Research at a Crossroads

Meta's product cycle had compressed dramatically. AI-led acceleration had shrunk design and engineering timelines from months to weeks. The research org hadn't kept pace. The traditional cycle — file a request, wait 4–6 weeks, receive a deck after the product had shipped — was no longer compatible with how Meta builds. Research was becoming either a bottleneck or an afterthought.

The leadership ask was direct: make research as fast and AI-native as the rest of the product stack. The research org's concern was equally direct: don't automate away the judgment that makes research useful in the first place. Both were right. The design role was to find the path that honored both.

Core User Problems

• Research outputs were scattered across team drives, email, and disconnected tools — knowledge silos with no central record
• Manual, repetitive pre-field work consuming ~61% of researcher time
• Fragmented tooling across six or more platforms (survey tools, docs, project trackers, research repositories)
• Limited access for non-researchers — PMs, designers, and engineers couldn't easily find past insights
• Trust fragility — the org needed AI to sharpen research thinking, not replace it

How might we make Research Studio the place where a study lives — not just another tool researchers talk to?

Alpha answered the wrong question well. It gave researchers a fast, capable chat interface — and they used it. But after each session, the output landed in the same fragmented sprawl as before: a discussion guide in Google Docs, findings in a deck, approvals in email. The AI was useful. The work was still scattered.

The Beta design question wasn't about velocity — it was about coherence. Could we make the project the permanent home of a study, so that every artifact, conversation, finding, and approval existed in one place from brief to shareout? After Alpha shipped in March 2026, the eval data confirmed the path: go narrower and deeper, centered on the project as the organizing unit.

Our Strategy

The Pivot: From Chat History to Research Record

Alpha looked like a search box with memory. You typed a question, got an answer, and a list of past conversations accumulated below. It was familiar — and it reproduced exactly the fragmentation it was supposed to solve. Chats weren't connected to studies. Artifacts lived elsewhere. There was no through-line from question to finding to shipped research.

The Beta pivot was structural: the study, not the conversation, became the unit of work. A Project is a permanent container that holds everything generated in the course of a research study — AI-generated artifacts, researcher-created documents, chats with the agent, synthesis, approvals, and the final shareout. Chat doesn't disappear; it becomes one tool within a project, anchored to the work rather than floating alongside it.

• Create or enter a project — start fresh or promote an existing chat into a project
• Generate pre-field artifacts — discussion guides, screeners, and research plans created by AI within the project context
• Conduct and synthesize — field notes, findings, and conclusions live in the project alongside the AI's reasoning
• Approve and share out — approvals tracked and distributed directly from the project, with full audit trail

Reframing Beta Scope Around Pre-Field Velocity

The most important design decision on this project wasn't a UI choice — it was the scoping reframe written into the Beta PRD. If we compressed pre-field from 3–4 weeks to 2–4 days, total lifecycle compression from 7–11 weeks to 3–5 weeks followed naturally.

• Four launch-blocking skills: Discovery, Planning & Design, Discussion Guide, Survey/Screener
• Three skills explicitly deferred: Quant Analysis, Qual Analysis, Report/Shareout
• Trade-offs written directly into the PRD so the whole team could hold them

Architecture as a Product Decision

Alpha's dependency on Mise recipes was framed by engineering as a latency problem. I reframed it as a product problem — the agent's capability ceiling was set by infrastructure choices, and that ceiling was capping what we could design for. Design has to know enough about the substrate to advocate for the right one.

• Search latency: Vector DB (3 min) → Laser KNN (300ms)
• Data coverage: GDocs 8% → 93% · GSlides 0% → 99.9%

Designing the Human Checkpoint Into Every Flow

"AI assists, not automates" had to do real work — not just sit in a deck. I worked with engineering and content design to translate the principle into concrete interaction patterns.

• Discussion Guide: agent generates, researcher edits — no "approve and ship" affordance
• Planning & Design: methodology recommendations include reasoning, alternatives, and an override path
• Compliance: auto-populates required forms; researcher reviews before submission
• Reasoning UI: visible retrieval, synthesis, and filtering steps at every turn

Shipped Metrics, and a Posture Shift That Matters More

Beta launched in May 2026. The qualitative outcome that matters more than any number: the research org has stopped framing AI as a threat to craft. Twenty-two Research Ambassadors across Meta drove Beta adoption inside their pillars.

The Data Layer Is Running Ahead of the UI

The data layer is the most strategically interesting part of this work — and it's already running ahead of the UI. In the Scale phase, PMs, designers, and engineers will be able to query it directly with appropriate guardrails.

The harder open question: whether researchers actually use the agent as a partner or revert to using it as a search box. The design principle is only as real as the behavior it produces.