AI Agent for WhatsApp: How It Works and What It Costs

An AI agent for WhatsApp is a program connected to your business number that reads incoming messages, answers them using your own price list, policies and product data, and hands the conversation to a human when it hits something it cannot close. It runs on the WhatsApp Business API, not on a phone in someone's pocket. It makes sense when you get enough repetitive questions that answering them eats real hours every day. If you receive fifteen messages a day and most of them are unique, a well-written set of quick replies in the WhatsApp Business app will serve you better and cost nothing.

What an AI agent actually handles

The useful work is narrow and boring, which is exactly why it works:

• Answering questions that have a factual answer: opening hours, address, delivery terms, warranty, whether a size is in stock, what a service includes.
• Qualifying leads: asking the three or four questions your sales team asks anyway, then writing the answers into the CRM as a new deal.
• Order and booking status: the customer sends an order number, the agent checks the database and replies.
• Taking a booking or an order and confirming it, with the slot or item actually reserved in your system.
• Routing: sending the conversation to the right person, with a short summary of what the customer already said.

What it should not do: invent facts, negotiate discounts, handle complaints from angry customers, or make promises about dates it cannot verify. A good agent is configured to say "I will pass this to a colleague" more often than owners expect, and that is a feature.

When a simple chatbot is enough

A rule-based bot with buttons — menu, options, links — is cheaper, faster to build and more predictable. It is the right tool when:

• Your questions fall into five or six fixed categories.
• You mostly need to route people to the right department or send a link.
• Wrong answers would be expensive: medical, legal, financial commitments.

An LLM agent earns its place when customers write in free form, in several languages, with typos and half-sentences, and when the answer depends on data that changes: stock, prices, schedules, order state. That is where buttons break down and a model that reads your catalogue does not.

When you should not build one

If your response time is already good and your volume is low, automation adds a layer of maintenance for no gain. If your price list lives only in your head and changes per customer, the agent has nothing to read and will guess — which is worse than silence. If your real problem is that nobody answers messages after 6 p.m., a shared inbox and one part-time person may fix it for less than a build.

Also be honest about the data. An AI agent on WhatsApp is only as good as the source you point it at. No structured catalogue, no stock data, no written policies — fix that first, then automate.

What the build involves

On WhatsApp there is a technical path you cannot skip. The agent needs the WhatsApp Business API (Cloud API) through Meta or a provider. That means a verified business, a phone number dedicated to the API, and a display name approved by Meta. A number already used in the consumer WhatsApp app has to be migrated; it cannot serve both.

Meta's rules shape the design. You can reply freely for 24 hours after a customer writes to you. Outside that window you may only send pre-approved template messages, and they are billed per conversation. So "we will just message everyone with a promo" is not how this works, and any contractor who promises otherwise is glossing over the rules.

The rest of the build is:

• A knowledge base the model reads from — catalogue, FAQ, policies — kept in a place your team can edit without a developer.
• Integrations: CRM, order database, booking calendar, payment links, 1C or your stock system.
• Handoff to humans in a shared inbox, so an operator can take over mid-conversation and see the history.
• Guardrails: topics the agent refuses, a limit on how far it improvises, logging of every exchange.
• A review loop for the first weeks, where someone reads transcripts and fixes bad answers.

What it costs

Three cost layers, and they are separate.

The build. At our studio, automation and AI agents start from 250 000 AMD, and connecting an existing system — CRM, 1C, a payment provider — starts from 149 000 AMD. Those are starting points; the exact figure follows a 30-minute brief, because the price depends on how many systems the agent has to talk to and how messy the data is.

Running costs. Meta charges per conversation under its own pricing, and the LLM provider charges per token. Both are usage-based and paid directly by you, not bundled.

Maintenance. Prices change, products change, the agent drifts. Support plans start from 9 900 AMD per month. Budget for someone to read transcripts monthly even if nothing breaks.

What to prepare before the first call

Pull one month of WhatsApp history and sort it. Count how many messages are the same five questions. Write down which questions have a single correct answer and which need judgement. Decide who takes over when the agent stops. Have your catalogue or price list in a file, not in screenshots. Confirm who in your company owns the Meta Business account — this blocks more projects than anything technical.

How to choose who builds it

Ask to see a live agent, not a demo video. Ask how it behaves when it does not know something. Ask where the knowledge base lives and whether your manager can edit it. Ask who owns the Meta account, the number and the conversation data — it should be you. Be careful with resellers who lock you into their platform: if you cannot export your history and move your number, you are renting your own customer relationships.

Frequently asked questions

Can I use my normal WhatsApp Business app number? Not at the same time. The number must be moved to the WhatsApp Business API, and after migration the app version of that number stops working. Many businesses keep the old number for existing contacts and start the API on a new one.

Will customers know they are talking to a bot? They will work it out, so say it up front and offer a human at any moment. Agents that pretend to be a person create complaints the moment they get something wrong.

How long until it is worth anything? The first version answers a limited set of questions well. Expect a few weeks of reading real conversations and correcting the agent before it covers the bulk of your traffic. Anyone promising a finished agent without that loop is selling a demo.