WhatsApp as an onboarding channel: lessons from Toshify

Why WhatsApp and not something else

Toshify supports drivers wanting to join Cabify: account opening, document verification, app onboarding, activation. The initial brief mentioned a website. Our audit showed three things.

First, 95% of candidates are already-active drivers (taxi, ride-share, competitor platforms) who want a quick process between rides. They don’t open a browser on their phone — they open WhatsApp.

Second, the historical onboarding operator (3 FTEs on the phone) capped at 60% of candidates called back within 24 h. Not from incompetence, from time-zone constraints. An asynchronous channel solves that.

Third, the WhatsApp Business API is mature, GDPR-compliant, and usable in production without hacks.

The stack

We picked battle-tested bricks: official WhatsApp Business API (via Twilio), Node.js backend, n8n orchestration, Gemini LLM with a documentary knowledge base (strict RAG), and push to Intercom, Supabase and our internal tool for hand-off to the sales team.

No autonomous multi-step agent, no complex vector memory, no local model. The stack is intentionally simple — that’s what makes it maintainable. Details in our AI chatbots service.

Three critical guardrails against hallucinations

When your agent speaks on behalf of a brand (Toshify) that itself speaks on behalf of a third-party platform (Cabify), a single wrong answer can trigger a dispute. Three guardrails are systematic.

Guardrail 1 — Strict RAG. The agent only answers from the documentary base provided by Toshify (driver policy, Cabify terms, administrative steps). If the question falls outside, it hands off.

Guardrail 2 — Confidence threshold. The LLM produces a relevance probability for the answer. Below 0.85, automatic human hand-off. Better to hand off an easy case than to hallucinate.

Guardrail 3 — Anti-speculation system prompt. The prompt explicitly contains: “If you’re not sure, say ‘I’m forwarding your question to a supervisor’. Never guess.” Basic but critical prompt engineering.

Result: zero hallucinations reported by Toshify in 8 months in production. Not because the model is perfect, but because the guardrails keep it from wandering off-topic.

Smart hand-offs: when to step in

The classic chatbot trap: handle everything or hand everything off. The right design is in between.

We identified four hand-off triggers: (1) the question falls outside the scope covered by the knowledge base, (2) the candidate expresses frustration (detected via LLM classification), (3) the profile matches a premium or disputed case (CRM fields), (4) the candidate explicitly asks for a human.

At scale, 95% of candidates are handled fully autonomously, and 5% are handed off to the supervisor. The Toshify team runs a single supervisor in active oversight, against three FTEs before. Details in the Toshify case.

GDPR and compliance (don’t skip this)

WhatsApp Business API is, on paper, GDPR-compliant. But that doesn’t free you from a few basics.

Explicit opt-in: the candidate must have consented before you write to them. The candidate themselves opening the conversation (via QR code or link) suffices technically.

Data retention: we set a 12-month retention for conversations, 24 months for qualification data in the CRM. Beyond that, automatic deletion.

Hosting: transcripts and structured data stored in Europe (France). That matters beyond GDPR: it reassures security audits on the Cabify side.

Argentina: law 25.326 is largely aligned with GDPR. We had the solution validated by a local lawyer before going live.

What we’d do differently

First regret: we underestimated Spanish diversity. Argentine Spanish (Río de la Plata) has voseo and a distinct lexicon from Iberian or Colombian Spanish. Our first prompts sounded Iberian. We reworked the system prompt to integrate the local tone, but it would have been simpler to handle upfront.

Second regret: we’d have spent more time scoping hand-offs. Initially, we underestimated atypical cases (drivers who had had an incident on another platform, for instance). We added these rules along the way, but a half-day of extra audit at scoping would have saved 2–3 iterations.

Third regret: none. WhatsApp was the right channel, the stack was the right one, the guardrails held. That’s rare.