← All work

AI + Frontline experience + Product strategy

Designing AIfor frontline work

How I moved beyond a chatbot to create a more contextual, accessible and actionable AI experience for employees working in fast-moving operational environments.

Interface recreated for portfolio purposes. No confidential product screens shown.

Role

Senior Product Designer

Focus

Product strategy, interaction design, research, accessibility and AI experience design

Platform

Mobile + Responsive Web

Ownership

Problem framing, interaction strategy, prototyping, validation and reusable UX patterns

Team

Product, Engineering, Research and Business Stakeholders

Faster task completion
34%Faster task completion
Higher successful task completion
23 ptsHigher successful task completion
Higher user confidence
39%Higher user confidence
Reusable AI interaction patterns
8Reusable AI interaction patterns

Illustrative validation metrics. Replace with verified project results before publishing.

01Context

AI had information.Employees needed context.

Frontline employees constantly move between customers, inventory, operational tasks, policies, exceptions and internal systems. The opportunity initially looked straightforward: make information easier to access through conversational AI. The work revealed a more important challenge.

Employees didn’t always need another conversation. They needed the right information at the right moment.

The design challenge became

How might we make AI useful during active work without forcing employees to stop what they are doing and have a long conversation with a chatbot?

Context by the numbers

Recurring frontline workflows explored
6+Recurring frontline workflows explored
Primary workflow categories
4Primary workflow categories
Primary product surfaces
2Primary product surfaces
Potential AI interaction models evaluated
5Potential AI interaction models evaluated

02Problem

The problem wasn’t accessto information. It wascontext switching.

  1. 01Stop the current task
  2. 02Search or open another system
  3. 03Formulate a question
  4. 04Interpret the response
  5. 05Verify whether it is correct
  6. 06Translate the answer into an action
  7. 07Return to the original workflow

Speed

Trust

Control

The product needed to help employees move quickly without removing their ability to understand, verify and control what happened next.

Example baseline usability metrics

Successful task completion
68%Successful task completion
Average time to complete priority tasks
3:40Average time to complete priority tasks
Confidence before taking action
3.1 / 5Confidence before taking action
Wrong-path or failed recovery rate
31%Wrong-path or failed recovery rate

03Exploration

The question changed.

Instead of asking how should the chatbot work? I started asking when should this experience even be conversational?

  • 01

    Conversational

    Best for: Ambiguous questions, discovery and exploration

  • 02

    Structured guidance

    Best for: Predictable multi-step tasks

  • 03

    Recommendations

    Best for: Decision support when AI can narrow options

  • 04

    Direct action

    Best for: Clear, low-risk and reversible tasks

  • 05

    Human escalation

    Best for: High-risk, uncertain or policy-sensitive scenarios

Interaction model comparison

Best use — Predictable, repeatable multi-step tasks

04Validation

What I needed to learn

  • Q01

    Can employees understand what the AI knows?

  • Q02

    Can they verify a recommendation quickly?

  • Q03

    Can they recover when AI is uncertain?

  • Q04

    Can the interaction stay understandable and accessible across different ways of using the product?

Validation at a glance

12

Participants

3

Rounds of usability testing

4

Priority workflows

30+

Interaction scenarios

2

Primary device types

Replace with verified study numbers before publishing.

Validation results

  • 01 · Task completion

    Baseline68%

    Validated design91%

    +23 percentage points

  • 02 · Time on task

    Baseline3 min 40 sec

    Validated design2 min 25 sec

    34% faster

  • 03 · User confidence

    Baseline3.1 / 5

    Validated design4.3 / 5

    39% improvement

  • 04 · Failed recovery

    Baseline31%

    Validated design12%

    61% reduction

Illustrative validation metrics. Replace with verified project results before publishing.

Finding → design response

  • Users hesitated when AI recommendations appeared without enough supporting context.

    Expose relevant evidence, source information and important context beside the recommendation.

  • Long conversational exchanges slowed predictable tasks.

    Move repeatable steps into structured UI.

  • Uncertainty was interpreted as system failure.

    Design explicit uncertainty, fallback and recovery states.

  • Dynamic AI responses created accessibility and predictability challenges.

    Create consistent semantic structure, focus behavior, status messaging and predictable actions.

05Decisions

The decisions thatshaped the product

Decision 01

Conversation would not be the default.

  • Chat-first for everything
  • Chat plus shortcuts
  • Structured UI with conversation on demand

What I considered

34%

Faster task completion

What we observed
Predictable tasks became slower when every action was hidden inside conversation.
What I decided
Use structured UI for repeatable tasks and conversation only when ambiguity requires it.
Why
Employees already knew what they wanted. Typing it out added work instead of removing it.
What changed
The assistant became a layer inside the workflow rather than a separate destination.

Decision 02

AI needed to show context.

  • Confidence score only
  • Expandable rationale
  • Evidence shown inline by default

What I considered

39%

Improvement in user confidence

What we observed
Users hesitated when recommendations appeared without enough explanation.
What I decided
Expose relevant context and supporting information before asking employees to act.
Why
Trust came from seeing the inputs, not from the tone of the answer.
What changed
Every recommendation now carries its source, timestamp and the data it was based on.

Decision 03

Verification before action.

  • Auto-execute with undo
  • Always confirm
  • Confirm by impact level

What I considered

XX%

Reduction in unintended actions · pending verified data

What we observed
Employees wanted more control when AI recommendations affected meaningful decisions.
What I decided
Add explicit review and confirmation before higher-impact actions.
Why
Undo is not a real safety net when the action reaches a customer or another team.
What changed
Actions are now tiered, and the interface only asks for confirmation where the stakes justify it.

Decision 04

Uncertainty became a designed state.

  • Hide low-confidence answers
  • Generic error message
  • Explicit uncertainty state with next steps

What I considered

61%

Reduction in failed recovery

What we observed
When AI did not have enough information, users often interpreted the experience as broken.
What I decided
Clearly communicate what the system knows, what it does not know and what the employee can do next.
Why
A visible limit is easier to work with than a confident wrong answer.
What changed
Uncertainty, fallback and escalation became first-class states in the pattern library.

Decision 05

Accessibility became part of the interaction architecture.

  • Audit at the end
  • Component-level fixes
  • Accessibility baked into the shared patterns

What I considered

What we observed
Dynamic AI output broke focus order and left screen reader users without status updates.
What I decided
Design keyboard navigation, semantic hierarchy, focus behavior, status communication, readable error states and predictable actions into the reusable interaction patterns.
Why
Fixing it once in the pattern is cheaper and more consistent than fixing it in every workflow.
What changed
Accessibility moved from a review step to a property of the interaction model itself.
Primary workflows keyboard accessible
100%Primary workflows keyboard accessible
Target accessibility standard
WCAG 2.2 AATarget accessibility standard
Critical accessibility issues after final validation
0Critical accessibility issues after final validation
Reusable accessible interaction patterns
8Reusable accessible interaction patterns

Example accessibility outcomes. Editable placeholders.

06Final experience

From question to action

  1. 01 · Understand

    The product recognizes relevant workflow context before requiring the employee to explain everything manually.

    3 systemsContext combined into one experience

  2. 02 · Guide

    The interface surfaces likely next steps instead of requiring employees to discover every action through conversation.

    5 → 2Choices reduced to two primary actions

  3. 03 · Verify

    Employees can understand why the recommendation was made before acting.

    4.3 / 5Average confidence after validation

  4. 04 · Act

    The employee stays in control and completes the task with fewer steps.

    34%Faster completion

  5. 05 · Recover

    When AI is uncertain, the interface provides a clear fallback or escalation route.

    88%Successful recovery

07Strategy

The bigger shift

We weren’t designing an AI assistant. We were defining where AI should participate in the work.

  1. 01Inform
  2. 02Recommend
  3. 03Guide
  4. 04Act
  5. 05Escalate

The right level depends on five factors

Context
What information does the system already know?
Confidence
How certain is the AI?
Risk
What happens if the recommendation is wrong?
Reversibility
Can the employee undo the action?
User control
How much authority should remain with the employee?

What this created

Instead of designing AI behavior independently for every new workflow, the project created a reusable decision model for choosing between conversation, structured UI, recommendations, automation and human escalation.

5

AI interaction models

8

Reusable interaction patterns

4+

Workflows supported

2

Primary product surfaces

1

Shared decision framework

08Outcome

The outcome

34%

Faster task completion

Common workflows required fewer steps and less context switching.

23 pts

Higher task success

Task completion improved from 68% to 91% during validation.

39%

Higher confidence

Users reported greater confidence before taking AI-supported actions.

61%Reduction in failed recovery

For employees

  • Less context switching
  • Faster access to relevant information
  • Clearer next steps
  • More confidence before acting
  • Better recovery when AI is uncertain

For the product

  • Reusable AI interaction model
  • Shared patterns across workflows
  • More consistent accessibility behavior
  • Clearer rules for AI autonomy

For the business

  • Potentially faster operational execution
  • Reduced dependency on manual information lookup
  • More scalable AI implementation across frontline workflows
  • More consistent product behavior across teams
Faster task completion
34%Faster task completion
Successful task completion
91%Successful task completion
User confidence
4.3 / 5User confidence
Successful recovery
88%Successful recovery
Reusable AI interaction patterns
8Reusable AI interaction patterns
Supported workflow types
4+Supported workflow types

Validation and scale metrics shown as editable placeholders until verified against final project data.

What I would measure next

  • AI recommendation acceptance rate
  • Task completion time
  • Recovery success rate
  • Escalation rate
  • User confidence
  • Accessibility task completion
  • Number of conversational turns required
  • Percentage of tasks completed without leaving the workflow
  • AI override rate
  • Repeat-use rate

09Reflection

What I learned

This project changed how I think about AI experiences. The best AI interaction isn’t always a conversation.

Sometimes the better experience is understanding the user’s context, showing the right information, providing a clear next step and knowing when the system should get out of the way.

The most important design decision wasn’t how the chatbot looked. It was deciding when conversation was useful, when structure was better and how much control AI should have.