AI Experience Architect · Conversational AI Strategy · Agent + Prompt Systems

Designing AI people can trust.

I turn complex services, knowledge, and operational requirements into AI experiences people can understand and use—from agent behavior and prompt systems to recovery, governance, analytics, and adoption.

Explore selected work Download latest portfolio (PDF)ebony.keira@gmail.com

7+ yearsConversational AI and AI experience

Public + enterpriseGovernment, Microsoft, Airbnb, Fox

End-to-endDiscovery through launch and iteration

Human-centeredTrust, accessibility, recovery, control

Selected work

Three problems. One systems mindset.

These cases show how I approach public-facing agents, enterprise support experiences, and operational enablement. Sensitive implementation details are intentionally generalized. Illustrative examples are labeled, and no unverified performance metrics are shown.

Phoebe: A public-facing AI assistant for the City of Phoenix

Designing a citywide service experience around resident needs—not government structure.

Copilot StudioAgent architectureGovernance

Design response

I reframed city services as human goals, then created governed interaction patterns that could scale across departments. Ambiguous requests triggered focused clarification; low-confidence or unsupported requests moved into designed recovery rather than dead ends.

Resident languageStart with the problem as the person describes it.

Interpret + clarifyDetermine intent and ask only what is needed.

Route + respondUse approved knowledge and service logic.

Recover + learnRedirect, escalate, and improve from evidence.

Illustrative interaction · not a production transcript

Resident: “There is water running down my street.”
Phoebe: “Is it coming from a hydrant, a pipe, or are you not sure?”
Resident: “Looks like a hydrant.”
Phoebe: “I can help you find the right reporting path. Is the water creating an immediate safety hazard?”

The smallest useful clarification avoids a speculative answer. Approved service knowledge guides routing; a safety concern changes the next step. Exact live scripts, URLs, and thresholds are omitted.

Representative acceptance checks

Recognize everyday phrasing and retain details already supplied.
Ground answers in approved sources; avoid invented policy.
Clarify ambiguity, recover from unsupported requests, and escalate hazards appropriately.
Use accessible language and review transcripts, failed intents, drop-off, and task completion after launch.

These are illustrative checks; no production metrics are claimed.

Analytics & Reporting Limitations

In the Microsoft Government Community Cloud (GCC) environment used for Phoebe, platform reporting limitations restricted access to reliable chatbot usage and performance data. This prevented consistent measurement of conversation volume, containment rates, and resolution rates during my work on the project.

Within these platform constraints, I owned conversation design, intent mapping, branching logic, knowledge content, guardrails, and citizen-facing self-service workflows in Copilot Studio. The work centered on clear service journeys, grounded responses, appropriate clarification, and recovery and escalation paths. Quantitative performance results are not presented because the available reporting did not support reliable measurement.

Product ownershipFrom discovery and architecture through testing and post-launch improvement.

Trust by designConfidence-aware behavior, guardrails, accessibility, and human handoff.

Visible outcomePhoebe was elevated to the Phoenix.gov homepage as a flagship digital service.

Enterprise support: Designing repeatable conversation systems

Making complex support interactions clear, recoverable, measurable, and platform-aware.

MicrosoftWashington DOLAnalytics

Design response

I created reusable conversation patterns instead of treating every response as a one-off. Transcript review and conversational diagnostics surfaced friction, inconsistent responses, failed intents, and weak handoffs; those signals drove redesign and testing.

ObserveReview language, transcripts, failures, and drop-off.

DiagnoseSeparate intent, content, routing, and platform issues.

RedesignRefine dialogue, prompts, knowledge, and recovery.

EvaluateTest successful paths, edge cases, and failure behavior.

Platform fluencyCopilot Studio, QnA Maker, dialogue logic, APIs, and routing systems.

Quality systemsReusable frameworks for tone, accuracy, recoverability, and consistency.

Cross-functional deliveryProduct, engineering, analytics, UX, and operations partnership.

UltraCare: Turning workflow chaos into an operating system

Mapping operational knowledge into CRM-ready workflows, role-based forms, and a foundation for future AI enablement.

Workflow designEnablementKnowledge systems

Design response

I treated the internal operation as a product experience: identify the next decision, put the guardrail where it is needed, and make completion observable. The platform-agnostic documentation defines fields, ownership, and status changes that can transfer into digital workflows once a CRM is selected and configured.

MapExpose steps, owners, systems, decisions, and gaps.

StandardizeDefine expected actions and verification criteria.

OperationalizeCreate forms, queues, checklists, and escalation paths.

EnableTurn the system into teachable, reusable knowledge.

Program thinkingConnected procedures, tools, roles, controls, and adoption.

UsabilityDesigned guidance around the moment of action and the next required step.

Paper to electronicCRM-ready forms, ownership, statuses, and workflows prepared for migration. The CRM and AI have not launched.

Training + product enablement

Make complex software teachable.

A hands-on work sample from Entertainment Partners, showing the same stakeholder-led method I bring to conversation design and operational enablement.

Scenechronize: Script Sides Settings

A software tutorial guiding users through creating script sides settings.

Instructional designCamtasia

Watch on YouTube ↗

What this demonstrates

Turning subject-matter expertise into clear user guidance, validating instructions with stakeholders, and owning a learning asset from script through final edit.

How I work

From ambiguity to delivery.

I connect human needs, product behavior, technical reality, and operational adoption so the experience holds together beyond the demo.

Discover

Understand user goals, workflows, and constraints.

Define

Agree on the desired outcome, requirements, and risk.

Architect

Design intent, context, knowledge, and recovery.

Prototype

Make the proposed behavior tangible.

Evaluate

Check expected paths, ambiguity, and edge cases.

Ship + learn

Partner in delivery and review evidence to improve the experience.

Capabilities

Strategy with production depth.

I work across the experience layer and the systems that make it reliable, measurable, and maintainable.

AI experience architecture

Journeys, agent behavior, state, context, task flow, and interaction models.

Conversation design

Intent, utterances, dialogue, clarification, recovery, escalation, tone, and accessibility.

Prompt and grounding systems

Prompt structure, response standards, approved knowledge, guardrails, and consistency.

Evaluation and analytics

Transcript audits, failed-intent analysis, edge cases, UAT, drop-off, and quality review.

AI governance

Responsible behavior, limits, uncertainty, human oversight, documentation, and reuse.

Enablement and adoption

Workflow standards, training, operational guidance, stakeholder alignment, and change support.

AI changes what a product can do. It does not change what people need: clarity, control, recovery, and a useful next step.

Contact

Let’s make complex AI usable.

Available for senior individual-contributor and manager-level opportunities in conversational AI, AI experience, implementation, enablement, and digital product strategy.

Email Ebony LinkedIn