Staff Designer, Conversation Design Lead · Agentic & Generative AI

Design that teaches
AI how to speak.

I design how AI behaves: the reasoning, trust, and orchestration decisions a system makes before it says a word, not just the words it says. It’s a discipline still being invented, and I’ve spent the last several years building it inside a company that didn’t have one.

View selected work Get in touch
Selected Work

Here’s where I designed
how AI behaves.

Two full case studies are open to everyone, no password needed. The other three quote internal research and unreleased work, so they're shared by request, and every card's Highlights panel is free either way.

01

Disambiguation Design

Shaping how the organization handles ambiguous requests, moving from guessing wrong or giving a giant generic response to asking the right clarifying question.

10+enterprise customers demanded the feature back within weeks of its removal
Conversation Design Intent Understanding Organizational Advocacy Framework Development
Full case study →
Designing Disambiguation
Full case study · open to everyone
The insight

Ambiguous requests are the default condition. 90% of user queries are medium to high ambiguity. A well-designed system doesn’t overwhelm users with every possible answer. It asks one focused question.

What I did

Held conviction on this problem for three years, built the research to prove it, and shaped how the organization thinks about clarification. The patterns I created improve intent understanding by being honest and human: communicating what the system knew and asking an intelligent, well-designed clarification.

Current status

V1 shipped late 2025. Design is now a named partner in architectural decisions from day one. The next version is in active development, with design defining the experience from the ground up.

The challenge

I identified this as critical three years before the organization was ready to hear it. When the feature was removed, customers demanded it back within weeks. That forced the org to reckon with what they’d dismissed.

Full case study →
02

Creating a Conversation Design Discipline at ServiceNow

Principles, training, patterns, and tools that shifted how a 1,000-person design org approaches AI behavior.

600+designers directly trained; enablement reached close to 2,000 people
Discipline Building Enablement Systems Thinking Thought Leadership
Full case study →
Building a Design Discipline
Full case study · open to everyone
The insight

Conversation design is a discipline. It requires linguistic patterns, behavioral principles about when to ask versus act, and design thinking. It’s not decoration on top of a system. It shapes how the system behaves.

What I did

Built the entire conversation design discipline from the ground up:

  • Principles & behavioral standards
  • Training curriculum
  • Reusable patterns
  • Evaluation frameworks & golden datasets
  • Engineering-design collaboration tools
Current status

600+ designers directly trained, with enablement materials and recordings reaching close to 2,000 people across design, PM, and engineering. Teams use the patterns for every conversational AI decision. The training curriculum is being adopted across ServiceNow as standard designer onboarding.

The challenge

The LLM wave hit and 1,000 designers suddenly needed to build conversational AI. Most had no training, no framework, no shared understanding of what good conversation actually looked like.

Full case study →
03

Making AI Behavior a Design Decision

Sustained advocacy that turned behavioral quality into architecture, from a designer’s opinion to the Otto PRD’s named source of truth.

~4×deflection: 11.4% vs. 2.75% for the legacy path, from early-access customer data
Advocacy Quality Rubric PRD Authorship Organizational Design
Full case study →
Making AI Behavior a Design Decision
Preview · no password needed
The insight

AI behavior isn’t decided by the model, it’s decided in the orchestration layer no one was designing or measuring. That’s a discipline gap as much as an execution gap: conversation design for agentic AI barely existed as a defined practice anywhere yet, and ServiceNow was later than most to start closing it.

What I did

Spent months building the evidence, partner relationships, and institutional infrastructure to make behavioral quality an architecture decision instead of a designer’s opinion, including authoring 11 of the 26 criteria in the org’s quality rubric.

Current status

My standards are the named source of truth in the Otto PRD, I lead a design-org OKR on conversation quality, and the designed experience deflects 11.4% of cases against the legacy path’s 2.75%.

The challenge

Getting a discipline recognized inside an organization that wasn’t looking for one, sustained advocacy across product cycles, not a single reframe that landed and stuck.

Full case study →
04

Designing Conversational Trust: Demonstrated, Not Performed

Making the case that trust, not tone, is the variable that determines whether an agentic AI experience works, proven on the two surfaces that shipped: what Otto says (tone) and what it knows (context).

+14 ptstone and empathy improved (73% → 87%) after the redesigned system prompt shipped
Trust Design Behavioral Specification PRD Authorship Systems Thinking
Full case study →
Designing Conversational Trust
Preview · no password needed
The insight

The instinct when an AI’s voice feels off is to make it sound friendlier. That aims at the wrong target. Trust, not tone, is what determines whether an agentic experience works, and it’s a downstream effect of demonstrated competence, not something you can write your way into.

What I did

Authored a five-part behavioral standard in the Otto Conversation Experience PRD, each requirement specified precisely enough that engineering could build directly from it, starting with a full rewrite of the system prompt.

Current status

Tone and context are shipped and measurably improving trust scores, the two surfaces this study proves in depth. The rest of the standard (memory, control, a response taxonomy) is authored and covered in separate work.

The challenge

Proving that warmer language wasn’t the lever, and that the org needed to invest in demonstrated reliability instead, a harder, slower case to make than a tone pass.

Full case study →
05

Agentic AI Processing Transparency

Redesigning how autonomous and agentic AI communicates what it’s doing, moving from technical logs to plain language that builds trust.

0escalations to disable processing messages since the redesign shipped, teams had been turning the feature off
Agentic AI Trust Design User Mental Models Language Architecture
Full case study →
Designing Processing Transparency
Preview · no password needed
The insight

Design for trust, not transparency. Users don’t need to see everything AI is doing. They need to see what builds confidence in what’s about to happen.

What I did

Reframed the problem from “how do we show everything” to “what builds trust,” then designed the systems that made that distinction real:

  • AI Blueprint concept
  • Three-tier threshold framework
  • Prompting guidance that generates processing language instead of hard-coding it
Current status

In active deployment across ServiceNow Otto. Technical log messages are gone from the surfaces where it shipped. Research participants now describe the language as clear, human, and trustworthy. The spec has reached 900+ recipients across the org. Step grouping model in development.

The challenge

Organizations default to showing everything. That creates overwhelm, not trust. The harder design problem is knowing what to hide and what to surface.

Full case study →
Alea Abrams
About

I make AI a better communicator.

I design how AI systems think and speak. As the Staff Designer leading ServiceNow's conversation design discipline, I build the behavioral frameworks that shape how entire product lines speak: architectural decisions, not surface-level polish. I’ve spent years moving conversation quality into architecture reviews, rather than UX passes after the decisions are already made. That’s settled now. I’m the design lead on the design org’s OKR to improve AI content quality, which makes that seat at the table a formal responsibility, not something I have to keep winning.

The problems I’m most drawn to are the ones no one has solved yet: how agents should communicate uncertainty, when AI should ask versus act, how users form accurate mental models of what a system can actually do. These aren’t UX problems. They live in the system, not on the surface. The stakes are high enough that I don’t think they belong anywhere else.

“Alea consistently pushes the boundaries of what great design looks like. Her ability to combine deep technical understanding with creative problem-solving is rare.”

“If you’re looking for someone who blends craft excellence, AI expertise, and a forward-thinking approach to innovation, Alea is that person.”

Tami McBride, Former Manager · ServiceNow
Prompt Design Agentic AI UX Organizational Strategy Conversation Design AI Mental Models Designing Uncertainty
Design Philosophy

Words are the interface.

01
Conversation is the oldest interface

Spoken language is at least 100,000 years old. Writing is around 5,000. Touchscreens, maybe 20. Conversation is how people have always navigated complexity and built trust. AI didn’t invent that, it inherited it. Designing it well means respecting what was already there.

02
Every word is architecture

Linguists call it pragmatics: the meaning between what’s said and what’s understood. In AI, that meaning is set by decisions most people never see, the prompt, the pipeline, the logic for what to say when the system isn’t sure. That isn’t polish, it’s the system. Someone is making those calls either way.

03
Language carries weight

An AI that uses jargon locks people out. One that hides uncertainty removes informed choice. One that presents a guess as fact isn’t making a tone mistake, it’s making an ethical one. Language has always governed access, and AI reproduces its builders’ assumptions at scale. I’d rather name that than not.

Writing & Talks

Speaking & writing.

2026
Why AI Conversations Go Wrong
The design failures that break trust in conversational AI, and how to fix them.
ServiceNow Knowledge 2026 · Workshop & Article
Read →
2026
The Promise and the Gap: Rethinking Accessibility in Conversational AI
Why accessible design for AI means rethinking transparency rather than adding features.
Essay · Accessibility & AI
Read →
2026
Should AI Sound Like Me?
Why "human-like" is the wrong design goal for AI conversation, and what should replace it.
Essay · Humanity & AI
Read →
2025
Stop Conversationalizing Your UI
When adding chat to a UI makes things worse, and when conversation actually belongs in design.
Essay · Conversation Design
Read →
2024
Mind Meets Machine: Leveraging Mental Models for Superior AI Interactions
How understanding user mental models transforms AI behavior from confusing to intuitive.
Chatbot & AI Conference · Talk & Article
Read →
Let's connect

Ready to design AI that earns trust?

Open to job opportunities, speaking engagements, and workshops.

Email LinkedIn Download Resume