WhatsApp Bots: Deterministic Flows or Generative AI? When to Use Each

Not all WhatsApp bots are the same. There is a massive difference between a bot that follows a fixed script and one that uses artificial intelligence to generate responses. Each approach has clear advantages, and choosing wrong can cost you money, customers, or both.

In this guide we explain both approaches, when to use each, and why the best solution for most companies is a combination of the two.

Table of contents

• Two approaches, one channel
• Flow bots (deterministic)
• Generative AI bots
• The hybrid approach: best of both worlds
• Direct Response: why it is critical for transactional data
• Guardrails: safety controls for your bot
• Decision guide: where to start
• Conclusion

Two approaches, one channel

When someone says “I want a WhatsApp bot,” they could be thinking of very different things:

Option A: A bot that follows predefined steps. The customer writes, the bot asks for a piece of data, looks it up in a system and gives the exact answer. Like an ATM: predictable, fast, no surprises.

Option B: A bot with artificial intelligence that understands natural language, answers open-ended questions and generates unique responses for each conversation. Like talking to a trained agent (but one who never gets tired).

Neither is “better” than the other. The key is knowing when to use each one.

Flow bots (deterministic)

A flow bot works like a decision tree. It has defined steps, clear conditions and pre-written responses. It doesn’t improvise.

How it works

1. The customer sends a message
2. The bot identifies the intent (can be by keyword, menu or basic classification)
3. The bot follows a sequence of steps: ask for data, validate, call an API, respond
4. The response is a fixed template with interpolated variables

Concrete example: balance inquiry.

Customer: I want to know how much I owe
Bot: To check your balance, I need your service number. What is it?
Customer: 44513
Bot: [calls internal API, gets balance]
Bot: Your current balance is $1,250.00. Due date: April 15, 2026.

Every step is predefined. The bot doesn’t “decide” what to respond; it executes a script.

Advantages

Zero hallucinations. The bot never invents data. If the API returns $1,250.00, that is what the customer gets. There is no risk of the model rounding, paraphrasing or making up a number.

No AI API costs. You don’t need an OpenAI, Anthropic or Google API key. The bot runs on pure logic. This matters if you handle thousands of conversations per month.

100% predictable behavior. You can test every flow end to end and know exactly what the bot will respond in each scenario. No variability.

Speed. Running a deterministic flow takes milliseconds. No waiting for a language model to generate a response.

When to use them

• Balance and account status inquiries
• Payment reference generation
• Reporting issues or incidents
• Scheduling appointments or services
• Any operation where the correct answer depends on data from an external system

Rule of thumb: if the answer contains numbers, dates, codes or references, use a deterministic flow.

Generative AI bots

A generative AI bot uses a language model (like GPT-4, Claude or Gemini) to understand the customer’s message and generate a response. It doesn’t follow a fixed script; it “thinks” and responds.

How it works

1. The customer writes a message in natural language
2. The message is sent to a language model along with a system prompt (instructions) and the conversation history
3. The model generates a unique response based on the context
4. The response is sent to the customer

Concrete example: question about cancellation policies.

Customer: Hi, I'm thinking about canceling my service. What options do I have?
Bot: I understand you are considering canceling. Before we proceed, I would
     like to understand your situation better. There are several options we
     can explore:
     
     1. If it's about the price, we have more affordable plans that might
        be a better fit for your budget.
     2. If there's a technical issue, I can help you resolve it.
     3. If you definitely want to cancel, I can connect you with an agent
        who will guide you through the process.
     
     What's your situation?

This response isn’t pre-written. The model generated it based on context and instructions.

The RAG pattern: answering with your documents

RAG (Retrieval-Augmented Generation) is a pattern where the bot searches your documents before answering. Instead of relying solely on the model’s general knowledge, it looks up information specific to your company.

How it works:

1. You upload your documents (PDFs, manuals, FAQs, policies) to the platform
2. The documents are processed and stored in a vector database
3. When the customer asks something, the bot searches for the most relevant fragments
4. The model generates the response using those fragments as reference

Example: a customer asks “What are your business hours?” The bot searches your documents, finds the hours section and responds with the correct information from your company, not generic data.

Advantages

Understands natural language. The customer doesn’t need to type exact keywords or navigate menus. They can ask however they want and the bot understands.

Handles unexpected questions. A deterministic flow can only answer what was programmed. An AI bot can handle questions you never anticipated, as long as it has the information in its documents or prompt.

Fluid conversations. It can follow up, ask for clarification and maintain context throughout the conversation. It feels more natural.

Learns from your documents. With RAG, the bot becomes an expert on your company without needing to program every answer.

Risks

Hallucinations. Language models can make up information that sounds convincing but is incorrect. This is unacceptable for financial or transactional data.

API costs. Every message to the model has a cost. At high volumes (thousands of conversations per day), this adds up.

Latency. Generating a response takes 1-3 seconds, versus milliseconds for a deterministic flow.

Unpredictability. The same input can generate slightly different responses each time. This can be an issue for regulatory compliance.

When to use them

• Frequently asked questions (FAQs)
• Product and service information
• Policies and procedures
• Complaints or feedback (where empathetic tone matters)
• Any question where the answer is in your documents

Rule of thumb: if the answer is informational (not transactional), generative AI is your friend.

The hybrid approach: best of both worlds

The reality is that most companies need both. Some questions are better solved with flows and others with AI. The solution isn’t choosing one or the other, but combining them intelligently.

How it works in CX Inbox

CX Inbox uses a hybrid approach with three layers:

Layer 1: Intent classification (LLM). When the customer writes, a language model classifies the message into an intent. It doesn’t generate the response; it only classifies. This takes a fraction of a second.

Layer 2: Deterministic execution (TypeScript). If the intent is transactional (check balance, generate payment, report issue), a TypeScript state machine handles the entire flow. It asks for data, validates, calls APIs and responds with fixed templates. Zero AI in the execution.

Layer 3: Generative response (LLM). If the intent is informational (general question, product information, complaint), the language model generates the response. If RAG is configured, it searches your documents first.

The principle is simple: LLM classifies, TypeScript decides, templates respond.

Example in action

A customer of a telecom company writes:

“I haven’t had internet since yesterday, my service number is 44513”

1. Classification: the LLM identifies intent = “report-issue” and extracts service_number = 44513 and description = "No internet since yesterday".
2. Deterministic execution: the state machine verifies the service number against the API, confirms the account holder’s identity, registers the issue report.
3. Response: a fixed template sends the ticket number and estimated resolution time.

If the same customer then asks:

“And how much would it cost to switch to a faster plan?”

1. Classification: the LLM identifies intent = “general-information”.
2. Generative response: the LLM searches the knowledge base (RAG), finds the available plans and generates a personalized response.

All in the same conversation, without the customer noticing the difference.

BYOK: bring your own key

A benefit of the hybrid approach is that AI is optional. You can start with deterministic flows only (zero AI API cost) and add AI later when you need it.

When you are ready, CX Inbox supports BYOK (Bring Your Own Key): you connect your own API key from OpenAI, Anthropic (Claude) or Google (Gemini). There is no markup on model usage. You pay the AI provider directly.

Direct Response: why it is critical for transactional data

We mentioned hallucinations earlier, but this point is worth expanding because it is one of the most expensive mistakes.

Imagine this scenario:

1. The bot queries the API and gets: balance = $2,347.50, reference = 8901-2345-6789
2. The bot passes this data to the language model so it can “write a friendly response”
3. The model responds: “Your balance is approximately $2,350 and your payment reference is 8901-2345-6789”

Seems harmless, but:

• The amount was rounded ($2,347.50 -> $2,350). If the customer pays $2,350, the payment doesn’t apply correctly
• In other cases, the model might drop a digit from the reference or change a number

Direct Response solves this. When a tool is marked as Direct Response:

1. The tool executes and gets the data from the API
2. The data is interpolated into a fixed template (not AI-generated)
3. The message is sent directly to the customer, without passing through the model
4. The model receives a notice: “The information was sent to the customer” and continues the conversation normally

Rule: any data that contains numbers, amounts, references, ticket numbers or dates should use Direct Response. Always.

Guardrails: safety controls for your bot

A bot that handles customers needs controls. Especially if it uses generative AI. These are the most important guardrails:

Word filters

You can define words or phrases the bot should never mention (competitors, confidential information, sensitive topics). If the model tries to include them, the message is blocked and escalated to an agent.

Retry limits

If the customer doesn’t provide valid data after N attempts, the bot should escalate to an agent instead of continuing to ask. Nothing frustrates a customer more than a bot repeating the same question.

Deterministic identity verification

Name and personal data verification should never go through AI. It is done with string comparison in code: the name the customer gives versus the name in the system. If it doesn’t match, the bot asks for verification or escalates.

Negative sentiment detection

The bot can detect frustration or anger in the customer’s tone and automatically escalate to a human agent before the situation gets worse.

Turn limits

If the bot conversation exceeds a certain number of turns without resolution (for example, 15), the bot escalates automatically. This prevents circular conversations.

Decision guide: where to start

Start with flows if…

• Your main use case is transactional (balances, payments, reports, appointments)
• You don’t want AI API costs right now
• You need exact, predictable answers
• Your conversation volume is high (AI costs at scale are significant)
• You operate in a regulated industry where every response must be controlled

Add AI when…

• You receive many questions that can’t be resolved with a fixed flow (varied FAQs)
• You have documentation the bot could reference (manuals, policies, catalog)
• You want the bot to handle the first interaction and only escalate complex cases
• Your support team is overwhelmed with repetitive but varied questions

The typical path

1. Month 1: Set up 2-3 deterministic flows for your most common operations (balance inquiry, issue report, order status)
2. Month 2: Measure results. If the bot is resolving 50%+ of conversations and you have questions it can’t handle, add AI
3. Month 3: Upload your documents (RAG), configure the bot prompt with your tone and policies, enable generative responses for informational intents
4. Ongoing: Review escalated conversations. If there are patterns, create new deterministic flows. AI covers what flows don’t

Conclusion

The “flows vs AI” debate is a false dichotomy. What you need is to know when to use each tool:

• Transactional data: deterministic flows + Direct Response. No exceptions.
• Informational questions: generative AI + RAG. With guardrails.
• Classification: AI to understand intent. Fast and cheap.
• Execution: TypeScript for business logic. Predictable and safe.

The hybrid approach isn’t a compromise; it is the correct architecture for an enterprise bot that handles real customer data.

Try CX Inbox free — 14 days, no credit card required. Set up your first hybrid bot: flows for transactional, AI for conversational.

---

Published: April 2026. This guide reflects the current state of conversational bot technology and CX Inbox capabilities.