Esval Industrial is a Colombian industrial SME that manufactures plastic packaging for cosmetics brands, and most of its B2B pipeline starts on WhatsApp. Over three months in 2026 we built EVA, an AI sales assistant that turned WhatsApp B2B channel automation from a manual, office-hours inbox into an auditable pipeline. EVA knows the catalog, qualifies the lead, detects commitments and payment proofs, and hands the conversation to the right advisor through WhatsApp's interactive buttons. Every step is logged, and 95 settings are editable from the panel without a code deploy. The platform was formally delivered on August 17, 2026 and is operated by Esval's own team. Here is how we approached it; to see the screens, read the full case study.
WebMinds is an AI and software development company based in Atlanta, Georgia, that builds AI agents, custom software and business automation for companies in the United States and Latin America, including several WhatsApp AI assistants in production in Colombia.
Key facts
- Client: Esval Industrial, a manufacturer of plastic packaging for cosmetics brands in Colombia.
- Duration: 3 months, with formal v1.0 delivery on August 17, 2026 and continuous operation with WebMinds support since.
- AI per message: 6 AI analyses on every inbound message: the catalog-grounded reply, lead extraction, commitment, payment and frustration detection, and a shared classifier.
- Control: 95 settings editable from the panel, from personality and hours to credentials, advisors and legal data.
- Engineering: 49 versioned migrations, 61 API endpoints, 45 operations scripts, 3 audit tables and 27 screens. About 85% of the code is backend, AI pipelines, orchestration and audit.
- Compliance: public privacy, terms and data-deletion pages under Colombian Law 1581.
The problem: the number-one lead channel had no business tools
WhatsApp is Esval's main lead channel. Until 2026 it was handled by three people answering manually, one chat at a time, during business hours. After hours, leads went cold. There was no centralized history, no audit trail and no way to see who was handling what.
Four issues stood out:
- No business tooling. Advisors answered from their own phones, with no shared history, no reassignment and no view of workload.
- A deep technical catalog. Most questions were about collapsible tubes, bottles, caps, materials, minimum order quantities and barrier properties by material. Nobody knows the whole catalog by heart.
- No prior infrastructure. There was no Meta Business setup, no approved templates, no data-protection policy under Law 1581, no catalog embeddings and no panel.
- Human judgment was non-negotiable. Commitments, payment proofs, frustration and complete customer data had to reach a specific person, with context.
That made the channel valuable and, without structure, risky: a request that arrived after business hours waited for the next working day, and the lead went cold.
Why WhatsApp as a B2B channel needed more than a chatbot
A typical chatbot answers and stops there. A B2B packaging sale involves minimum quantities, material specifications, quotes, payments and named advisors. We needed:
- Catalog-grounded answers that cite real specifications, not general knowledge about plastics.
- Routing to a specific person, chosen by the customer or by customer type, with only that advisor notified.
- Detection of sensitive moments: commitments, payment proofs and frustration.
- Auditability: a record of every escalation, takeover and button tap.
- Business ownership: the marketing lead and advisors needed to change rules without asking an engineer.
That is the kind of system our AI & Automation practice builds, delivered as a full web platform through our Web & App Development service.
What we built: WhatsApp B2B channel automation with EVA
The platform lives at app.esval.com.co. EVA is not a chat widget on top of a model. The interface is the small part. Underneath are 49 migrations, 61 endpoints, six AI analyses per message, 45 versioned operations scripts, an hourly scheduled job and a panel with 95 settings.
EVA, the AI agent
Six AI analyses run on every inbound message: the catalog-grounded reply plus lead extraction and commitment, payment and frustration detection, supported by a shared classifier. Generation, vision and embeddings run on Gemini over the official WhatsApp Cloud API.
A catalog knowledge base
The product catalog, technical sheets and sales procedures are embedded in pgvector. This is retrieval-augmented generation (RAG): before answering, EVA retrieves the relevant passages and answers from them. Image-only PDFs are read with vision, and Esval updates the catalog on its own through CSV uploads.
The escalation router, with advisor choice
The customer picks their advisor with WhatsApp interactive buttons, or EVA routes by customer type. Only the chosen advisor is notified, by WhatsApp and email. There is a safety net: if the chosen advisor is removed before the customer taps the button, the system falls back to the next active one. If the customer types instead of tapping, EVA closes the prompt and keeps the conversation going.
Independent cooldowns
Payment, commitment and escalation alerts each have their own window, so an active conversation never floods an advisor's phone.
A B2B lead board
Five stages built for Esval's sales process (new, advisor requested, quote, won and lost), with automatic assignment and a personal view for each login.
Strictly idempotent
Meta resends a message if it does not get a fast 200 OK. Every inbound message is keyed by its WhatsApp ID, so a retry never generates a second AI reply.
Credentials checked before they are saved
Gemini, Meta and Resend keys are tested against the real services before saving, with clear messages for an expired key, exhausted credits or a quota limit.
Hourly jobs on GitHub Actions
The hosting plan only allows daily cron jobs, so inactive conversations are closed every hour by a GitHub Actions workflow that calls a secret-protected endpoint.
Audit, email and compliance
Every escalation, takeover and button tap is recorded. The platform sends six branded transactional email flows, publishes privacy, terms and data-deletion pages under Law 1581, and includes an embeddable web widget.
How we deliver: Hybrid AI Coding
A senior WebMinds engineer steers AI agents (Claude Code) for implementation, with QA on every change and a simulator that runs the exact production pipeline. Architecture and every call with the client stay human. The stack is Next.js and TypeScript, Supabase with pgvector, Gemini, the WhatsApp Cloud API, Resend, Vercel and GitHub Actions.
Is your B2B pipeline stuck in personal phones? We can audit your WhatsApp channel and show you what an auditable setup looks like. Get your free assessment.
Timeline
- May 2026: project confirmed, with Esval brand, proposal and data model.
- July 2026: catalog ingestion with vision OCR, branded email per alert type and legal pages under Law 1581.
- July 31, 2026: customer-driven advisor selection with WhatsApp interactive buttons.
- August 2026: self-service catalog upload and real-time unread badges.
- August 17, 2026: formal delivery of v1.0, with a user manual and panel-editable settings.
- Since then: in continuous operation with WebMinds support.
Results: a real B2B pipeline on WhatsApp
As documented in the Esval case study:
- A real B2B pipeline on WhatsApp. What used to be a manual, after-hours black hole is now a measurable pipeline: catalog-grounded replies, every lead on a board, every escalation logged and the right advisor brought in at the right moment.
- The business in the driver's seat. With 95 editable settings, Esval changes the assistant's personality, rules, cooldowns, credentials, advisors and legal data on its own. Engineering stays with the engineers; the business stays with the business.
- Auditable from day one. Every escalation attempt, manual takeover and button tap is recorded with its outcome per channel, and customer data is handled under Colombian data-protection law.
For other WhatsApp AI assistants we have built in Colombia, see the vehicle-catalog sales assistant in the Industrias Qatar case study and the IT consultancy assistant in the Netthplus case study. For the broader picture, see our sales automation use case.
FAQ
Can WhatsApp work as a serious B2B sales channel?
Yes, if it has business structure behind it: a shared history, a lead board, routing to named advisors and an audit trail. For Esval, EVA answers from the catalog, qualifies the lead and routes it, and every step is logged.
Can a WhatsApp AI agent hand a conversation to a specific advisor?
Yes. With EVA, the customer chooses an advisor using WhatsApp interactive buttons, or EVA routes by customer type. Only the chosen advisor is notified, by WhatsApp and email, and there is a fallback if that advisor is no longer active.
Can the business change the AI's behavior without developers?
At Esval, 95 settings are editable from the panel, including the assistant's personality, sales mode, hours, holidays, cooldowns, credentials, advisors, campaigns and legal data, with no code deploy.
How does the AI know the product catalog?
The catalog, technical sheets and sales procedures are embedded in a pgvector knowledge base. EVA retrieves the relevant passages before answering, reads image-only PDFs with vision and picks up CSV catalog updates uploaded by the team.
Does a WhatsApp AI agent comply with Colombian data-protection law?
For Esval we published privacy, terms and data-deletion pages under Law 1581, and customer data is handled under Colombian data-protection law. Every escalation and takeover is audited.
How long does a project like this take?
Esval's platform took three months from confirmation in May 2026 to formal delivery on August 17, 2026, starting with no prior Meta Business setup, templates or catalog embeddings.
Talk to us
If WhatsApp is where your B2B buyers start and you want a pipeline you can measure and audit, we can help. Get your free assessment, or book a call with our team. WebMinds is based in Atlanta and serves companies across the US and Latin America, in English and Spanish.