AI with WhatsApp: the channel your customers already use, with responses that reference your data
WhatsApp is not just a web chat in a different colour. It requires templates, 24-hour windows, opt-in, and well-designed human escalation.
The problem
WhatsApp is the preferred channel in Spain and Latin America for communicating with businesses. However, most manage it like a shared mobile phone: unassigned messages, delayed responses, and no connection to orders or schedules.
Cheap 'WhatsApp bots' are often rigid button trees that frustrate users. When they add generative AI without rules, they invent prices, confirm non-existent appointments, or respond outside business hours without notice.
The WhatsApp Business API has strict rules: templates for proactive messages, a 24-hour window for free responses, quality limits, and the risk of blocking if you spam. Ignoring this can jeopardise your business line.
Without integration with CRM, ERP, or calendar, every conversation starts from scratch. The customer repeats their ID, order, and problem. The agent wastes time, and the experience feels small-scale, even if you generate millions.
Scaling from 50 to 500 daily conversations with the same team is impossible without intelligent automation. However, poor automation —incorrect responses in such a personal channel— damages the brand more than not responding.
Customers perceive WhatsApp as a conversation with a person. A generic response or an endless menu breaks that expectation more than in email or web forms. The quality bar is higher, even if the message is automated.
Salespeople use personal WhatsApp to close deals —and you lose history, metrics, and control when they leave. Centralising in Business API with AI is not bureaucracy: it’s a commercial asset that remains in the company.
Organisational change matters: support, IT, and business must agree on what to automate and what requires human judgement. Without that agreement, the project generates internal friction even if the technology works.
In retail and hospitality, the customer sends a photo of the receipt, audio with the inquiry, or shared location. A button bot is not sufficient; a multimodal pipeline with validation before acting is needed.
Commercial conversion via WhatsApp is high when the response is immediate and personalised. Every hour of delay on a hot lead measurably reduces the likelihood of closing.
WhatsApp groups and broadcast lists have different rules than 1:1 conversations. Mixing use cases without design creates blockages and confusion for the customer.
RUMAZA does not sell licenses: we build a system that you can measure, maintain, and scale. If the core of the problem is not automatable with available data, we’ll tell you in the first meeting —saving months and budget.
Hours and holidays: the bot must know when to escalate to a human and what out-of-office message to leave without promising impossible deadlines.
Integration with payments and links: sending a payment link or secure follow-up requires domain validation and not exposing PII in long URLs.
Comparing three quotes without a common specification is useless: scope, integrations, and acceptance metrics must be identical to decide with criteria.
The total cost includes Meta conversation, API provider, hosting, template maintenance, and time for agents who continue to handle escalations —everything needs to be modelled.
Iteration with real data from the first fortnight in production: adjusting thresholds, prompts, and rules with customer metrics, not lab assumptions.
What is AI with WhatsApp (no fluff)
It is a system that receives messages via WhatsApp Business API (directly or with a provider like Twilio, 360dialog, MessageBird), interprets intent, queries your systems, and responds in text, buttons, or lists —within Meta's rules.
AI provides flexible understanding: the customer writes 'I want to change the delivery of order 8842' and the system extracts the number, queries the carrier, and responds with real options. You are not dependent on a menu of 'press 1, press 2'.
A healthy flow combines: webhook reception, message queue, AI engine with tools (orders, appointments, FAQs), business rules (hours, language, escalation), and a human tray for cases outside of trust.
Proactive messages —appointment reminders, shipping notifications— use approved templates. Responses within the 24-hour window can be conversational. Mixing both without design generates rejections and costs.
It is not WhatsApp Web with ChatGPT stuck on it. It is production architecture: retries, idempotency, logs, cost control per conversation, and GDPR compliance for data traversing the channel.
Types of messages you must design separately: reactive response within window, utility template (order status), marketing template (with explicit opt-in), and escalation message to human with internal summary.
AI interprets text, transcribed audio, and images (receipt photo, error capture). Each modality has different validation rules before acting.
Typical integration: Meta webhook → your backend → queue → AI worker → API response → log in CRM. Without that chain, the first traffic spike can crash the service or duplicate messages.
Gradual deployment: pilot with one channel or one type of inquiry, measurement for two weeks, expansion based on data —not a big bang that saturates the team and the customer.
Consent and opt-in: recording when the user accepted communications, message category, and opt-out option. Without traceability, audits and compliance fail.
Analytics by conversation: source (ad, QR, web), detected intent, resolution, time, and cost. This way, you optimise templates and flows with data, not intuition.
A/B testing of messages: two variants of reminder template measure no-show and cost. Continuous optimisation like in marketing, applied to operations.
RUMAZA criteria: specific problem, accessible data, success metric, and closed scope. Without these four pillars, there is no project —only an experiment that profits the consultant and costs the client.
Rate limiting per user prevents abuse and costs if someone spams the bot or a competitor stress-tests the line.
Bidirectional synchronisation with CRM: incoming message creates or updates contact; human agent sees the same record as in email.
Evolutionary maintenance —new intents, providers, languages— is budgeted separately from the MVP to avoid surprises or zombie projects.
Testing with real users before go-live: 20 recorded and reviewed conversations prevent surprises on the public launch day.
Post-launch support with direct channel and agreed SLA: critical incidents during business hours resolved the same day —no eternal tickets.
We document assumptions, known limits, and expansion plan in the delivery —total transparency about what the system does today and what remains for a phase two if the numbers justify it.
When it makes sense
- More than 30% of your inquiries already come via WhatsApp —with volume and data to justify it.
- You need responses outside business hours without losing leads —with volume and data to justify it.
- You have repeatable processes: appointments, orders, status updates —with volume and data to justify it.
- You want to unify WhatsApp with CRM instead of using loose mobiles —with volume and data to justify it.
- The volume exceeds what an agent can handle with quality —with volume and data to justify it.
- You can obtain opt-in and comply with WhatsApp Business policies —with volume and data to justify it.
What can be built
Order and shipping inquiries
Customer identifies order; the bot queries e-commerce and responds with tracking. Sends proactive template when the package ships. Includes logs, trust thresholds, and human review in the initial phase until metrics are calibrated in production.
Appointment scheduling
Understands preferred date, checks calendar, proposes slots, and confirms with a reminder 24 hours in advance via template. Includes logs, trust thresholds, and human review in the initial phase until metrics are calibrated in production.
Lead capture and qualification
Guided questions, basic scoring, and automatic creation in CRM. Escalates to sales if the lead meets criteria. Includes logs, trust thresholds, and human review in the initial phase until metrics are calibrated in production.
Support with human handoff
AI resolves FAQs; if it detects frustration or a sensitive topic, it passes the conversation to an agent with a summary and data already collected. Includes logs, trust thresholds, and human review in the initial phase until metrics are calibrated in production.
How RUMAZA would build it
Possible technologies
- WhatsApp Business API
- Twilio / 360dialog / Meta Cloud API
- Python / Node.js
- OpenAI / Anthropic
- Redis / RabbitMQ
- PostgreSQL
- CRM and e-commerce via REST
Application scenarios
Sales or reception handle personal WhatsApp
Customer inquiries mixed with private messages. Business/API channel, applicable templates, and connection with stock, appointments, or CRM are advisable.
Manual confirmations and reminders
Clinic, academy, or service with appointments confirmed by phone or manually. Reminder and rescheduling can be automated while respecting opt-in and hours.
Inquiries outside business hours
Customers write at night or on weekends, and no one responds until Monday. An automated flow can provide basic information or collect data for the team.
Common mistakes
- Using personal WhatsApp or Business App for volume that requires API
- Sending proactive messages without approved templates
- Not informing that it is an automated assistant when applicable
- AI without limits that confirms irreversible actions (orders, payments)
- Ignoring the 24-hour window and losing conversations
- Not measuring conversation quality: just counting messages sent
- Not reviewing the project after 90 days with real metrics and adjusting or closing what does not contribute.
Frequently asked questions
Do I need WhatsApp Business API or is the app enough?
The app is suitable for a few manual chats. From ~50 conversations/day or if you want AI and integrations, you need API with an approved provider. We define this in scope based on your systems, volume, and legal constraints —without promising generic figures.
How much does it cost per conversation?
Meta charges by conversation category (marketing, utility, service). The provider adds a margin. We calculate expected cost in the design. We define this in scope based on your systems, volume, and legal constraints —without promising generic figures.
Can AI send audios or images?
It can receive and process images (OCR, classification). Sending media depends on the flow and templates. We define this on a case-by-case basis. We define this in scope based on your systems, volume, and legal constraints —without promising generic figures.
How does the conversation pass to a human?
Triggered by a keyword, low model confidence, or user request. The agent sees the complete history in the panel. We define this in scope based on your systems, volume, and legal constraints —without promising generic figures.
Is it legal in Europe with GDPR?
Yes, with legal basis, user information, DPA with provider, and data minimisation. We do not send complete history to the model without control. We define this in scope based on your systems, volume, and legal constraints —without promising generic figures.
Does it work with multiple agents?
Yes. Assignment queue, states (open, pending, closed), and synchronisation with CRM. We define this in scope based on your systems, volume, and legal constraints —without promising generic figures.
Related guides
Is WhatsApp overwhelming in your company?
Tell me the volume, what they ask, and what systems you have. I will propose flow, API, and closed scope.