I designed the AI Companion not as a feature, but as a platform capability that could scale across multiple products and recruiting workflows.
At Employ, I led the design of the first AI Companion for recruiters across a suite of three products (JazzHR, Lever, Jobvite) serving 22,000+ customers. The work moved from an undefined AI mandate to a launched product and a shared AI foundation the company continues to build on.
Principal Product Designer leading AI product design end-to-end. Also led the Design Centre of Excellence — the governance layer that shaped how the broader design team practiced AI design.
AI Companion vision and North Star
Interaction architecture
Trust architecture framework
AI pattern library
Accessibility-compliant components (WCAG)
Executive vision presentation for the Board
Executive pressure to ship AI. No product vision, no design strategy, no shared patterns across three acquired products with different design DNA.
Turn an AI mandate into a platform capability, not a bolt-on feature.
Reduce recruiter cognitive load across the hiring workflow.
Establish a governed foundation the whole product suite could build on.
An agentive AI Companion designed as a platform capability, anchored on a trust architecture and delivered through a governed AI pattern library shared across JazzHR, Lever, and Jobvite.
Key Goal title summary
Turn a vague AI mandate into a coherent platform capability — one that reduces recruiter cognitive load, earns trust across three distinct product audiences, and gives every team a shared foundation to build on.
- Recruiters were not asking for more AI features. They were drowning in cognitive load: switching screens, re-entering data, and making judgment calls on incomplete information.
- At the same time, Employ had three acquired products with different design DNA converging on a shared AI strategy. Building AI feature-by-feature would have produced three different mental models for what AI means on the platform. A UX problem, an engineering problem, and a trust problem.
- The real question was not what AI features should we build. It was what does a recruiter need to trust before they act on an AI recommendation, and how do we design that once for the whole platform?
Reframe the brief.
Stakeholders arrived with a feature list: draft job descriptions, suggest candidates, send emails. Discovery reframed it. Instead of "build an AI assistant," the brief became: help recruiters make better decisions, faster, without losing trust in the process. That reframe reordered every downstream decision.
Partner with the executive layer early.
I worked directly with the CPO and two VPs of Product to translate design decisions into business language. Not "here's the flow," but "here's the risk we eliminate, the trust we earn, and the retention we protect." That vocabulary secured board investment for a strategic AI direction rather than a feature roadmap.
Design the interaction architecture in three layers.
- Context layer — the assistant continuously interprets the recruiter's workflow: the role, pipeline data, past actions.
- AI capability layer — generative text, candidate summarization, workflow recommendations, risk flagging with reasons.
- Action layer — the assistant helps the recruiter take the next step immediately, reducing tool switching.
Anchor the system in trust architecture.
Four thresholds became the evaluative lens for every AI interaction:
- Transparency — the AI explains its reasoning.
- Accuracy — it is right often enough to be relied on.
- Reversibility — errors are easy to undo.
- Control — the recruiter can override at any point.
Codify AI-specific interaction patterns.
A separate pattern library for the AI experience layer, governed through the Design Centre of Excellence: confidence indicators, thinking states, suggestion cards with visible reasoning, correction flows, escalation paths to human review, and reusable "AI assists / human decides" and "AI acts / human can undo" flows.
Design for confidence states, not just success states.
High-confidence outputs get direct presentation. Low-confidence outputs get softer language, visible reasoning, and a low-friction override. Graceful override is not a failure state — it is the system respecting the human's judgment.
Lead through the Design Centre of Excellence.
Rather than managing designers directly, I built the CoE: design principles, review process, quality standards, contribution model for new AI patterns, and mentorship structures. Other product teams could propose new AI patterns, but they had to demonstrate they had tested for trust and accessibility before entering the library. That model kept the library honest and made platform-level thinking a shared practice, not a personal one.
Move fast without design debt.
Dual-track: iterate on core flows in parallel with the accessibility-compliant component set. Design QA checkpoints throughout the build. POC to working prototype in under two weeks; from there to MVP inside the same tenure.
Designing AI, not just advocating for it. The AI Companion is agentive. It acts, not just responds. That required a new interaction vocabulary. Every state was mapped: confident recommendation, low-confidence suggestion, conflicting signals, no data available. Each got its own interaction treatment.
The output was not a feature. It was a trust interaction model and a governed AI pattern library, sitting alongside the standard design system, that any product team on the platform could build against without reinventing the mental model.
The patterns launched with the AI Interview Companion on June 3, 2025. They continued into the AI Screening Companion (Nov 2025), the Talent Fit Dashboard, and the agentic capabilities released after my tenure. Evidence that the foundation was durable, not just delivered.
The trust model and interaction grammar were shared across JazzHR, Lever, and Jobvite. Density, complexity, and pace flexed per product audience.
Ship fewer AI moments with governed patterns; refuse a feature-first sprawl that would have set the platform back two years.
The assistant surfaced at decision points where recruiters were already pausing or switching tools. Pushing adoption harder would have burned trust early.
The next chapter is agentic.
Systems that plan and act across longer horizons of a workflow, with humans staying in the loop where judgment and accountability live. The design questions get harder: how does an agent surface intent, ask for permission at the right altitude, and hand off gracefully when it is uncertain? The trust architecture built here is the starting point for those answers.
Title / Summary
Designing for AI requires a different interaction vocabulary than traditional software. You are designing for behavior and judgment, not features and states.
Trust is not a section of the UI.
It is the operating system of an agentic product. Every state, every override, every confidence indicator either builds it or spends it.
Frameworks outlast projects.
The trust architecture and the pattern library continued to shape releases after my tenure ended. That is the shape of impact I aim for at the platform level.
The reframe was the highest-leverage moment in the project.
Everything else compounded from it. Next time, I would spend even more time shaping the brief with executives before the first design artifact appears.
Title/Summary
If I had one more cycle, I would have instrumented adoption of the AI pattern library across product teams. I have a strong qualitative read on where it went; a quantitative one would have made the platform case even sharper to leadership.
