AI models for chatbots define how a bot understands each message and chooses its reply. In 2026, AI customer service combines language models, a knowledge base, and tools, connected through the WhatsApp Business API and with handoff to humans.
This guide compares chatbot types by model, shows how to choose a model for WhatsApp, and covers Meta’s rules for AI, 2026 costs, ongoing evaluation, and data protection.
Key Takeaways
- Rules, NLP with intents, LLMs, RAG, and agents with tools solve different problems and usually work together in the same chatbot.
- Choosing a model for WhatsApp weighs cost per use, latency, quality in the customer’s language, privacy, and control over made-up answers.
- Meta’s terms prohibit using conversation data to train third-party AI models and require a clear path to a human agent.
- From October 1, 2026, chatbot replies above the 1,000 free service messages per phone number become billable by Meta.
What Are AI Models for Chatbots?
Every chatbot has a channel, a conversation logic, and a model that interprets what the customer writes. Separating these layers avoids confusion when comparing solutions, quotes, and vendor promises.
Model, Orchestration, and Channel
The model is the component that classifies or generates text. Orchestration decides when to query a knowledge base, call a system, or pass the conversation to a person. The channel is where the customer is, such as the website or WhatsApp.
The AI models used in customer service also include speech and image recognition models, which transcribe voice notes and read photos sent by the customer. Even so, text remains at the center of the conversation.
Changing the model does not change the channel, and changing the channel does not require another model. So the useful question is not just which AI to use, but how the three layers fit into the company’s AI customer service on WhatsApp.
What Has Changed in Recent Years
Until recently, most business chatbots followed menus and fixed flows, sometimes with an intent-recognition layer. Generative language models made it possible to answer free-form sentences without anticipating every variation.
The next step is agents, which, beyond conversing, query inventory, calendars, or the CRM and record the result. When they specialize in one industry, they are called vertical AI agents, and on WhatsApp they are now subject to specific Meta rules.
Chatbot Types by AI Model: Rules, NLP, and LLMs
The first three types of AI models form the foundation of most chatbot projects. Each one responds differently, and the difference shows with the first customer who goes off script.
Rule-Based and Decision-Tree Chatbots
A rule-based chatbot follows a flow designed in advance: a numbered menu, buttons, and fixed answers. There is no learning, because someone wrote every path, which makes the behavior predictable and easy to audit.
It works well for triage, invoice copies, and opening hours. On WhatsApp, reply buttons, lists, and WhatsApp Flows make this model friendlier, because the customer taps options instead of typing numbers.
NLP with Intents and Entities
Classic natural language processing (NLP) classifies a sentence into an intent, such as “track order,” and extracts entities, such as the order number. The answer is still written by the business, and the model only decides which one to use.
This was the approach behind many natural-language bots built before generative AI. It requires example phrases for each intent and constant maintenance, because new questions fall into “I didn’t understand” until someone adds them.
Generative Language Models (LLMs)
Large language models (LLMs) generate the answer from instructions and the conversation history. They understand typos, slang, and compound questions without every variation having to be registered.
The most widely used are language models from large technology companies, accessed through an API, and there are also open models that run on the company’s own servers. The choice between them depends on cost, control over data, and the team’s capacity to maintain infrastructure.
The price of this flexibility is hallucination: the model may invent a lead time or a rule with complete confidence. That is why LLMs in customer service rarely work alone, as the next section shows.
RAG and Agents: Models That Query Data and Take Actions
RAG and agents do not replace LLMs. They give the model access to reliable information and to the company’s systems, and that is where the chatbot stops just talking and starts solving.
RAG: Answers Grounded in a Knowledge Base
Retrieval-augmented generation (RAG) searches for relevant passages in company documents, such as policies, catalogs, and FAQs, and passes them to the model before it answers. The technique was described in a 2020 research paper.
Quality depends more on the knowledge base than on the model. Outdated, duplicate, or contradictory documents produce wrong answers that look right, so curating the knowledge base is an ongoing task with a defined owner.
Agents with Tools
An agent receives a list of tools, each with a name, description, and parameters, and decides when to call them. In the open Model Context Protocol standard, tools let the model query databases and call APIs.
In practice, the agent checks the calendar, issues an invoice copy, or updates a customer record, as the guide to AI agent CRM integration details. The same specification recommends keeping a human in the loop, with the ability to deny tool calls.
Comparing AI Models for Chatbots
The table sums up what each type delivers and where it tends to fail. The most stable projects combine several of them in the same flow.
| Model | How It Responds | Strength | Limitation |
|---|---|---|---|
| Rules and decision tree | Fixed flow with buttons | Predictable and inexpensive | Stalls off script |
| NLP with intents | Classifies and uses a ready answer | Full control of the text | Intent maintenance |
| Generative LLM | Generates free text | Understands varied phrasing | Can make up information |
| RAG | Generates from knowledge base passages | Answers grounded in documents | Depends on curation |
| Agent with tools | Queries and acts in systems | Resolves the request | Requires permissions and auditing |
A common hybrid design uses buttons for repetitive tasks, RAG for questions, and tools for actions, always with an agent available for cases the AI should not decide.
How to Choose an AI Model for WhatsApp Customer Service
There is no best model in the abstract. The choice depends on volume, the type of question, and the risk of each wrong answer. Choosing the platform is a separate step, covered in the guide to WhatsApp AI tools.
Cost per Use and Latency
Models accessed through an API usually charge by tokens, the chunks of text that go in and come out. Long histories and large knowledge bases in the context make each answer more expensive, so summarizing the conversation and retrieving only useful passages lowers the bill.
On WhatsApp, customers expect a conversational pace. Larger models tend to be slower, and a common strategy is to use a smaller model to classify and a larger one only for complex questions. Measuring real response time matters more than the spec sheet.
Quality in the Customer’s Language and Tone Control
Testing must happen in the language customers actually write in, such as Brazilian Portuguese or American English, and include WhatsApp abbreviations, transcribed voice notes, product names, and regional expressions. A set of real questions with expected answers makes it possible to compare models fairly.
Tone comes from the system instructions, which define the persona, the limits, and what to do when information is missing. Well-written instructions matter as much as the chosen model, and any change to them should go through the same test set.
Privacy, Hallucination, and Control
Before contracting a model, check a few points:
- Data: where conversations are processed and how long they are stored.
- Training: whether the vendor uses messages to train models, which conflicts with Meta’s terms.
- Control: whether it is possible to limit topics, require an internal source, and force a transfer to humans.
Against hallucination, the defenses are RAG, short answers, an instruction to transfer when data is missing, and ready-made answers on sensitive topics. Prices, lead times, and commercial terms are safer when they come from a system rather than from generated text.
How AI Connects to WhatsApp Through the Official API
The AI model does not talk to WhatsApp directly. Between the two sits the WhatsApp Cloud API, which receives and sends messages, and a server that bridges the model and the company’s systems.
Webhooks, the Cloud API, and the 24-Hour Window
On the WhatsApp Business Platform, every incoming message and every delivery status reaches the server by webhook. The server passes the text to the model, receives the answer, and sends it through the Cloud API.
The server should validate each event’s signature and discard duplicates, because Meta resends for up to seven days when it gets no acknowledgment. Outside the 24-hour window opened by the customer, only approved templates can be sent, and the AI cannot write freely.
Coexistence, Unofficial APIs, and Handoff
With WhatsApp coexistence, the same number stays on the WhatsApp Business app and on the API, and messages sent from the app remain free. The team replies from the phone while the agent serves through the API.
Tools that imitate WhatsApp Web violate the terms and raise the risk of a ban, as the guide to the unofficial WhatsApp API explains. With the official API the risk is lower, and handoff follows the practices of human and AI customer service.
Integration with Company Systems
To take action, the server exposes only the necessary tools to the model, such as looking up an order or creating a lead, each with its own credentials. Integrating WhatsApp with the CRM is the classic example, alongside calendars and management systems.
Logs of each call, with input, output, and timestamp, make it possible to audit what the AI did on the company’s behalf. Without this history, it is hard to explain to the customer or the team why an action was taken.
Meta’s Rules for AI on WhatsApp Business
The WhatsApp Business Solution Terms and the WhatsApp Business Messaging Policy include specific points for anyone using AI models. They apply to any business connected to the API, not only to those that develop AI.
General-Purpose AI Providers and the Brazil Exception
The terms prohibit providers of general-purpose AI, LLMs, or assistants from using the Business Solution when AI is the primary functionality rather than incidental or ancillary. There is an exception for users with Brazilian numbers (+55), in effect since March 11, 2026; the exception for the European Economic Area ended on May 12, 2026 (Meta’s official page).
In Brazil, since March 11, 2026, these providers pay for each non-template message delivered to +55 numbers, according to the page on AI provider pricing. The rule does not change billing for businesses that use AI in their own customer service.
Conversation Data and Model Training
A business cannot allow Business Solution data to be used to create, train, or improve third-party AI models. Fine-tuning a model used exclusively by the business itself is allowed under the terms.
In practice, the contract with the model vendor must bar the use of conversations for training. A vendor that trains models on the data it receives creates a direct conflict with Meta’s terms.
Human Escalation and Sensitive Data
The Messaging Policy allows automation as long as there are prompt, clear, and direct escalation paths to a human. Three precautions address much of the risk:
- Visible handoff: a keyword or button to talk to an agent at any time.
- No sensitive identifiers: do not request or send full payment card numbers, bank account numbers, or national ID numbers.
- Logging: record when and why the AI transferred each conversation.
Costs of AI Models for Chatbots on WhatsApp in 2026
The total cost adds up two independent bills: what the model vendor charges for processing and what Meta charges for delivered messages. Ignoring either one distorts the project budget.
What Meta Charges from October 1, 2026
Meta charges per delivered message, by message category and by the recipient’s country, according to Meta’s official pricing page. In Brazil, for example, from October 1, 2026, Meta charges R$ 0.3217 per marketing message and R$ 0.035 per delivered utility, authentication, or service message.
Each phone number gets 1,000 free service messages delivered per month. Because chatbot replies are service messages, the AI’s volume uses up that allowance. The details are in the guide to WhatsApp API pricing, and the cost calculator (Brazil, in Portuguese) simulates spend for Brazilian numbers.
Model Cost per Conversation
The model vendor charges separately, usually by tokens. Long answers, multiple tool calls, and extended histories multiply the cost, even when the conversation produces few messages on WhatsApp.
A simple estimate starts from the number of conversations per month, the average number of messages the AI sends in each, and the average tokens per answer. With these three figures, both bills can be projected before choosing the model.
Complete answers in a single message lower both bills, because every delivered message above the free allowance is charged. Meta’s own Business AI agent follows a different logic, with token-based billing separate from delivery.
Evaluating and Monitoring AI Models
A model that passed testing can get worse with a new version, an outdated knowledge base, or a changed instruction. That is why evaluation is ongoing, not a one-time step before launch.
Metrics to Evaluate the Model
The indicators below show whether AI models for chatbots actually resolve requests or just converse:
- Resolution without a human: conversations completed by the AI without a transfer.
- Escalation rate: when and why the conversation was transferred.
- Corrected answers: errors flagged by the team or reversed in the system.
- Tool failures and latency: calls with errors and time per call.
- Satisfaction: CSAT or NPS collected at the end of the interaction.
There are no official benchmarks for these rates, so what matters is the week-over-week trend. A sample of conversations reviewed by people reveals errors the numbers miss, such as the wrong tone or outdated information.
Traditional customer service metrics and a short satisfaction survey on WhatsApp complete the picture.
Prompt Injection and Excessive Agency
OWASP lists prompt injection as the top risk for LLM applications: messages or external content that try to change the model’s instructions. OWASP itself states that it is unclear whether foolproof prevention exists.
Excessive agency arises when the agent has too many tools, permissions, or too much autonomy. Mitigations include few tools, least privilege, authorization in the backend, call limits, and human approval for high-impact actions.
Data Protection and Deploying AI Models for Chatbots
Choosing and connecting the model is half the work. The other half is handling personal data within the law and launching the chatbot in stages, without opening everything at once.
LGPD: Automated Decisions and Data Minimization
Under Article 20 of Brazil’s data protection law (LGPD, Law 13,709/2018), individuals may request a review of decisions made solely through automated processing that affect their interests, including consumer and credit profiling. If asked, the controller must disclose the criteria used. Similar rules apply under GDPR in the EU.
The principles of purpose limitation, necessity, and security call for the chatbot to collect only what it uses. For business-initiated messages, WhatsApp opt-in must be recorded, and questions about legal basis should go to the legal team or the data protection officer.
Step-by-Step Deployment
A lean roadmap reduces rework:
- Map the most frequent questions and actions in current customer service.
- Define the model type for each one: rule, RAG, or tool.
- Build the knowledge base, the instructions, and the agent’s permissions.
- Test with real conversations in the customers’ language before opening to the public.
- Launch with human handoff and review the metrics every week.
The guide on how to create AI agents for WhatsApp details each step, and smaller businesses can start with a narrower scope, such as frequent questions only, before adding actions in other systems.
ConverZap: AI Agents on the Official API
ConverZap connects the AI agent for WhatsApp to the official API with coexistence, integrates databases and CRMs to take actions, and brings together template-based messaging and customer service dashboards.
Businesses that want to define the right AI model for their own customer service can talk to the ConverZap team and assess the scenario based on conversation volume and the systems they already use.
FAQ: AI Models for Chatbots
There is no single best one. Predictable flows work with rules, varied questions call for an LLM with a knowledge base, and requests that require action call for an agent with tools. Testing with the company’s real questions decides the choice.
Generally, no. The model improves when the team updates the knowledge base, the instructions, and the examples. On WhatsApp, Meta’s terms also prohibit using conversations to train third-party AI models.
Not necessarily. Ready-made platforms connect the model to the WhatsApp API with visual configuration. Coding comes in when the agent needs to call in-house systems that have no ready integration.
Yes. Open models on the company’s own servers give more control over data but require infrastructure, updates, and monitoring. Quality in the customers’ language needs to be tested as with any other model.
Yes, with a transcription step before the language model. The official API delivers the audio as media, the system transcribes it, and the model replies in text or, if configured, in generated audio.
It depends on the scope. A question-answering agent with a knowledge base is ready sooner than an agent that takes actions in several systems, which requires integrations and permission testing. Starting with a limited use case shortens the path.
