Four AI products, two principles: transparency and agency. 38.8% media list creation, 57.5% search adoption.

PR teams monitor coverage using Boolean search queries—expert-level syntax that most PR professionals don't speak fluently. The Search Agent translates natural language monitoring goals into valid Boolean queries. You say what you're looking for; the agent builds the search syntax.

The design problem was making expert-level output accessible to users with no search expertise, while preserving the power and precision of Boolean logic. This required deep behavioral specification. I provided feedback on guardrails covering: intent inference (understanding what the user is actually trying to find), entity handling (recognizing brands, topics, keywords), query construction standards (building syntactically correct searches), transparency requirements (showing the user what query was built and why), and failure modes (how the agent recovers when it doesn't understand the request, or when the query returns no results). The rigor of that underlying specification is what makes the surface experience feel trustworthy.

57.48% of prompts led to viewing coverage on search results screen. 44% returning user rate.

You describe the journalists you want to reach in plain language—"Tech journalists who cover startups and AI" or "Automotive journalists with international audience reach"—and the agent assembles a targeted list in seconds.

The design problem was making AI-generated recommendations feel trustworthy enough to use for real outreach, where the stakes are professional reputation. The solution was radical transparency. Every recommendation comes with a detailed match explanation: a summary of why we matched them, their top topics (e.g., automotive, motorsport), their audience scope (consumer, international, trade), and their top entity mentions in recent coverage (e.g., Porsche, Ferrari). Every result can be refined by topic, outlet, and location. The conversational thread model lets users iterate and refine rather than starting over from scratch.

38.78% of sessions led to media list creation. 44% of users returned within 30 days.

When a PR team sends a pitch, they need to know which journalists are most likely to care. The Optimized Pitch Recipients Agent evaluates every journalist in the Muck Rack database against a pitch and surfaces the most relevant recipients with reasoning.

The design challenge was building enough transparency that users would trust the recommendations without feeling replaced by the AI. We showed the reasoning for each match. Users could accept or reject recommendations. The AI informed; the user decided.

This release also included a set of small but significant overdue UX improvements to the pitching flow itself—restructuring the flow to be more user-friendly and reducing the overall number of steps by 20%.

Customer reaction at launch was immediate: "The pitch recipient optimization tool is KILLER." That kind of unsolicited feedback is a better signal than any adoption metric. Among recommended recipients, click-to-open rates doubled compared to non-recommended ones. Email open rates improved by 10%.

When responding to inbound inquiries from journalists—requests for comment, crisis comms—PR teams need context fast. The Media Brief Agent generates a complete journalist briefing document in seconds: background on the journalist, their coverage themes, potential questions they might ask, and suggested response angles. All without leaving the workflow.

My role as director was ensuring all four products operated on a consistent principle and felt like part of the same system. The through-line: two principles, applied consistently across all four. Transparency: make AI reasoning visible. Agency: keep the human in control. In practice, that means every AI-generated output is accompanied by reasoning. Every recommendation can be overridden. The AI proposes; the user decides.

Those principles sound simple. Executing it consistently across four different products, four different user contexts, and four different risk profiles: media discovery, list building, pitch optimization, and crisis response. That's where the design work actually lived. The consistency wasn't about aesthetics. It was about trust. If users trust the AI to be transparent in one context, they're more likely to trust it in another. If they understand how one agent reasons, they're faster to understand the next one.

To make that consistency operational, I worked with my team to build two foundational artifacts.

The first was a platform-wide AI governance framework: 11 behavioral guardrails applied uniformly across every agent on the platform. These aren't suggestions. They're requirements. The guardrails cover scope enforcement, tool transparency, intent-led interaction, capability disclosure, state preservation, error recovery, latency expectations, voice and tone, and bias mitigation. Every agent ships compliant with all 11. No exceptions carved out for timeline pressure.

The second was the Agent Specification Template: a mandatory pre-development source of truth used by product, design, engineering, and QA before any agent moves into build. It covers agent purpose, scope, adjacent agents, input and output types, agent-specific guardrails, edge cases, adversarial and misuse scenarios, failure handling, UX patterns, and pre-launch checklists. The template exists because agentic products fail in specific, predictable ways. Mishandled edge cases. Unclear failure states. Inconsistent transparency. Getting every stakeholder aligned before the first line of code is how you prevent those failures at scale, not patch them after launch.