Netthplus is a Colombian IT consultancy covering SAP, cybersecurity, cloud, DevOps, RPA and IT talent, with a wide portfolio and a small sales team. In July 2026 we launched an AI commercial assistant on WhatsApp for B2B sales. It handles first contact in Spanish, grounds every answer in Netthplus's own technical documents, qualifies the lead and alerts a human advisor the moment the conversation calls for it. In production it replies in 6.1 seconds on average, around the clock. This article explains how it works and why we built it the way we did. To see the screens, read the full case study.
We are WebMinds, an AI and software development company based in Atlanta, Georgia, working with companies in the United States and Latin America. Netthplus runs on our WhatsApp AI platform, white-labelled for them at app.netthplus.com, with code and domain owned by the client.
Key facts
- Client: Netthplus, an IT consultancy in Colombia (SAP FI/CO and S/4HANA, cybersecurity, cloud, DevOps, RPA, AI analytics and IT staffing).
- Speed: 6.1 s average reply time end to end in production, including retrieval and four AI detectors; 14 s at p95.
- Knowledge: 115 knowledge chunks from 11 Netthplus documents, all indexed in pgvector.
- Handoff: 3 escalation triggers (model tag, keyword safety net and commitment detector), alerting the advisor by WhatsApp and email in parallel.
- Security: 12/12 tables under row-level security with 26 policies; server-side keys never reach the browser.
- Live since: July 8, 2026, in continuous operation since.
The problem: a broad technical portfolio and a small sales team
Netthplus sells complex services. A prospect may ask about SAP modules, consultant profiles, certifications, a cybersecurity engagement or IT staffing in the same week. Before the project, every inbound WhatsApp enquiry had to be typed by hand, and nights and weekends quietly lost pipeline to competitors.
The challenges were concrete:
- Repetitive questions. Most inbound questions were about services, SAP modules, consultant profiles and certifications, and each one was answered individually.
- No coverage outside office hours. Prospects writing at night or on weekends went elsewhere.
- Scattered history. Each advisor kept chats on their own phone, with no consolidated leads and no organizational memory.
- No single channel for marketing. Profiles and business cards pointed to whoever was on duty, so campaigns had no consistent destination.
For B2B companies in Colombia and Latin America, WhatsApp is often the first place a prospect writes, even for technical services. The question is not whether to be there, but how to answer well, fast and with a record.
Why a generic chatbot would not work for B2B IT services
A scripted bot can answer "what are your hours?" It cannot answer a detailed question about SAP S/4HANA with information that matches what the consultancy actually offers. For Netthplus, we needed:
- Answers grounded in real documents. The assistant must cite the actual portfolio, not general knowledge about SAP or cloud.
- Handoff that does not fail. A B2B lead that asks for a meeting is worth more than any automated answer. Escalation had to survive edge cases, including prospects who stop replying halfway.
- Lead capture. Every conversation should produce a structured record a sales team can work.
- Compliance. When the assistant collects personal data, it must cite a data-policy link under Colombian Law 1581.
- Resilience to model changes. LLM providers release and promote models often, and a sales channel cannot go silent when that happens.
That combination is what our AI & Automation practice builds, together with the messaging and email integrations that connect the assistant to the people who close.
What we built: an AI commercial assistant on WhatsApp
Behind every reply runs a retrieval pipeline, LLM generation, four secondary AI detectors, multichannel notifications and row-level security on every table. Roughly 65% of the effort is backend. We delivered five systems.
The WhatsApp AI agent
A Meta Cloud API webhook with signature verification receives each message. Generation runs on Gemini, grounded in retrieval. The assistant understands voice notes and customer images and replies in Colombian Spanish.
A document knowledge base with hybrid search
Netthplus's PDFs and office files are split into chunks, embedded at 768 dimensions and served from pgvector. Search is hybrid: full-text plus semantic. That matters for technical vocabulary, where an exact term like a module code should match even when semantic similarity alone would miss it. Scanned PDFs are read with vision. Today the base holds 115 chunks from 11 documents.
Human escalation with three independent triggers
Three triggers can bring in a person: a tag the model emits, a keyword safety net, and a commitment detector. When any of them fires, the sales advisor is notified in parallel by an approved WhatsApp template and by email, with the latest messages attached.
Escalation that survives abandonment
The assistant flags the lead both when it asks for contact details and when it confirms them, with a six-hour cooldown. A prospect who stops halfway through still reaches an advisor.
Lead extraction and a lightweight CRM
A second AI pass pulls name, email, interest, estimated value and status from each conversation into a lead board. Source channel and notes come along, so sales can see where interest is coming from.
Answer Meta first, think later
The webhook returns 200 OK to Meta in under two seconds and keeps processing in the background. Slow model calls never trigger Meta retries or leave a prospect waiting.
Survives the next model release
When a newly promoted Flash model saturated for days, a pinned stable version with a cascading fallback kept the assistant answering: retries per model, then a fallback chain, with every recovery logged. We also found that reasoning tokens count against the output budget, which cut replies mid-word. A larger budget and logging of every non-normal finish solved it.
Operations dashboard and social entry points
The dashboard has 21 screens: a real-time inbox, a training studio with a chat simulator that runs on the production pipeline, analytics, and a wa.me link and QR generator per channel. Every campaign, bio, email signature and business card points to the same assistant, with a prefilled message that shows where each lead came from.
Compliance the client controls
The data-policy link the assistant cites when it collects personal data under Colombian Law 1581 is editable from the panel, so Netthplus can update it without a deployment.
How we deliver: Hybrid AI Coding
Senior WebMinds engineering steers AI agents for implementation, with independent QA on every change. Architecture, product judgment and client conversations stay human, and every production finding becomes a documented fix and a reusable diagnostic script. The stack is strict TypeScript and Next.js, Supabase (Postgres, Auth, Realtime) with pgvector, Gemini, the WhatsApp Cloud API, Resend, Tailwind CSS and Vercel.
Losing B2B enquiries after hours? Let us look at your WhatsApp channel and your documents. Get your free assessment.
Timeline
- March–June 2026: platform foundation, including RAG, leads, catalog media, simulator, web widget, social links and hybrid search.
- July 8, 2026: launch for Netthplus with Meta compliance, Netthplus branding, legal pages and the first real conversation.
- July 2026: two-turn escalation, keyword safety net and a template-language fix after early feedback.
- August 21, 2026: fast acknowledgement to Meta and a cascading model fallback.
- September 2026: client-editable data-policy link and a help guide for sharing the assistant on social media.
- Ongoing: continuous operation and support.
Results: from hours to seconds
The Netthplus case study documents three outcomes:
- From hours to seconds. Enquiries that used to wait in a manual inbox now get an answer in about six seconds on average, around the clock, grounded in Netthplus's own documents. People step in when the assistant escalates, alerted in parallel by WhatsApp and email.
- One channel for all of marketing. wa.me links and downloadable QR codes per channel (Instagram, Facebook, email signatures, business cards) give marketing one assistant to point every campaign at, with a prefilled message that shows where each lead came from.
- Knowledge and pipeline, consolidated. The portfolio lives in one searchable knowledge base, and conversations and extracted leads live in one secured database: auditable, exportable and ready for their CRM.
We also built WhatsApp AI sales assistants for a Colombian manufacturer of tow hitches and bike racks in the Industrias Qatar case study and for a plastic packaging manufacturer's B2B pipeline in the Esval case study. For more on how we work with consultancies and service firms, see professional services.
FAQ
What is an AI commercial assistant on WhatsApp?
It is an AI agent on the official WhatsApp Cloud API that handles first contact with prospects, answers from the company's own documents, qualifies the lead and hands the conversation to a salesperson when needed. For Netthplus, it answers questions about SAP, cybersecurity, cloud and IT staffing.
Can a WhatsApp AI assistant answer technical B2B questions accurately?
It can when its answers are grounded in the company's documents. Netthplus's assistant searches 115 chunks from 11 of its own documents with hybrid full-text and semantic search before replying.
How fast does a WhatsApp AI assistant reply?
In production, Netthplus's assistant replies in 6.1 seconds on average end to end, including retrieval and four AI detectors, with a p95 of 14 seconds.
Can the AI hand a prospect to a human salesperson?
Yes. Three independent triggers (a model tag, a keyword safety net and a commitment detector) notify the advisor by WhatsApp template and email in parallel, with the latest messages attached.
What happens if the AI model has an outage or a bad release?
A pinned stable model version with a cascading fallback keeps the assistant answering. Retries happen per model, then through a fallback chain, and every recovery is logged.
Who owns the code and the data?
For Netthplus, the platform is white-labelled at app.netthplus.com with code and domain owned by the client, and conversations and leads live in a secured database under row-level security.
Talk to us
If your team sells complex services and your prospects already write on WhatsApp, we can help you answer them in seconds without losing the human touch where it counts. Get your free assessment, or book a call. WebMinds is an AI and software development company based in Atlanta, Georgia, serving the US and Latin America in English and Spanish.