Designing Trust and Flow in AI-Assisted Test Planning (2025)

Designing Trust and Flow in AI-Assisted Test Planning (2025)

Client
Confidential (Developer Tools Domain)
Timeline
6 Weeks
Role
UX Designer (concepts, flows, prototypes; partnered with engineers and PMs)
Tools
Figma, FigJam

Test planning is the hidden bottleneck of quality assurance: repetitive, manual, and disconnected from developer flow. I designed the Progressive Assistance Model (PAM), a framework for contextual AI that amplifies developer intent instead of replacing it. The redesign cut planning time nearly in half and made trust in automation measurable.

6 manual steps 4 AI-assisted
| ~55–75% faster (modeled)

‍

Disclaimer 📌

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.

Challenge 🧩

Developers spent ~20% of planning time searching for context and re-entering duplicate test data.
The experience broke cognitive flow and eroded confidence in tooling.

“I copy test steps from past cases every time.” — Developer pain point

This fragmented process creates recurring issues:

  • Breaks focus through constant context switching
  • Increases errors from duplication and missing coverage
  • No reuse of similar tests across projects
See how the 6-step workflow plays out, with pain points at each stage.
  1. Hunt through PRDs to extract acceptance criteria.
    ⚠️ Context switching
  2. Infer scenarios from criteria and personal heuristics.
    ⚠️ Guesswork
  3. Manually draft titles and descriptions.
    ⚠️ Manual drafting
  4. Specify environment and parameters.
    ⚠️ Setup overhead
  5. Self-review for coverage and duplicates.
    ⚠️ Rework
  6. Share for peer review.
    ⚠️ Peer latency

‍

Pain Point → Research Insight → Design Goal 🎯

From stakeholder conversations and workflow analysis, key themes emerged. I aligned each pain point with supporting research insights and proposed design goals to turn them into actionable solutions.

‍

1
❗ Pain Point

Fragmented copy-paste across tools

Developers switch between PRDs, test plans, and trackers to move the same details.

🔎 Research Insight
“Most planning time is spent reformatting and pasting from other docs.”
Design Goal

AI in the flow of work

Surface AI help directly in context. No extra tabs.

2
❗ Pain Point

Review bottlenecks hide issues

One-by-one reviews slow releases and let problems slip through.

🔎 Research Insight
“Reviewing 100+ cases manually is unrealistic. We skim and miss things.”
Design Goal

Faster reviews

Batch actions and previews make large reviews quick and manageable.

3
❗ Pain Point

Speed vs. oversight trade-off

Teams want speed but also need visibility and trust in results.

🔎 Research Insight
“Automation helps, but I need to see what changed and why.”
Design Goal

Keep developers in control

All suggestions stay editable, reviewable, and rejectable. AI proposes, humans decide.

‍

Exploring AI Entry Points in the Workflow 🚪

I mapped potential moments where AI could assist during test creation through the lens of discoverability and focus.

The goal was to make assistance feel available yet invisible.

  • Global “Ask AI”

    Universal access for power users anywhere in the experience.

    • Always available
    • May compete with the primary task flow
  • Chat panel

    Expandable guidance and deeper exploration without blocking the workspace.

    • Persistent yet optional
    • Can be overlooked when collapsed
  • Inline icon

    Context-aware help near acceptance criteria, aligned to intent.

    • Quick and easy to find
    • Risk of visual noise if overused
Iterated concept sketch showing global entry, inline trigger near acceptance criteria, and a side chat panel.
Exploratory sketch of AI entry points in the test planning UI, comparing global, inline, and chat panel access.

‍

Iterations 🧠

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.

‍

‍

Designing the Progressive Assistance Model (PAM) 🔷

After realizing that AI visibility needed to align with user attention instead of competing with it, I created the Progressive Assistance Model (PAM), a scalable UX framework that defines when, where, and how AI appears within developer workflows.

‍

‍

Refining AI Behavior 🧭

After defining the Progressive Assistance Model (PAM), the next step was to make AI feel like a partner rather than a command line.

Each refinement focused on the same issues that sparked the project: fragmented workflows, slow reviews, and a lack of trust.

The goal was simple: turn intelligence into presence and create help that feels human.

Intent Detection

The inline icon appears only when a developer hovers over acceptance criteria. A quiet cue that says “I am here if you need me.”

Progressive Reveal

Suggestions appear only after the developer asks for them. The user keeps control and the system does not interrupt.

In-Context Editing

Developers refine or regenerate test cases in the same panel. No tab switching. No lost focus. Just flow.

Batch Selection

Multiple cases can be reviewed and accepted together. Repetition turns into confident, high-speed decisions.

Assistance that appears on demand, adapts to intent, and scales with volume.
Acceptance criteria
Test cases
Coverage 60% Changes 0
Idle
Create test for empty input
Applied ✓
Validate edge cases for length
Applied ✓
Handle network retry logic
Applied ✓
View Test Suite →

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.

‍

Projected Efficiency Impact (Modeled) 📈

To validate its value, I modeled the redesigned flow against the original six-step manual process.

Per Case projected
~55–75% faster
Manual
≈ 90–150 s
AI-assisted
≈ 35–45 s
20-Case Batch projected
~75–85% faster
Manual
≈ 30–50 min
AI-assisted
≈ 5–7 min

These projections were estimated through modeled task durations informed by internal discussions and observed workflow patterns.

Projection Basis
  • Review ≈ 5–8 s per case
  • 20–30% require 15–20 s edits
  • AI generation once per batch (≈ 2–6 s after trigger)
  • Batch accept and add to Test Plan/Suite ≈ 6–12 s

‍

User Testing 🧪

Beyond modeled efficiency gains, I ran dry runs with developers and PMs to validate comprehension and trust.

Observation

Developers quickly understood inline assist cues without onboarding.

Insight

Clear discoverability validated the PAM “inline first” principle, showing that subtle inline cues were self-explanatory and trustworthy.

Observation

Stakeholders emphasized contextual triggers felt less intrusive than global prompts.

Insight

Reinforced that assistance works best when aligned with focus; AI should adapt to attention, not compete for it.

Observation

One participant noted the chat panel “felt helpful without interrupting flow”.

Insight

Confirmed the perceived value of AI as a partner within flow; intelligence integrated, not imposed.

“Feels like gentle help exactly when I need it.”— Participant, design review
“Helpful without interrupting my flow.”— Participant, internal session

‍

Reflection & Next Steps 🚀

Designing contextual AI taught me that speed only matters when trust comes first.

By introducing assistance progressively, we built confidence before automation and transformed friction into flow.

PAM Integration Across the Developer Tools Ecosystem

The Progressive Assistance Model (PAM) became a reusable pattern I later incorporated into other areas of the ecosystem:

Wiki search assistance
Bug triage recommendations
Pull request descriptions
Engineering tools dashboards

‍

Let's make something cool together. 🚀

Share a few details, and I’ll be in touch.
Thanks for dropping a line! I'll be in touch shortly.
Oops! Something went wrong while submitting the form.
Back to Top ↑

© 2025 Kendra De'Anne