Implementing an AI agent on WhatsApp for a company requires 5 concrete steps: enable WhatsApp Business API, design the conversation, connect an LLM, integrate with your systems (CRM/ERP) and launch with monitoring. Total: 4-6 weeks for a functional MVP. This guide explains each step with the technical decisions you must make.
Applies to SMBs and mid-size companies in Colombia that want to automate WhatsApp support without replacing their human team — instead, freeing them from repetitive tasks.
Step 1 — Enable WhatsApp Business API
The official WhatsApp API (NOT the mobile app) is a prerequisite for automation. 3 ways to enable it:
- Meta Cloud API (direct with Meta, free service but charges per conversation): cheapest, requires technical team for initial integration.
- Twilio: popular provider, friendly UI, ~15-20% margin over Meta rate.
- 360dialog: LATAM-focused, strong Spanish support.
Meta typical cost: $0.025-0.08 USD per initiated conversation (country-based). Colombia: ~$0.04 USD per utility conversation.
Step 2 — Design the conversation (prompts and intents)
70% of an AI agent's success comes from the prompt and conversation flow, not the model. Prepare:
- Intent list: typical 8-15 intents to start (query product, place order, book appointment, support, etc.).
- Per-intent flows: what to ask, what to validate, when to escalate to human.
- System prompt (agent persona): tone, language, what it can/cannot answer, how to handle complaints.
- Fallbacks: what to say when it doesn't understand. Key to not losing clients.
Step 3 — Pick and connect the LLM
Top 4 production enterprise models in 2026:
- Claude Sonnet 4.5 (Anthropic): best at instruction following, excellent Spanish, medium cost. Recommended for rule-driven B2B agents.
- GPT-4o (OpenAI): best known, wide support, good latency.
- Gemini 2.5 Pro (Google): price-competitive, strong Workspace integration.
- Open-source (Llama 3.3, Mistral): on-premise, zero variable cost, but requires own GPU and MLOps expertise.
For SMBs we recommend starting with Claude Sonnet 4.5 or GPT-4o-mini (10x cheaper than full GPT-4o) until validating volume and requirements.
Step 4 — Integrate with CRM/ERP
An agent that only replies text is a FAQ chatbot. A real AI agent creates, reads and updates systems:
- CRM (HubSpot, Zoho, Bitrix24): create leads, update opportunities, read client history.
- ERP (Odoo, Dynamics, SAP): query inventory, create orders, check pricing.
- Calendar (Google Calendar, Outlook): book appointments with real availability.
- Payment gateway (Wompi, PayU, Stripe): generate payment links.
Modern architecture uses Model Context Protocol (MCP) or native LLM function calling. Each integration = 1-3 dev days if the system has a REST API.
Step 5 — Launch with monitoring
An AI agent without monitoring degrades in weeks. Before launch implement:
- Conversation dashboard: see all interactions, filter by intent and outcome.
- Human intervention alerts: when agent doesn't know, auto-escalate to operator.
- Quality metrics: resolution rate, NPS, response time, cost per conversation.
- Auditable logs: every agent decision logged for debugging and compliance.
Typical 6-week timeline
| Week | Milestone |
|---|---|
| 1 | Discovery, API+provider choice, systems access |
| 2 | Conversation design + prompts + flows |
| 3-4 | Backend development + integrations |
| 5 | Internal testing + prompt tuning |
| 6 | Pilot with 10-20% real traffic + monitoring |
Typical pitfalls to avoid
- Launching without human handoff configured.
- Not measuring monthly LLM cost before scaling volume.
- Using generic prompt without company-specific context.
- Not versioning prompts (every change can break behavior).
- Trusting "model improves on its own": it doesn't, without active monitoring.
Next step
If you want to implement an AI agent on WhatsApp for your company, schedule a free assessment with Expansión TI. We have a proprietary product (Klerk Agent) that accelerates time-to-market to 2-3 weeks instead of 6. See also the service detail at AI Implementation.