The WhatsApp Test

Caribou recently published a report on how millions of entrepreneurs and self-employed people rely on WhatsApp to earn a living. This blog builds on conversations that co-authors had about “what comes next” as WhatsApp becomes not only a place to talk to friends and strangers but also to interact with the world’s knowledge (via generative AI), and perhaps to complete tasks (via agentic AI).

—

Amina runs a small clothing shop in Ibadan, Nigeria. She never finished secondary school, but she has built a steady business selling ankara fabric and ready-made garments to customers in her neighborhood. Like millions of entrepreneurs around the world, WhatsApp is her primary business communications tool. She takes orders through messages, coordinates with suppliers in Lagos, and posts photos of new stock to customer groups.

Amina is a persona drawn from our research with entrepreneurs across India, Nigeria, Kenya, and Pakistan, including 7,800 surveys across two waves of fieldwork, 40+ in-depth interviews, and self-documented video diaries from eight micro-entrepreneurs. 

The promise of Agentic AI is that Amina will soon ask her phone: “Which of my three suppliers has the best price on this fabric right now?” An AI agent searches her WhatsApp conversations with all three suppliers, compares their latest quotes, and tells her. Or: “Remind me to reorder when I’m down to five units of this style.” The agent monitors her inventory messages and prompts her at the right time. Or: a customer messages at 2am asking about sizes available. The agent responds accurately based on Amina’s previous messages, without waking her.

Unlike generative chatbots that respond to prompts, agentic AI systems can search, compare, schedule, transact, and coordinate. They represent AI moving from answering questions to completing tasks.

There are hints of this future in what is already deployed: the WhatsApp Business API supports automated responses and basic chatbot interactions. It’s early days for true agentic AI that acts autonomously across contexts, and these systems are largely untested in the context of micro- and small enterprises. 

Some will argue that Amina, and millions of people like her, need basic infrastructure (like roads, reliable electricity, and more affordable data), not AI. They’re not wrong. But the enthusiasm around agentic AI is spilling into use cases like Amina’s. To encourage uptake and effective use, Designers, platform owners, and policy makers should apply principles that match her needs and her context. That context may be WhatsApp.

The chat thread is the lingua franca of the WhatsApp experience. It’s easy to imagine that agentic AI could support many core business tasks via chat, letting entrepreneurs focus on relationships and growth. A business run via WhatsApp imposes real constraints, and Amina faces most of them at once. 

Amina’s challenges aren’t edge cases. They’re closer to default conditions for the global majority. Yet most early agentic AI systems are gaining traction among groups with more resources. They assume reliable connectivity, powerful devices, text literacy, cheap data, and open platforms. The inclusive digital economies community risks pinning our hopes and the success of Amina’s business on systems that work in demos but fail in Ibadan.

Richard Heeks documented this pattern as the design-reality gap in e-government projects: when designers assume resources users do not have, or over-prescribe how the systems are to be used, systems fail regardless of their sophistication or the good intentions of all involved. Agentic AI risks repeating these failures and at a greater scale and speed.

But the design-reality gap is closeable. If designers treat WhatsApp’s constraints as parameters rather than problems, they can build systems that work within real conditions. Five design principles, drawing on the Principles for Digital Development, applied to agentic AI in resource-constrained contexts, can help close the gap.

The Principles for Digital Development remind us to design for inclusion. Enable users to interact with agents through voice messages, not just text. Speech recognition and synthesis should work reliably in multiple languages and accents, and handle the conversational style of voice messages: pauses, corrections, and informal phrasing. For entrepreneurs like Amina, voice is the native mode of communication, and any system requiring precise text prompts may exclude her, regardless of its sophistication.

When an agent chooses a supplier, schedules a task, or sends a message, users should understand why. Show exactly how much data each interaction consumes, allow users to set limits and control background activity, and provide clear value metrics: “This saved you 30 minutes today” or “This comparison saved you 500 naira.” Transparency builds trust and enables informed decisions.

Train agents on local business terminology, pricing conventions, and negotiation styles. This reflects two Principles: to design with people and to understand the existing ecosystem, building for how business actually works in specific contexts, not imposing external models. A system designed for Western retail won’t understand how Amina negotiates with suppliers or manages customer credit. 

Build systems that can queue tasks, cache critical information, and sync intelligently when connectivity returns. Rather than requiring constant cloud access, agentic AI should handle common tasks locally and synchronize results when possible. Minimize data usage and computational demands to be kind to users’ wallets and device limitations. The Principles encourage these, building for sustainability.

People thrive in digital spaces where they feel protected. Automated systems should reinforce safety, rather than introduce new uncertainty. Users want to be able to distinguish legitimate agent behavior from malicious impersonation, with clear accountability when agents make harmful decisions. To do so, the digital systems built for them must anticipate and mitigate harms. 

Yet businesses of every size depend on being discoverable. Recommendation and routing patterns should not sideline small enterprises. If agentic systems begin to shape who gets seen by surfacing some businesses while burying others, visibility mechanisms must work for small-scale entrepreneurs, not just retail giants.

WhatsApp is thus a critical proving ground. If agentic systems can work here, under these constraints, it would signal genuine progress toward practical AI for the global majority. However, working within WhatsApp’s constraints means working within Meta’s control. Platform power and proprietary APIs shape what’s possible. The future of messaging software is path-dependent and shapable, and paradigm shifts like agentic AI present moments when we can act with intention and clarity to influence that path. There are different possible versions of WhatsApp’s future. Better versions likely open APIs to enable third-party agents, maintain a commitment to end-to-end encryption, and/or design for including entrepreneurs like Amina from the start rather than as an afterthought. Worse versions centralize agentic capabilities in Meta’s own AI, extract more data, and/or optimize for wealthy markets while treating the global majority as externality.

Through our conversations about the research and “what’s next,” we’ve come to see WhatsApp as a threshold test for agentic AI. The micro-enterprises we studied would be more likely to use agentic AI that works on a Tuesday afternoon when the power is out, data is expensive, and three suppliers are waiting. If agentic AI can meet that standard—operating under connectivity constraints, on low-end devices, with voice-first interfaces—it can serve the global majority. If it can’t, agentic AI may remain yet another tool for the already-privileged. Whether designers, platform owners, and policymakers make the choices necessary to cross that threshold will be the next real test.