AI in customer service became an easy promise. Reply faster, serve 24 hours a day, cut queues, sell more, qualify leads, recover opportunities, and take the load off the team.

All of that can happen. But the opposite can happen too.

In Merge's recent sales conversations, clear signs of an increasingly common problem showed up: poorly configured service AI doing more damage than having no AI at all. Internal reports pointed to cases of AI ending conversations on its own, routing customers to the wrong department, answering questions about delivery time and shipping without enough basis, and confusing customers who sent audio messages into thinking they were talking to a person.

Get MergeNews in your inbox

Every week, news about WhatsApp, AI, CRM, and sales automation to turn customer service into sales.

The problem is almost never "AI" in the abstract. The problem is putting AI on top of a bad process.

When a company has no history, categories, ownership, handoff rules, a knowledge base, or a clear limit on what can be answered, AI just speeds up the confusion. Instead of organizing service, it distributes errors at scale.

The first mistake is using AI as an FAQ with a nice voice. The customer doesn't just want a polite answer. They want a resolution. If the AI replies generically and doesn't grasp intent, urgency, or stage in the journey, it just swaps the old script for smoother-sounding text.

The second mistake is not setting limits. There are questions AI can safely answer: hours, simple status checks, commercial policy, required documentation, location, payment methods, FAQs, and initial data collection. There are questions that require a human: negotiation, exceptions, sensitive complaints, conflict, uncertain deadlines, legal matters, health, delicate financial issues, or commercial promises.

The third mistake is not connecting AI to history. A customer who already bought, filed a complaint, received a proposal, or talked to a salesperson shouldn't be treated as if they're starting from zero. Without a CRM, AI has no context. And without context, it can repeat questions, miss the stage, or send the customer to the wrong place.

The fourth mistake is automating conversation closure without care. Closing a ticket may look efficient on a report, but it can be frustrating for the customer. A conversation closed too soon can turn into a new message, rework, a complaint, or a lost sale.

The fifth mistake is measuring only speed. An AI can reply in seconds and still make the experience worse. The right metric isn't just response time. It's resolution, correct handoff to a human, satisfaction, conversion, rework, avoided errors, and cost per useful interaction.

Opinion Box's own 2025 WhatsApp in Brazil survey shows why this care is necessary. It found that 59% of consumers say they dislike it when companies use automated replies. That doesn't mean automation is bad. It means bad automation gets noticed fast. On WhatsApp, customers can tell when a company is trying to push a bot to avoid actually helping them.

At the same time, the channel remains strong. According to the same survey, 82% of users already communicate with brands on WhatsApp and 75% use the app to ask questions or request information. In other words: there's demand for service on this channel. The opportunity lies in serving better, not just automating more.

The arrival of Meta Business Agent and AI agents on CRM platforms reinforces this direction. The market is moving away from fixed chatbots and toward agents that classify, recommend, log, query data, and execute tasks. But the more autonomy AI gains, the more governance matters.

A useful AI on WhatsApp needs four things: context, limits, handoff, and metrics.

Context to know who the customer is, where they came from, what they asked for, and what stage they're at. Limits so it doesn't answer things it doesn't know. Handoff to bring in a human when the case requires it. Metrics to find out whether it's actually resolving things or just replying.

The human's role also changes. They don't need to join every conversation, but they need to be in the right ones: negotiation, exceptions, complaints, qualified leads, sensitive cases, and final decisions. AI prepares the service; the human resolves what requires judgment.

This is where Merge comes in: AI on WhatsApp shouldn't be a pretty layer on top of a mess. It needs to be connected to the CRM, to history, to queues, to owners, and to the next step. Without that, the company gains speed and loses control.

The best AI isn't the one that talks the most. It's the one that knows when to answer, when to ask, when to log, and when to call in a person.

Practical action for businesses

  • Define which topics AI can answer and which should go to a human.
  • Connect AI to the customer's history, source, stage, and owner.
  • Create clear handoff rules for urgency, exceptions, complaints, and purchase intent.
  • Review automatically closed conversations to identify errors and rework.
  • Measure resolution, conversion, correct handoff, and satisfaction — not just response time.
Sources