3,000 target users · 30% faster campaign setup · 40% lift in trust scores
IAS Agent: AI-Powered Ad Verification Assistant
Role: Lead Product Designer · Timeline: 6 months (2025 - 2026) · Platform: Signal
3,000 users engaging with IAS Agent · 30% shorter campaign setup time · Trust scores up from ~40% to ~80%
Finalist, 30th Annual Webby Awards - AI, Marketing & Branding
THE PROBLEM
Mid-market advertisers and independent agencies were losing hours to fragmented reporting, manual campaign setup, and inconsistent brand safety enforcement across regions - friction that slowed activation and put compliance at risk.
I started with Pendo analytics to find where users dropped off, then used interviews and usability testing to understand why.
THE APPROACH
IAS Agent addressed this as an intelligent hub sitting above the UI - a consistent entry point for navigation, product questions, and task execution, wherever a user was working. Rather than treat this as a one-off feature, I authored the Agent Interaction Model - a shared framework adopted by Signal's Product, Design, and Engineering teams for how AI agents behave across the platform: transparent about its reasoning, honest about uncertainty, consistent wherever it appears. IAS Agent became its first grounded implementation.
To make these principles concrete and reusable, I also built an IAS Agent–specific pattern library, connected to Signal's broader design system. Other designers have since picked it up as a foundation for their own AI workflows, extending consistency beyond IAS Agent into other product areas.
HOW I GOT THERE
Exploring the options
Early concepts ranged from a pure chat interface to a guided query builder. Pure chat left users staring at an empty field, and a query builder felt too rigid for everyday questions. The hybrid model, structured prompts you can refine in plain language, took the best of both.
Testing and what changed
Early testing showed users didn’t trust answers they couldn’t check. I redesigned the response so users could see exactly what the agent had interpreted and the sources behind it, and change either one. Trust scores rose from roughly 40% to roughly 80%.
BEFORE
AFTER
Same question, same answer. Adding (1) footnote markers in the answer, (2) a numbered source list and (3) a "Related article(s)" box moved user trust from roughly 40% to roughly 80%.
Turning principles into standards
The Agent Interaction Model only works if every team applies it the same way. I turned its principles into six standards that Product, Design, and Engineering could build against.
01
One Agent, one identity
The same icon, color, and drawer everywhere. Page depth shows through the header label and welcome content, never a new persona.
04
Cross-page answers say so
The response states the switch and deep-links to the right page. Detection must clear a 90% precision and recall bar before launch.
02
Present, never in the way
One fixed launcher location, minimized by default. Proactive nudges are rare and dismissible.
05
Acknowledge, then redirect
If nothing clears the confidence threshold, the agent says so and offers search or human support. It never fabricates an answer.
03
Context shown, never assumed
The header names what the agent is drawing on, and one control returns to general help without losing the conversation.
06
Conversations persist
History carries across navigation, with a light "now viewing" marker when the context changes.
DESIGNED FOR WHAT'S NEXT
If a page ships a separate specialist agent: confirm before jumping, carry the user's intent forward, and hide the handoff behind one loading state.
Phasing the roadmap
I started by prioritizing the highest-impact use cases, then weighted feasibility more heavily once we were building. We shipped smaller, lower-risk wins first, like tag setup, which de-risked engineering and built momentum for harder work like cross-platform reporting.
Working with Product and Engineering
Feasibility shaped the sequence. I worked with Engineering to size each phase, and that conversation is why the roadmap starts small.
KEY DECISIONS
I chose hybrid input over pure chat - structured prompts paired with natural language cut the blank-page anxiety of an empty text field. Every response is also grounded and sourced - citations the user can click through to verify, never an unsupported claim; this transparency alone lifted trust scores from roughly 40% to roughly 80%, measured via a CSAT-style survey before and after the change shipped in H1 2026, and became non-negotiable for enterprise customers.
When context was missing, the agent said so and defaulted to a safe, general answer rather than guessing at specifics. And when a request fell outside what it could confidently handle - an unclear prompt, or simply outside the user's access - it offered a clear next step instead of a dead end; a hallucinated answer, or a silent failure, was never acceptable. These decisions later became the standard for how AI features get built across the organization.
THE SOLUTION
IAS Agent is the agentic AI layer inside Signal, handling up to 90% of core workflows - from routing requests to configuring reports - through conversational prompts and proactive alerts. It also serves as a grounded reference point, answering conceptual and how-to questions directly from Help Center documentation.
STRATEGIC TENSIONS
Two trade-offs shaped the work. Transparency vs. simplicity - showing AI reasoning without overwhelming novice users. Feasibility vs. impact - shipping smaller, lower risk wins like tag setup before tackling harder cross-platform reporting.
IMPACT
NORTH STAR METRIC
3,000
USERS ENGAGING WITH IAS AGENT
40% - 80%
USER TRUST SCORE, BEFORE VS. AFTER
30%
REDUCTION IN CAMPAIGN SETUP TIME
UP TO 90%
OF CORE WORKFLOWS SUPPORTED BY IAS AGENT