WhatsApp AI agent for business: sales and support answered from your own knowledge

What is a WhatsApp AI agent, and how is it different from a menu bot?

A menu bot is the "press 1 for sales, press 2 for support" experience moved into chat. It works while the customer follows the script and breaks the moment someone writes "hi, I have a question about last week's order" in one line. A WhatsApp AI agent reads that line, understands that it is an order question, looks the order up if the system is connected, and answers in plain language. It handles free text, typos, voice notes transcribed to text, and the very common case of three questions in a single message.

The second difference is where the answers come from. A menu bot only knows the branches someone drew. An agent answers from your company's knowledge: the product catalog, the price list, the delivery policy, the FAQ your team keeps in a document and, when it makes sense, live data from your CRM, your booking calendar or your order system. The third difference is judgment: it qualifies a lead with a few questions, books a slot, and recognizes when the right move is to stop and hand off to a human with the full context attached.

Can one WhatsApp AI tool answer questions, quote, book and hand off to staff?

Yes, but rarely as a single product you subscribe to. Off-the-shelf WhatsApp chatbot platforms handle common questions well. Once the workflow also has to read your stock, apply your pricing rules, book into your calendar and pass a ready buyer to a person, it becomes an integration build on the WhatsApp Business Platform connected to your own systems. That build is what a custom agent is.

A service business that sends estimates. The agent answers the common questions, collects the details of the job (photos, measurements, address) and drafts an estimate from your price list or your quoting sheet. It connects to that pricing source and to your CRM. A person approves the estimate before it goes out, and anything that sounds like a complaint or a serious issue goes straight to staff with the history attached.

A retailer that checks stock and flags VIP buyers. The agent answers product questions from the catalog and checks availability in your inventory or e-commerce system in real time, never from memory. It reads purchase history in your CRM or store to recognize high-value customers under a rule you define, then notifies a salesperson who continues the chat personally.

A clinic or studio that books appointments and sends a payment link. The agent offers real slots from your calendar, books the appointment and sends a payment link generated by your existing payment provider; it never handles card data. Reminders and follow-ups outside the customer service window go out as approved templates to customers who opted in. Clinical questions, reschedules with exceptions and refunds go to a person.

An academy handling enrolment enquiries on WhatsApp and email. The same knowledge base about courses, dates, fees and requirements answers on both channels, so staff stop repeating themselves. The agent asks the few questions that separate a curious visitor from a serious applicant, records the answers in your CRM or admissions sheet, and hands applicants who are ready to enrol to an advisor right away, with the conversation summarized.

When is an off-the-shelf WhatsApp chatbot platform enough?

If you only need answers to frequent questions, a greeting, business hours and simple routing to the right person or department, a subscription to a WhatsApp chatbot platform is cheaper and faster than a custom build. You can configure it yourself without an engineer, the provider maintains it, and a shared inbox for your team is often included. If that is your case I will say so on the first call, because I do not resell tools and a custom agent would be money badly spent.

The platform stops being enough when the answer depends on data that lives outside it: stock that changes by the hour, prices with rules and exceptions, a calendar several people share, customer history in your CRM. Most platforms offer connectors, and for a standard setup they may be all you need. When the connectors do not reach your systems, or the handoff loses the context and your staff ask the customer everything again, the missing piece is integration work. It can sit on top of the platform you already pay for, replacing nothing.

How does the WhatsApp Business channel work behind the scenes?

Personal WhatsApp and the free WhatsApp Business app do not give software a way to read and answer messages. What you need is the WhatsApp Business Platform, which Meta operates through its Cloud API, or the same platform reached through a business solution provider (BSP) that adds a dashboard, billing in local currency and support. Either way, your company number is registered under a verified Meta Business account, incoming messages arrive at a webhook you control, and your system replies through the API.

Three rules shape how the agent behaves. First, when a customer writes to you, a customer service window opens (24 hours at the time of writing) during which you can reply freely. Outside that window you can only start a conversation with a pre-approved message template, which is how appointment reminders and order updates go out. Second, templates are reviewed by Meta and messages are priced by category. Third, people must have opted in to hear from you; a number copied from a list is not consent. These policies and prices change. Before a build I check Meta's current documentation, and I recommend you do the same rather than trusting any blog post, including this page.

Which WhatsApp AI agent use cases actually pay off?

The agents that pay for themselves share a pattern: a high volume of repetitive conversations, an answer that lives in a document or a system, and a real cost of slowness. If a lead who writes on Saturday gets an answer on Monday, you lost part of them. If a client asks "where is my order" ten times a day, someone in your team is doing lookups instead of selling.

The internal assistant is the use case people forget. A number your own team messages to ask "what is the lead time on product X" or "what did we quote client Y last month" pays back fast, because the team is already in WhatsApp all day and the alternative is interrupting the owner. Post-sale follow-ups are the other quiet winner: templates for reminders, reviews and reorders go out on schedule, and the agent handles the replies instead of leaving them unread.

What does the architecture of a WhatsApp AI agent look like?

In plain words: a message arrives at a webhook, an HTTPS endpoint Meta calls whenever someone writes to your number. The webhook hands it to the agent runtime, which is an n8n workflow or a small Python service depending on how much custom logic you need. The runtime loads the conversation memory for that phone number, retrieves the relevant fragments of your documents from a Postgres database with pgvector, and sends everything to the model (Claude or OpenAI through their APIs) together with instructions about tone, what it may promise, and when to escalate.

The model returns a reply, a tool call (look up an order, check the calendar, create a lead) or an escalation. Replies go back through the API. Escalations flag the conversation, notify a person in Slack or in WhatsApp itself, and pause the agent on that thread until the human is done. Every message in and out, every tool call and every escalation is written to an audit log you can query, so when a customer says "your bot told me X" you can see exactly what was said and why.

What guardrails does a WhatsApp AI agent need?

A WhatsApp agent talks to customers in your name, so the guardrails matter more than the model. The first rule: it never invents prices, stock or availability. If the answer is not in the retrieved documents or in a system it can query, it says it will check and escalates. The second: anything that looks like a commercial commitment (a formal quote, a discount, a delivery date you have not confirmed) goes to a person for approval before it is sent. The agent drafts, the human approves.

The third: it detects frustration, repeated questions and an explicit "I want to talk to a person", and hands off instead of looping. The fourth is personal data. WhatsApp conversations contain names, addresses, ID numbers and sometimes payment details; the agent stores only what the process needs, the log lives in your database and not in a third-party tool, and retention is a setting you control. Add rate limits, a kill switch per thread and global, and a daily review of escalated conversations during the first weeks.

What does a WhatsApp AI agent build look like, week by week?

Week one is discovery and plumbing. We map the conversations you actually get (I ask for two weeks of real chats, anonymized), decide which ones the agent handles and which go straight to a person, and set up the channel: Meta Business verification, the number, the webhook and the first template. Your documents are collected and loaded into the retrieval database. By the end of the week you see the agent answer your ten most common questions on a test number.

Week two is the agent itself: the prompts, the tools (calendar, CRM, order lookup), the escalation logic and the approval queue. We iterate by replaying real conversations against it. Week three is hardening and handoff: monitoring and alerts, the audit log view, the kill switch, load testing, documentation, the exported workflows and code in your repository, and Loom videos walking through every piece. It runs in your Meta account, your n8n or your cloud, with your API keys. If something I built breaks in the following 60 days, I fix it at no cost.

How much does a WhatsApp AI agent cost?

The reference build is the AI Intake Agent with knowledge base, published on this site at USD 4,900 fixed, with an optional USD 1,200 per month for operation and iteration after the guarantee period. That is the scope described on this page: the channel set up, the agent answering from your documents, lead qualification and booking, escalation to a human, and the audit log. Larger scopes, for example an agent that also drives quoting, follow-ups and internal support across several numbers, are quoted as a single fixed price after a discovery call, with the AI Operating System at USD 6,800 as the upper reference point.

What is not inside the fixed price: the messaging fees Meta or the BSP charge you, the LLM API usage and the hosting. All of those are billed to you directly, in your name, with a cap declared in the proposal, so you own every account and there is no markup on tools. I do not bill hourly for builds and I do not resell software.

When is a simple auto-reply enough, and when do you not need me?

If your volume is a few conversations a day and every answer is the same three lines, the greeting and away messages in the WhatsApp Business app plus a quick-reply list solve it for free. If your questions are truly a fixed menu (opening hours, address, three services at fixed prices), the no-code flow builder your provider already offers is cheaper than a custom agent, and I will tell you so on the first call.

A custom agent earns its cost when conversations are varied, the answers live in documents or systems, speed changes revenue, and someone is spending hours a day inside one phone. If that is your situation, the 15-minute call is where we confirm it with your real numbers before anyone writes a proposal.