AI WhatsApp Bot in 15 days: what's realistic, real architecture and pricing (2026)

The problem we hear over and over: appointments and orders are lost because nobody answers WhatsApp outside business hours, and half the workday goes to answering the same 8 questions — hours, prices, location, availability. The solution everyone asks for sounds magic: an AI WhatsApp bot that understands the customer, books in the calendar and alerts a human when needed.

At SprintMarkt we solved this problem from the inside: we built an autonomous AI salesperson over WhatsApp for our own sales CRM — official Meta Cloud API integration, unified conversation inbox, Claude-powered replies and human handoff. This post contains no invented client case and no magic ROI: it has the exact architecture we use, the Meta rules nobody tells you about before you start, a realistic 15-day plan and price ranges if you want to commission it.

Technical architecture (no fluff): two integration paths. The one we use: Meta's official Cloud API directly — a webhook (FastAPI in our case) receives every inbound message, Claude classifies intent and drafts the reply, and the conversation is stored in the CRM; no middlemen or platform fee, you only pay Meta's per-message/conversation cost (cents, per their current pricing). The alternative: a BSP like 360dialog or Twilio (€25-50/mo) that handles registration if you don't want to deal with Meta Business. Rest of the stack: Claude (Haiku for intent classification — cheap and fast —, Sonnet for complex replies), an indexed knowledge base (Supabase pgvector is plenty for dozens of documents) and your calendar or CRM via API. Everything fits on a €5-10/mo VPS.

The Meta rules that shape the WHOLE project (we learned these operating ours): (1) The 24-hour window — you can only reply with free-form text within 24h of the customer's last message; after that, only Meta-pre-approved templates. Design flows assuming this from day 1. (2) New numbers start with a limit of ~250 business-initiated conversations per 24h, which rises in tiers as Meta trusts you; business verification in Meta Business speeds it up. (3) If Meta detects spam patterns — many identical messages in a row, users blocking you — it throttles your sends even within your limit. The bot is for ATTENDING conversations; in Spain, cold first contact via WhatsApp is also prohibited by the LSSI (art. 21).

Day 1-5 — Discovery, flows and number registration: map the business's 10-12 most frequent conversations (new appointment, change, cancellation, prices, hours, location, urgent cases, payment...) and reduce them to 4 main intents with a simple decision tree. Document prices and policies in a living document — that becomes the RAG knowledge base. And request the number registration on day ONE, not day ten: Meta verification can take 24-48h and it's the only step that doesn't depend on you. Close the week with the webhook connected and a "hello world" replying.

Day 6-10 — RAG, integrations and the prompt: index the knowledge base (pgvector is enough), connect calendar/CRM via API with create/cancel/find functions. The system prompt carries the business personality — and the secret is not abstract instructions but real few-shot examples of how your best employee replies. Plus the golden anti-hallucination rule, written literally into the prompt: "if you can't find the data in the base, do NOT invent it: route to a human". We reinforce it with an automatic post-response check.

Day 11-13 — Human handoff (the CRITICAL part): auto-reply limit (if the bot loses the thread twice, hand off), a keyword to request a human ("I want to talk to someone" = immediate handoff), and a push notification to the team with the whole conversation in view. In our CRM the bot ships switched off behind a feature flag: first the unified inbox worked with humans replying, then the AI gets turned on. We recommend the same to any business — silent mode before hero mode.

Day 14-15 — Controlled testing + minimum metrics: test with a small group of real customers who know the business, collect every complaint (tone, local terms, uncovered cases) and fix it in the prompt or the base — each fix takes ~30 minutes. Before go-live, set up a minimal dashboard: conversations/day, % auto-resolved, human handoffs, average response time. Without those 4 metrics you won't know whether the bot helps or hurts.

What does NOT fit in 15 days (don't let anyone sell you otherwise): fine-tuning models with your data, deep ERP integrations, polished multi-tenant for several businesses, and mass outbound campaigns — which additionally require approved templates, Meta tier limits and, in Spain, prior consent. A good customer-service bot in 15 days is realistic; a "digital employee that does everything" is not.

Indicative pricing if you commission it from us (SprintMarkt 2026 rates): standard clinic/practice (1 number, 1 calendar, ~30 intents): €2,500-3,500 setup + €150-220/mo. Legal practice / consultancy (more legal nodes, escalation to a partner): €3,500-4,500 + €200/mo. Restaurant with bookings and dynamic menu: €2,500-3,200 + €130/mo. Real estate with catalog + lead qualification: €4,000-5,500 + €250/mo. Includes full setup, 1 month of post-launch support and 2 prompt-tuning sessions in the first 3 months.