Defining the North Star AI vision for WhatsApp Business

Many AI capabilities, no unified vision

WhatsApp Business was at an inflection point: multiple AI capabilities were being built in parallel (a business-focused agent, a general assistant, several interface modalities), but there was no unified vision for how they should come together for small-business users.

Define what AI should be for small businesses

From the outset the goal wasn't to ship a feature — it was to set a North Star: an evidence-based answer to what AI should be for SMBs on WhatsApp, plus the framework to get the team there. Concretely, I set out to answer four questions:

• How do SMBs perceive and differentiate the business agent from the general AI assistant?
• Which modalities should AI take — traditional UI flows, voice agents, or messaging-based assistants?
• What jobs-to-be-done are highest priority for AI across the SMB workflow?
• Where does AI build trust vs. create anxiety for business owners?

Designing for a vision, not a feature

• Vision-level ambiguity: the "right answer" wasn't yes/no — it was a strategic framework to guide years of product development.
• Concept abstraction: testing future-state concepts (voice agents, avatar assistants) with SMBs who had no prior familiarity required careful stimulus design to avoid leading or confusing participants.
• Cross-market validity: Mexico and India have very different commerce patterns, digital maturity, and language needs — yet findings had to converge into a single recommendation.
• Multi-stakeholder alignment: results had to inform PM (roadmap), Design (modality), Engineering (infrastructure), and Leadership (investment) — each with different decision frames.
• Competing internal visions: different teams held different hypotheses; the research had to provide objective evidence to resolve debates, not confirm any one team's priors.

Comparative concept testing across modalities

Research & concept design

• Framed the study around two levels: immediate BizAI opportunities (improve now) and broader AI opportunities across WAB (build next).
• Designed concept stimuli for three modalities: a traditional wizard-guided UI, a voice/avatar agent, and a messaging assistant.
• Mapped jobs-to-be-done across the full SMB workflow: profile setup, catalog management, customer chat, ads/broadcast, CRM, and sales/support.
• Structured a comparative evaluation — reacting to multiple modalities for the same task — instead of testing concepts in isolation.

Execution & synthesis

• Conducted 16 generative interviews across the two markets, probing AI perceptions, trust, mental models, modality preference by task complexity, customization needs, and willingness to centralize into one assistant.
• Surfaced the core confusion: SMBs conflate the business agent with the general assistant at entry points.
• Mapped modality preference to task complexity: wizard flows for routine tasks; voice/avatar for complex "teaching moments"; users want to switch.
• Synthesized the strategic recommendation: one unified AI assistant with tiered customization > multiple separate bots — organized so each insight linked to a concrete monetization or retention pathway.

The interaction decisions I shaped

Beyond findings, this study made concrete decisions about how people would interact with the AI:

• Modality by task complexity: wizard-guided UI for routine setup, voice/avatar for complex "teaching moments," with smart defaults and the ability to switch — not a single-modality bet.
• Trust & mental models: a distinct, business-first agent identity to resolve the agent-vs-assistant confusion and reduce anxiety at the entry point.
• Human-in-the-loop boundary: the AI assists but never "closes the deal" — hand-holding that lowers risk over autonomy that raises it.
• Interaction architecture: one unified assistant with tiered customization (pre-trained default vs. custom) — a decision about the agent's behavioral model, not just a feature toggle.

What I learned

What worked: the two-tier recommendation structure made findings actionable across time horizons; cross-market convergence strengthened the recommendations; and framing every finding through business impact kept PM and leadership engaged.

What I'd do differently: add a quantitative validation phase to strengthen the "one assistant" recommendation; also test with experienced users (all participants were new to the agent), and coordinate earlier with adjacent research.

What I learned: how to design research that shapes long-term vision (not just evaluate features), the power of comparative concept testing, and how to translate qualitative insight into strategic frameworks leadership can act on.

What I owned

Design research that set product & design direction

This was generative design research and product strategy — the upstream design work that decides what a product should become before any screen is finalized. As design researcher and strategist, I designed the concept stimuli for three AI modalities, mapped jobs-to-be-done across the SMB workflow, ran a comparative evaluation, and synthesized the findings into a North Star framework and a set of design principles that now guide modality, branding, and monetization decisions for WhatsApp Business AI.