.
Developers refine or regenerate test cases in the same panel. No tab switching. No lost focus. Just flow.
Multiple cases can be reviewed and accepted together. Repetition turns into confident, high-speed decisions.
Demo of progressive assistance for test case planning. Inline assist appears near acceptance criteria, the panel opens for generation, then a review pass with edit and add-to-suite. HUD chips show coverage and changes. Coverage animates to show improvement.
To estimate potential impact, I modeled the redesigned flow against the original six-step manual process.
These projections were estimated through modeled task durations informed by internal discussions and observed workflow patterns.
Even with conservative modeling, the redesigned flow suggested a meaningful reduction in repetitive setup time by shifting developers from manual drafting into focused review and oversight.
Beyond the modeled workflow analysis, I ran dry runs with developers and PMs to evaluate comprehension, perceived trust, and whether the assistance model felt intrusive or supportive.
Developers quickly understood inline assist cues without onboarding.
InsightClear discoverability validated the PAM “inline first” principle, showing that subtle inline cues were self-explanatory and trustworthy.
Stakeholders emphasized contextual triggers felt less intrusive than global prompts.
InsightReinforced that assistance works best when aligned with focus; AI should adapt to attention, not compete for it.
One participant noted the chat panel “felt helpful without interrupting flow”.
InsightConfirmed the perceived value of AI as a partner within flow; intelligence integrated, not imposed.
Designing contextual AI taught me that speed only matters when users understand what the system is doing and can stay in control.
By introducing assistance progressively, the experience built confidence before automation, turning repetitive planning work into reviewable, human-directed flow.
The Progressive Assistance Model (PAM) became a reusable pattern I later incorporated into other areas of the ecosystem: