AI Chat on Website or WhatsApp: Where to Start Your First Pilot

AI Chat on Website or WhatsApp: Where to Start Your First Pilot

If you are wondering where to start with AI chat on your website or WhatsApp, the answer for most Brazilian companies that already serve customers by message is WhatsApp. Customers are already there, and the phone number helps you identify who is writing. Website chat makes more sense when your audience is new and anonymous, arrives from search or ads, and asks questions that depend on the page they are looking at.

The decision comes down to four factors: customer profile, conversation volume, identification and integration effort. This article compares the two channels on those points and ends with a plan for the pilot.

Key takeaways

• Start in the channel where your customers already talk to you today, not the one that looks more modern.
• WhatsApp usually wins on repeat customers and identification. The website wins with new visitors and page context.
• On WhatsApp, the effort is in channel rules (API, number, templates). On the website, it is in the interface, authentication and consent.
• A good pilot covers one flow, in one channel, with clear limits, a record of decisions and a handoff to a person.
• Measure your current service before turning the AI on. Without a baseline, you cannot decide whether the pilot worked.

What is the real difference between website chat and WhatsApp?

Website chat is a conversation started by the visitor inside your page. The session is usually anonymous and ends when the person closes the tab. You control the interface, the hours, the opening text and what happens afterward.

WhatsApp is a persistent conversation, with history, inside an app people open every day. The contact continues the next day and the next week. That changes the kind of question you get: on the website, pre-sale questions dominate (how it works, who it is for, how long it takes). On WhatsApp you also see follow-ups, duplicate invoices, rescheduling and questions from people who are already customers.

There is also a governance difference. On the website, the rules are yours. On WhatsApp, part of the rules belongs to the platform, such as how you contract the API, how outbound messages are approved and the policies against abuse. It is worth reading the comparison of WhatsApp Business API vs app before deciding how the channel will be run.

Customer profile and volume: where does the conversation really happen?

Before looking at tools, look at your real conversations. Pull a sample of recent ones and note where the customer came from, what they asked and how many times they came back. A few signals help you read the result:

• Customers who already ask for everything on WhatsApp, including those who call and then send a message, show the channel is already their preference.
• Many contacts from people who never bought, arriving from search or ads, show the website plays a lead capture role.
• Repeated questions about status, deadlines, duplicate invoices and opening hours point to a good flow to automate, in any channel.
• Highly varied questions that need case-by-case analysis point to a bad flow to start with.

On volume, the practical rule is simple: AI is worth the integration effort when there is repetition. If the team gets few conversations a day and all of them are different, the pilot costs more attention than it returns. If there are many similar questions, peaks at certain hours or a queue outside business hours, the gain shows up quickly.

On WhatsApp, volume is measured by the inbox you already have. On the website, it depends on traffic and on how many visitors actually open the chat. A site with good traffic and few conversations started may signal that chat alone will not fix a conversion problem. In that case, the first step is to review the page, not to install an agent.

If after-hours service is your main pain, the design of what to answer and what to defer matters more than the channel. There is a specific guide on after-hours customer service with AI.

Identification: who is on the other side, and what can the AI say?

Identification defines what the agent is allowed to say, and this is where the channels differ most.

On WhatsApp, the number arrives with the message. You can match it against your customer records and recognize the customer without asking for everything again. But a number is not proof of identity: devices are shared, SIM cards change owners and someone can write on behalf of another person. For personal information such as address, amounts and contract data, ask for an additional confirmation, such as a code or a detail only the customer knows.

On the website, the visitor is anonymous by default. The safest rule is that the agent only talks about public information (services, hours, general rules) until the person authenticates. If the flow needs customer data, use a login or a verification code sent to a contact already on file.

In both channels, data protection care applies. Define the purpose of the use, collect only what is necessary, tell the person they are talking to an AI system, offer a path to a human and record what was consulted. The ANPD, Brazil's data protection authority, is the official reference for guidance on the LGPD.

One security point applies to both: any open channel receives hostile text, including attempts to make the agent ignore your rules. OWASP's material on risks in LLM applications helps you build the pilot's threat list. The practical rule is to give the agent only the actions it needs, starting with read access.

How much integration effort does each channel require?

The effort is different in nature. On WhatsApp, channel rules and operations weigh more. On the website, the interface and authentication weigh more.

On WhatsApp, the work usually includes:

• Defining which number will be used and whether it will sit on the official API or in the app.
• Verifying the business with Meta, when required, and contracting a provider.
• Receiving messages by webhook and handling audio, images and documents.
• Creating and getting approval for templates for outbound messages, with customer consent.
• Connecting the service inbox so conversations can be handed to a person.

The technical walkthrough is in how to integrate AI into your company's WhatsApp. If the idea includes contacting customers on the company's initiative, first read about templates and opt-in.

On the website, the work usually includes:

• Installing the chat component and adjusting its look to your standards.
• Hosting the service that talks to the model and protecting that API.
• Handling consent, cookies and the privacy notice.
• Passing the page context to the agent, such as the product or service the person is viewing.
• Defining the queue and hours for handoff to an agent.

Both channels share a piece of work that is often the largest of all: organizing the knowledge base, writing the agent's rules and connecting the systems that answer the questions (CRM, ERP, calendar). If that work is done, switching or adding channels later is much cheaper. If it is not, it is the real first project.

In general, the website is faster for standing up a demo. WhatsApp tends to deliver value sooner because the conversation and the demand already exist. Anyone who picks a channel only because it is easy to install usually ends up testing the wrong thing.

Decision guide: which channel for the first pilot?

Use the lists below as a quick filter. Count how many statements describe your company.

Choose WhatsApp first if:

• Most of your current service already happens on WhatsApp.
• Customers are repeat buyers and ask follow-up questions.
• The customer's number is already in your records and can be matched with your system.
• There is a repetitive, well-defined flow, such as status, scheduling or duplicate invoices.
• Someone is available to take over the conversations the agent hands off.

Choose the website first if:

• Most contacts come from new visitors with no history.
• Questions depend on the page, such as a specific product, plan or service.
• You do not have an organized WhatsApp number for support, and creating that process would be the biggest obstacle.
• The goal is to capture qualified leads, not to serve existing customers.
• You want full control of the interface and rules, without depending on a messaging platform.

Illustrative scenario: a parts distributor has a small sales team. The salespeople spend much of the day answering order status questions and duplicate invoice requests on WhatsApp, while the website receives few contacts. The first pilot goes to WhatsApp, limited to order status lookup in read-only mode, with identity confirmation and handoff to the responsible salesperson when a question falls outside the scope. Website chat is left for a second phase, using the same knowledge base, focused on public questions about the catalog and lead times.

If the filter ends in a tie, prefer the channel where you can better measure today's service. Whoever can measure the before can judge the after.

How to run the pilot: map, limit, pilot, hand off

The method we use in customer service projects follows four moves. They apply to any channel.

• Map: record how service works today, who answers, which questions repeat, where delays occur and what the starting point is for comparison.
• Limit: define which data the agent can read, which actions it can take and what it must never do, such as promising deadlines, granting discounts or changing customer records.
• Pilot: put a single real flow live, with a record of each agent decision and a success metric agreed before you start.
• Hand off: design the handover to a person, with the conversation history, and after the evaluation decide whether to expand, adjust or stop.

In practice, this becomes an execution plan:

• Step 1: choose a flow with repeated questions and a verifiable answer.
• Step 2: gather the real conversations for that flow and write the correct answers with the people who handle them today.
• Step 3: define the identification required and the data the agent may consult, starting with read access.
• Step 4: write the handoff rules, including complaints, requests for a human, out-of-scope topics and any doubt on the agent's part.
• Step 5: test with real cases before opening to the public, including out-of-scope questions and attempts to get around the rules.
• Step 6: open to a share of contacts, review conversations often and adjust the rules.
• Step 7: compare against the starting point and decide. There is a framework for this in how to measure an AI automation pilot.

For the agent's structure, Anthropic publishes good material on building effective agents. The central guidance is to start simple, with well-defined flows, and only increase autonomy when there is evidence it is needed.

This article separates what comes from official documentation (channel rules, data protection), what comes from implementation experience (pilot order, flow selection) and what is a hypothesis to test in your operation (which channel performs better). The third category can only be settled by the pilot.

Frequently asked questions

Is it worth having a chatbot on the website and on WhatsApp at the same time? Only after one of the two works well. Two channels open at once double the conversation review and make it harder to understand what is working. The better path is to reuse the knowledge base and rules from the first channel when the second one comes in.

Do I need the official WhatsApp API for a pilot? For AI operation in a company, the official API is usually the more stable path, because it separates the automation from the service inbox and allows outbound messages with templates. The regular app has limits that show up quickly. The article on API vs app details the differences, and using unofficial solutions increases the risk of the number being blocked.

How long should the pilot last? Long enough to cover a normal demand cycle, including at least one peak and a few unusual situations. Instead of fixing a deadline in advance, define the decision criterion: what must happen for you to expand, adjust or stop. This avoids pilots that run indefinitely without a conclusion.

Does the customer need to know they are talking to an AI? We recommend yes, right at the start of the conversation, with a clear path to reach a person. This is consistent with the transparency principle in the LGPD and reduces misunderstandings. For the legal details of your case, consult your data protection officer or the company's legal counsel.

Sources

• Anthropic: Building effective agents
• OWASP Top 10 for LLM Applications
• ANPD: Autoridade Nacional de Proteção de Dados