“There’s been a lot of noise lately around ‘AI agents,’ and suddenly job descriptions are asking for ‘AI‑agentic experience.’ This can make designers wonder:
‘Do I need to learn something entirely new to stay relevant?’”

No. If you’re a UX designer, you already think in the mental models required to design intelligent, autonomous systems. AI agents don’t introduce a new discipline; they expand what you already do.

What Is an AI Agent?

An AI agent is a system that takes action on behalf of the user, not just responds.
It observes context → interprets intent → decides → acts.

Examples:

• A health app that schedules your follow‑up appointment

• A tutoring system that adjusts the lesson when the student struggles

• A finance bot that optimizes savings automatically

This isn’t about screens.
It’s about behaviors.

Designers Are Already Equipped for This

Everything UX designers do is already relevant to agentic systems:

• Understanding user needs and motivations

• Mapping journeys, flows, and edge cases

• Creating predictable, trustworthy experiences

• Designing clarity, transparency, and usable behaviors

The foundation stays the same.

What Changes: You’re Designing Behavior, Not Just Interface

Design shifts from:

• “What screens do I create?”
  → “What decisions should the system make?”

• “What happens when the user taps X?”
  → “When should the agent act, pause, ask, or stop?”

Traditional UX vs. Agentic UX

The Real Skillset Designers Need (You Already Have It)

Not a new craft — expanded mental models.

1. Think in Intents, Not Tasks

What is the user trying to accomplish?
Agents need intent, not clicks.

2. Design the Decision‑Making Logic

You’re defining policies and reasoning paths, not flowcharts.

• What does the agent prioritize?

• How does it choose between multiple valid next steps?

• When should it infer vs. ask?

Example (travel agent):

• “Cheap flight” → prioritize price

• “I need to get there fast” → deprioritize layovers

• Confidence <70% → ask instead of acting

3. Add Constraints and Guardrails

Designers already do this: friction, confirmations, access control, error prevention.

In agentic systems, guardrails become behavioral boundaries:

• Actions the agent must never take

• When it must pause and confirm

• How it signals uncertainty

• What data it can use

• What must be explainable

Guardrails are design.

4. Communicate the Agent’s Reasoning

Trust comes from explainability.

Designers shape:

• How actions are explained

• How uncertainty is conveyed

• How the agent reveals its “thinking”

• When control returns to the user

5. Evaluate Outcomes, Not Screens

Traditional usability:
“Can users find the checkout button?”

Agentic usability:
“Did the agent book the right flight based on ambiguous input?”

The test object is now behavioral correctness.

UX Isn’t Diminished. It’s Expanded.

We still care about:

• clarity

• confidence

• trust

• comprehension

• error prevention

• ethical behavior

UX becomes more systemic, strategic, and outcome‑oriented.

Designers aren’t behind — we’re the ones who should guide agentic systems.

What You Might Actually Need to Learn

What’s new (but learnable):

• Basic AI literacy (LLMs, confidence scores)

• Co‑designing decision logic with engineers

• Designing for ambiguity (“I need to travel soon”)

These are extensions of what designers already do.

Closing Thought

AI agents aren’t a new skill designers must chase.
They’re an evolution of what we already do:
understand people, shape behaviors, and design for safe, meaningful outcomes.

When systems act on behalf of users, the cost of bad design is no longer frustration — it’s broken trust.

Designers aren’t behind. We’re the ones who should define how these systems behave.

Try This Today

Pick one feature you’re working on and ask:

• “If this could act autonomously, what decision would it make?”

• “What would it need to know to make that decision safely?”

• “When should it stop and ask the user for help?”

You just practiced agentic design.

If you gave nine different AI tools the same enterprise design challenge, would they solve the same problems? Would they prioritize the same features? Would any of them create accessible interfaces?

I decided to find out.

As design leaders, we’re seeing AI tools proliferate across our organizations. But most conversations stay surface-level: “AI is transformative” or “AI will replace designers.” I wanted to move past the hype and understand something more fundamental: how do different AI models actually approach complex design problems?

So I created a detailed design brief for “ConnectIQ,” an AI-enhanced contact center platform for companies with 500–5000+ employees, and tested it with nine AI tools:

Same prompt. Same requirements. Nine completely different results.

The most critical finding? Every single tool, all nine, failed accessibility.

The Experiment

The Challenge

ConnectIQ needed to serve three user types with conflicting needs:

• Customer service agents handling 100+ daily interactions

• Supervisors managing teams of 10–20 agents

• Operations managers forecasting and optimizing at scale

The design had to balance innovation with enterprise reality: real-time performance (<200ms), security compliance (SOC 2, GDPR, HIPAA), legacy system integration, and change management for thousands of users.

This tests whether AI understands:

• Multi-persona complexity

• Enterprise constraints

• Strategic vs. tactical thinking

• Innovation within operational limits

The Method

Text AI Phase:
Five models received the same detailed prompt asking for product strategy, IA, key features, workflows, design principles, AI philosophy, and success metrics. Target: 1,500–2,500 words.

Visual AI Phase:
Four tools received a minimal prompt: “Design the Agent Workspace and Supervisor Dashboard for ConnectIQ.”

I evaluated everything systematically: feature depth, enterprise understanding, creativity, usability, and critically, accessibility.

What the Text AI Models Revealed

Each model had a distinct “personality” and approach.

ChatGPT: The Balanced Generalist

Delivered 2,500 words of tightly structured analysis with hierarchical bullets. Balanced strategic thinking with tactical detail.

Unique contributions:

• ABAC (Attribute-Based Access Control)

• “Evidence-First Cards” UI pattern

• Best adherence to length constraints

Limitations: Conservative positioning, standard naming, less focus on human factors.

Claude: The Creative Humanist

2,800 words of narrative prose with novel frameworks and human-centered emphasis.

Unique contributions:

• “EQ Layer” for human states

• “Just-in-Time not Just-in-Case”

• “Human-in-the-Loop Guarantee”

Limitations: Less enterprise detail, fewer workflows, fewer modules.

Perplexity: The Research Scholar

2,600 words grounded in real enterprise system behavior.

Unique contributions:

• Edge computing for latency

• “AI authority levels” governance

• Separate AI performance metrics

Limitations: Less memorable positioning, longer latency targets.

Microsoft Copilot: The Technical Architect

4,000+ words, reading like a full PRD.

Unique contributions:

• Monte Carlo scenario analysis

• Causal inference models

• Aggressive latency targets

• Explicit bias testing

Limitations: Overly long, less creative, overwhelming detail.

Gemini: The Clarity Advocate

2,400 words with the clearest problem–solution framing.

Unique contributions:

• Graceful degradation

• Reduce effort, increase confidence

• Intent-centric positioning

• Non-judgmental interface

Limitations: Less technical innovation, fewer memorable concepts.

The Pattern: Strategic vs. Tactical Split

• Strategic thinkers: Claude, Gemini

• Balanced: ChatGPT, Perplexity

• Tactical implementer: Copilot

What they all missed: Accessibility, change management, user research, mobile, offline, internationalization.

What the Visual AI Tools Revealed

Four tools were selected for enterprise relevance.

🔸Figma Make: Industry Standard Polish

Polished, professional interfaces with a three-panel layout.

Weakness: Contrast violations, color-only indicators. Score: 8/15.

🔸Lovable: Full-Stack Prototyper

Generated working React code, not just mockups.

Weakness: Orange/red badges, color-coded bars. Score: 8/15.

🔸v0.dev: Power User Interface

Dark theme, high-density, enterprise-ready features.

Weakness: Severe accessibility violations. Score: 3/15.

🔸Claude Artifacts: Accessibility Leader

Light theme, whitespace, clarity-first.

Strength: Best accessibility baseline (12/15).
Weakness: Still not fully compliant.

The Universal Failure: Accessibility

All four visual tools made similar mistakes:

• Red/orange text on dark or light

• Color-only indicators

• Missing icons/patterns

• Color-only progress bars

Impact:
8% of male users (colorblind) would struggle. Low-vision users would fail. Enterprise audits would fail.

Beautiful ≠ accessible.

The Critical Finding: AI’s Accessibility Crisis

Every single AI tool failed accessibility. All nine.

Text AI Models

• Zero mention of WCAG

• Zero mention of screen readers

• Zero mention of colorblindness

• Zero mention of keyboard navigation

Visual AI Tools

• v0.dev: 3/15

• Lovable: 8/15

• Figma Make: 8/15

• Claude: 12/15

  ---

The Impact

In enterprise software, these failures mean:

• Legal risk

• Excluded users

• Failed procurement

• Expensive remediation

Accessibility is not optional.

What Design Leaders Need to Know

1. Tool Selection is Strategic

Use text AI for strategy, documentation, research.
Use design AI for visuals.
Use development AI for functional prototypes.
Never use AI alone for final decisions or accessibility.

2. Multi-Tool Synthesis is Optimal

1. Generate

2. Compare

3. Synthesize

4. Add what AI missed

5. Validate with users

3. Accessibility Requires Human Vigilance

Assume AI outputs are not accessible.

4. Skills Design Leaders Need Now

• Prompt engineering

• Critical evaluation

• Tool selection

• Synthesis

• Accessibility advocacy

• Ethical judgment

  ---

Key Takeaways

About AI Tools:

• Each has distinct strengths

• Design AI ≠ Development AI

• Multi-tool synthesis wins

• AI amplifies process but cannot replace it

About Accessibility:

• All 9 tools failed

• Text AI ignored it

• Visual AI violated contrast

• Human oversight is essential

About Design Leadership:

• Synthesis is the superpower

• Tool selection is strategic

• Accessibility advocacy is required

• AI accelerates but doesn’t replace human-centered design

  ---

What’s Next

AI will improve, but the insights remain:
AI amplifies design leadership, accessibility requires humans, and synthesis across tools produces better outcomes.

Let’s Connect

If you’re navigating AI transformation in design leadership, I’m interested in comparing notes.

Areas of focus:

• Evaluating AI tools

• Building AI-literate design practices

• Addressing accessibility gaps

The Mission

Why I decided to befriend banking bots

In preparation for a conversational AI–focused role, I dove headfirst into the world of banking chatbots. Because nothing says “fun weekend project” like analyzing how well AI can handle your money questions, right?

The Goal: Crack the code on conversational banking

Evaluate existing banking AI, identify what makes users want to throw their phones, and figure out how to make money conversations actually… conversational.

What I Did

Competitive Benchmarking

Deep-dived into major banking AI personalities like Erica, Eno, and friends to see who’s winning the “least annoying bot” award.

Experience Mapping

Traced user journeys from “check my balance” to “why did I spend $47 on artisanal soap again?”

Design Heuristics

Created principles for banking bots that don’t make people miss the good old days of waiting on hold.

Opportunity Spotting

Found ways AI could evolve from “balance checker” to “financial life coach” (minus the judgment).

The Conversational Landscape

Not all bots are created equal

The Players I Analyzed

Banking Natives: Erica (BofA), Eno (Capital One), KAI, Eva (HDFC)
Voice Assistants: Alexa, Google Home, Siri
Specialized Bots: Mental health, education, automotive
IVR Systems: Those phone trees we all love to hate

Conversational Banking Experience Map

Complete mapping across simple queries to complex financial planning conversations.

1. Simple Query Scenario

Quick information retrieval and basic banking tasks.

2. Account Management Scenario

Modifying account settings and personal information.

3. Loan Application Scenario

Complete loan application and approval process.

4. Investment Planning Scenario

Portfolio development and investment strategy implementation.

Meet Erica: The Overachiever

Bank of America’s AI that actually works

25M+ Users
676M Interactions in one year
Launched 2018

What Makes Erica Special

Beyond the basic “what’s my balance” stuff.

Proactive Alerts & Insights

✓ Duplicate charges detection
✓ Unused subscription alerts
✓ Spending behavior analysis
✓ Overdraft risk warnings

Natural Conversation Design

✓ Voice and text input
✓ Human-like follow-ups
✓ Context awareness
✓ Conversational memory

Personal Finance Education

✓ Behavioral spending tips
✓ Financial concept explanations
✓ Personalized recommendations
✓ Credit education content

Design Principles That Work

The secret sauce for banking AI

1. Clear Onboarding – Set expectations from hello

2. Graceful Fallbacks – When AI fails, fail beautifully

3. Tone & Personality – Professional but not robotic

4. Brevity – Nobody reads novels in banking apps

5. Input Flexibility – People say things in many different ways

6. Anticipation – Predict what users need next

   ---

Why Conversational AI Needs Special Heuristics

Traditional UI heuristics aren’t enough for conversational interfaces.

1. Clarity of Intent Recognition

The system should clearly understand and confirm user intent.

Do:
– “I understand you want to transfer $500. To which account?”
– Offer disambiguation options

Don’t:
– Guess user intent
– “I don’t understand” with no alternatives

2. Conversational Context Awareness

Maintain awareness of history and preferences.

Do:
– “Your checking account ending in 1234…”
– Remember preferences

Don’t:
– Ask for the same info repeatedly
– Lose context in long conversations

3. Graceful Error Handling

Provide helpful recovery options.

4. Appropriate Personality & Tone

Professional, warm, consistent.

5. Transparent Capability Boundaries

Communicate what the AI can and cannot do.

6. Progressive Disclosure

Start simple, reveal more on demand.

The Big Opportunities

Where banking AI could actually get exciting

Predictive Financial Insights

Move from “here’s your balance” to “here’s how to improve your finances.”

Hyper-Personalization

Every user’s financial situation is unique. AI should reflect that.

Emotional Intelligence

Recognize when users are stressed and respond accordingly.

Key Insight

Banking AI must balance convenience with trust.
Users want quick answers, but they also want to feel confident their money is in good hands.
The sweet spot: proactive, helpful, but never presumptuous.

What I Learned

Plot twists in the world of banking bots

Conversational AI in banking isn’t just about answering questions quickly. It’s about building digital relationships that feel trustworthy, helpful, and genuinely valuable.

The Sweet Spot

The best banking AI doesn’t replace human advisors - it enhances them.

Bottom Line

There’s a massive opportunity to elevate banking AI from a transactional assistant to a proactive financial partner.

The technology is there; we just need better design thinking.

Research Note

This was a self-directed exploration into conversational AI in banking.
All insights are based on publicly available information and user experience analysis.

“I want to help teams move faster together; not through more process, but through shared clarity, trust, and communication.”

The goal: reduce friction without adding bureaucracy. Clear communication is a speed advantage.

The 4‑Step Collaboration System

Listen → Align → Adjust → Agree

Use this staircase as the weekly rhythm when joining or reshaping a team.

Why This Matters

Fast-moving teams don’t struggle because of a lack of talent - they struggle because of misalignment.

This playbook reduces friction between Design, Product, and Engineering by making:

• Expectations visible

• Decision-making shared

• Handoffs predictable

• Feedback continuous (not reactive)

It’s intentionally simple - no heavy ceremonies.

Week 1: Listening & Learning

Short 1:1 conversations reveal where collaboration breaks today - expectations, timing, formats, and decision-making.

For Developers (ICs)

• What slows you down the most when you get designs?

• How do you prefer to receive designs: prototype, tickets, screenshots, or walkthrough?

• What’s your current process when designs don’t match technical constraints?

• How do you prefer to give feedback on feasibility?

• If you could change one thing about design–dev collaboration, what would it be?

For Dev Manager / Lead

• What do you need from design so your team feels confident going into a sprint?

• When things break down, what usually went wrong earlier?

• How should disagreements be handled?

• What has worked well in the past?

For Product Managers

• Where does friction happen today?

• How can we reduce rework during execution?

• How should we document design–dev decisions?

For Everyone

• What tools are non-negotiable?

• How should we make decisions during urgent pivots?

Output after Week 1:
Top 3 collaboration pain points + protocol for urgent decision-making.

Week 2: Quick Wins

Weekly Design–Dev Sync (15–20 min)

Before sprint planning → walk upcoming designs → devs flag risks → avoid mid-sprint surprises.

Enhanced Handoff Checklist

Every design handoff includes:

• Prototype link

• Final flow (start → end + alternatives)

• Edge cases (empty, error, loading)

• Responsive behavior (if relevant)

• Specs/redlines (if precision matters)

• Assets + interaction notes

• Technical constraints noted

• Priority level: MVP vs. Polish vs. Delight

• Definition of done

Optional: short Loom walkthrough (<5 min)

Resource Reality Framework

• MVP: Must work; polish deferred

• Polish: Customer/demo-ready

• Delight: Added when time allows

Output after Week 2:
Clearer handoffs + shared expectations + fewer blockers.

Week 3: Feedback & Adjustments

Ask the team:

• What worked better this sprint?

• What’s still frustrating?

• If we fix one thing next sprint, what should it be?

• How did MVP/Polish/Delight work in practice?

Output after Week 3:
Refined systems…

Writing Better Prompts

Tip 01 — Start with the who and what, not the how

Get clear on three things before you open Lovable: who is this for, what does it do, and what’s the single most important thing it needs to get right.

• ❌ “Build me a task tracker”

• ✅ “Build a weekly task tracker for freelance designers who need to log client hours and generate a summary at the end of each week. Keep it minimal and focused.”

Tip 02 — Use spatial and structural language

Describe where things go, how they’re grouped, and what the visual hierarchy should be.

• ✅ “Use a left sidebar for navigation, a main content area with a card-based layout, and a primary action button anchored to the top right.”

Tip 03 — Iterate in small, focused moves

Build the structure first, then refine section by section. Prompt 1: layout. Prompt 2: hero details. Prompt 3: features section.

Tip 04 — Reference real products for design direction

Name a product whose style you admire, or upload a screenshot directly as a visual reference.

• ✅ “Style this dashboard like Linear’s interface: clean, minimal, with generous whitespace and a monochrome palette.”

Tip 05 — Name your colors and fonts up front

Prepare a palette and font pairing before you start building, then bring those specific values into Lovable.

• ✅ “Use background #f7f5f2, text #1a1a1a, accent #2c6e63. DM Sans for body, Playfair Display for headings.”

Thinking Like a Designer

Design 01 — One feature, one prompt.

Get the homepage right first, then add the dashboard, then settings. Keeps Lovable focused.

Design 02 — Hierarchy beats decoration.

When something doesn’t look right, make the important element larger, the secondary smaller, and add more spacing. Size, weight, and spacing do the heavy lifting.

Design 03 — Constrain your palette.

2–3 colors max: one background, one for text, one accent. Every color should have a job.

Design 04 — White space is your friend.

If your build feels cramped, tell it: “Add more padding between sections and increase the spacing between cards.”

Design 05 — Know (and avoid) Lovable’s default look.

React + Tailwind + shadcn/ui defaults to white backgrounds, zinc/slate grays, rounded cards, and system fonts. It’s competent but generic. Override it with your own palette, fonts, and personality.

Your Toolkit

Color: Coolors · Realtime Colors

Fonts: Google Fonts · Fontpair

AI prep: Claude, ChatGPT, Copilot or Gemini to brainstorm palettes and refine prompts

Inspiration: Dribbble · Mobbin (real app screenshots by flow and pattern)

The workflow: Decide on palette, fonts, and layout before opening Lovable — then paste those specific values into your prompts.

Before You Build Checklist

1. Define your user (one sentence)

2. Define the core action (one thing your app must do well)

3. Pick 3 hex values: background, text, accent

4. Choose a font pairing (heading + body)

5. Find a reference product/screenshot

6. Sketch your layout (even on paper)

7. Write your first prompt - structure only, no styling yet

8. Layer in style in a follow-up prompt