Notes ·
What an AI assistant on WhatsApp really costs to run
Short answerModel fees are the smallest line: about US$3 a month for our whole production system, under one US cent per AI search. What actually costs money is WhatsApp's own template fees, hosting, the build, testing, and staff time.
Most AI quotes we see for small businesses are priced as if the model were the expensive part. In our experience it is the cheapest. Below are our own numbers from a WhatsApp-based sales system a small team uses every day: a shared inbox, AI-drafted replies, rule-checked search over a few hundred records, document reading, reminders and video interviews.
1. Model fees: a few dollars a month
We log every model call with its exact token counts and cost, so these are measured, not estimated.
| What | Measured cost |
|---|---|
| One AI search (read the conversation, pick from people who pass the rules) | about 0.4–0.8 US cents |
| Main model, all daily use (drafts, searches, summaries) | about 4 US cents a day |
| Reading PDFs, embeddings and fallback models | about 43 US cents a week |
| Total | roughly US$3 a month |
Two things keep it low. First, each kind of task goes to a fast, inexpensive model, with a second and third model as automatic fallbacks. Second, the expensive part of search, checking every record against the customer's conditions, is done by plain code, which costs nothing per call. The model only reads the handful of records that already passed.
Price is not the only test. We rejected two newer models that were cheap enough but took 10 to 30 seconds to answer. A salesperson waiting 30 seconds per draft costs more than any model fee.
2. WhatsApp's own fees
On the official WhatsApp Business Platform, replying within 24 hours of the customer's last message is free. Outside that window you can only send pre-approved templates, and Meta charges for those per message, by category (marketing, utility, authentication) and by the customer's country. For a sales team that follows up a lot, this can be larger than the AI bill. Design so that most messages happen inside the window: answer fast, and use templates only for things that genuinely need them, like a reminder before an appointment.
3. Hosting
Our entire stack (inbox, database, search, video rooms, phone bridge, backups, monitoring) runs on one cloud server in Singapore. Video is the only heavy traffic: a four-person, 45-minute interview sends out about 2 GB, so thirty interviews a month is about 60 GB, well inside what most cloud providers include.
4. The costs nobody budgets for
- Testing. Before any change goes live, we replay more than 60 real past requests on a staging copy. One heavy day of testing once cost about US$5 in model fees, more than three weeks of production. Testing is worth it; just budget for it, and use a separate key or a balance alarm so tests cannot starve production.
- Review time. We keep a person approving every outgoing message. That is a cost, and it is the right one until drafts are being sent unchanged for weeks.
- Keeping it running. Backups, monitoring, updates, and someone who answers when a provider changes its API. This is where most of the long-term cost of any system sits, AI or not.
What to ask any vendor
- Can you show me the cost of each model call, per task, after launch?
- What happens when the main model is slow or down?
- Which decisions are made by code, and which by the model?
Quick answers
How much do AI model fees cost for a small business on WhatsApp?
In our production system serving a small sales team, model fees ran at about US$3 a month in October 2026, and a single AI search cost under one US cent.
Are WhatsApp Business messages free?
Replies within 24 hours of the customer's last message are free. Template messages, which are what you must use outside that window, are charged per message by Meta, at rates that depend on the template category and the customer's country.
What is the biggest hidden cost of an AI assistant?
People's time, and testing. One day of thorough testing on our staging system cost more in model fees than three weeks of production use.