Should You Build Your Own WhatsApp AI With the Gemini or GPT API?

The question comes up in almost every call with a company that has its own developers: "Why don't we just connect WhatsApp to the Gemini API ourselves?" It is a fair question. The models are good, the APIs are open, and a first demo can work in an afternoon.

You can build your own WhatsApp AI on the Gemini or GPT API, and the model will be the cheapest part of it. A reply costs a fraction of a cent. The cost is in everything around the model: connecting it to your orders and CRM, handing over to people, following Meta's rules, protecting customer data, testing it in Arabic and running it every day. Build it if you have engineers with time to own it for years. Otherwise, work with a team that already runs it.

The model is the cheap part

Here is the arithmetic for one reply. A typical customer service reply sends the model your instructions, the relevant answers from your knowledge base and the recent conversation, about 3,000 tokens in, and gets about 300 tokens back.

Google lists Gemini 3.5 Flash-Lite at 0.30 dollars per million input tokens and 2.50 dollars per million output tokens on the paid tier. That reply costs about 0.17 cents. At 1,000 replies a day, the model costs around 50 dollars a month.

Prices change often and differ between models, so pull the current rates from Google or OpenAI before you plan. The conclusion rarely changes: for a commerce business, the model bill is small next to Meta's message charges, the engineering time and the people who keep the system right.

That matters if your margin per order is thin. Do not count the cost per message. Count the cost per conversation that ends in an order: Meta's charges, the model, and the time your team still spends on the chats that need a person.

What you are really building

The demo that answers three questions is perhaps a tenth of the work. A system you can leave running during a sale needs all of this:

• The WhatsApp connection. A WhatsApp Business Platform account, webhooks, message templates for anything sent outside the 24 hour window, and the rules for each message category and its price.
• The knowledge. Prices, stock, branches and delivery rules that change every week, kept current without a developer.
• The actions. Taking an order, booking an appointment, updating a delivery. This is where a chatbot becomes useful, and where most of the work is. We compared 13 tools on exactly this question.
• The handover. WhatsApp's Business Messaging Policy requires a prompt, clear and direct way to reach a person when automation answers. Someone has to receive that chat with its full history.
• The language. Customers in Lebanon and the Gulf write in Arabic, English and Arabizi, often in one message. Testing that properly is its own project.
• The mistakes. Every AI gives wrong answers sometimes. You need a way to catch them fast, before a customer screenshots one.
• The insight. Reading across thousands of chats to see what people ask for, what you do not sell and why they do not buy. Without this, you have an answering machine.

None of these is hard alone. Together, they are a product that needs an owner.

AI should talk. Code should act.

The most important design choice is not which model to use. It is what the model is allowed to touch.

In the system we run, the AI handles the conversation: it understands the question, finds the answer and writes the reply. Everything that changes a record, such as creating the order, saving the address or updating the CRM, is done by ordinary, deterministic code that the AI calls with checked inputs. The AI never writes to your database directly.

This makes the system predictable. An order is either valid or rejected by rules you can read. Nothing is created because a model misread a sentence, and every action can be traced afterwards.

The data question

Companies that consider building usually do it for one reason: they do not want customer data going to a third party. That is the right instinct. Here is what each path really means.

The free tier of an AI API is the wrong place for customer chats. Google's Gemini API terms say that content sent to unpaid services can be used to improve Google's products and may be read by human reviewers, and they tell users not to submit sensitive, confidential or personal information. The paid tier is different: Google states that prompts and responses are not used to improve its products.

Self hosting a small model keeps the data in house, but you take on its quality, its Arabic, its uptime and its updates. You still need everything in the list above.

A well designed partner setup can send almost nothing. In plumcut's setup, we do not store raw customer profiles. Only the phone number is used to match the customer in your own system, through APIs you provide, and identifying details are removed from the conversation before it reaches any model. For businesses that want zero customer data to leave, the assistant can run on your FAQs alone.

Our guide to customer data safety has ten questions to ask any vendor, including us.

What Meta allows

One worry we hear is that Meta has banned AI chatbots on WhatsApp. It has not banned yours. Since 15 January 2026, the WhatsApp Business Solution Terms stop providers of general purpose AI assistants from using the platform when the AI is the main product. A business answering its own customers is a different case, and the terms let a business use an AI provider as its third party service provider.

Build or partner?

Both are reasonable. The mistake is to compare a three question demo with a running system and decide that building is cheaper.

Where plumcut fits

plumcut builds and runs an intelligent automation system on WhatsApp that answers, sells and books, connects to the systems you already use, and shows you what customers actually want. We choose the lowest cost model that still answers well for each job, and we change it as prices and models move, so you never have to.

If your team is weighing a build, send us your real customer questions. We will show you what the full system looks like on your own data, and you can decide with both options in front of you.

Sources

Checked on 8 November 2026.

• Google AI for Developers, Gemini Developer API pricing: Gemini 3.5 Flash-Lite at 0.30 dollars per million input tokens and 2.50 dollars per million output tokens on the paid tier, and whether data is used to improve products on each tier.
• Google AI for Developers, Gemini API Additional Terms of Service: use of content sent to unpaid services, human review, and the instruction not to submit personal information.
• WhatsApp, WhatsApp Business Solution Terms: the AI providers clause and the permission to retain an AI provider as a third party service provider.
• WhatsApp, Business Messaging Policy: the 24 hour customer service window and the required escalation paths for automation.
• Meta for Developers, Pricing on the WhatsApp Business Platform: per message charges, including service messages after the first 1,000 per number per month since 1 October 2026.