AI

AI agents on WhatsApp: what they do well and where they need people

delpra™ team6 min read

AI agents on WhatsApp have become the promise of the moment: they serve customers alone, sell alone, never sleep. Part of that is true. The other part depends on rules almost nobody shows in the ads. This article separates the two, based on what we have learned building this kind of agent.

What an agent actually is

A traditional chatbot follows a button script: press 1 for this, 2 for that. An AI agent understands the question the way the person writes it, consults the sources you have opened to it (catalog, calendar, CRM, frequently asked questions) and answers in natural language. It can also take actions: record the contact, book a time, notify the salesperson.

The important difference is not that the conversation is friendlier. It is that the agent can resolve cases the button script did not foresee, as long as it has access to the right information and clear rules about what it may do.

What an agent does well

  • Answering right away, at any hour. The study "The Short Life of Online Sales Leads" (Harvard Business Review, 2011) showed that companies that try to make contact within 1 hour are nearly 7 times more likely to qualify the lead than those that wait 1 hour longer. An agent covers nights and weekends, when nobody on the team is available. We go deeper in what it costs to be slow answering a lead.
  • Handling repeated questions. Opening hours, address, how it works, what is included. The kind of question that eats up the team's time and has a known answer.
  • Qualifying. Understanding what the person is looking for, by when and under what conditions, before involving a salesperson.
  • Recording. Creating or updating the CRM record with a summary of the conversation, without depending on someone remembering.
  • Routing with context. Passing the conversation to the right person with the history attached, so nobody starts from zero.

What it should not do

Problems with agents almost always come from giving them freedom where there should be a rule.

Making up prices, deadlines or terms

Language models write plausible answers even when they do not know. In customer service that is a direct risk: a discount that does not exist, a wrong date, a broken link. The rule we use is simple: price, date, time and link can only come from an official source checked at that moment (the price list, the calendar, the system). If the source did not answer, the agent says it will confirm and calls a person.

Negotiating and deciding exceptions

A discount outside the price list, an unhappy customer, a situation that is not in the manual. That is work for people. The agent can recognize the case and route it quickly, which is already a lot.

Pretending to be a person

Besides being a matter of honesty, it is a practical matter: when the customer finds out, trust is gone. The agent should introduce itself as an automated service and offer a path to talk to someone.

Insisting with someone who asked it to stop

If the person asked not to receive messages, the agent needs to record that and respect it. This applies to any communication and is required by data protection laws such as Brazil's LGPD and by WhatsApp's own policies.

When to hand off to a person

Handing off to the team is not a failure of the agent, it is part of the design. Define the triggers before going live:

  1. The person asks to talk to someone.
  2. The subject is a complaint, a cancellation or billing.
  3. The question involves a price or term the official source does not cover.
  4. The agent did not understand after two attempts.
  5. The lead is qualified and ready to talk to the salesperson.

And define what happens at the handoff: who receives the alert, where, and what goes with it. The minimum is a summary of the conversation, the contact details and the reason for the handoff. Without that, the customer repeats everything and the time saved disappears.

A good agent is not the one that solves everything alone. It is the one that solves what it can and knows when to call the team.

Rules for going live safely

Use the official API

WhatsApp has an official API for businesses, Meta's Cloud API. Tools that simulate WhatsApp Web outside the rules may work for a while, but they put the number at risk of being blocked. For an operation that depends on the channel, the savings are not worth it.

Respect the 24-hour window

Under WhatsApp's rules, a business can reply freely for 24 hours after the customer's last message. Outside that window, only with approved message templates. The agent needs to know this rule so it does not try to send what will be rejected.

Have a test mode

Before talking to customers, the agent talks only to a list of approved numbers, from your own team. In this mode it does not write to production systems or send anything outside. This is where the strange answers show up, and it is better that they show up for you.

Measure the cost per conversation

Every AI response has a cost. It is usually small per conversation, but it grows with volume and with the size of the context the agent carries. Track this number from day one: it guides technical choices and avoids surprises on the bill.

Log and review

Read conversations every week at the beginning. The questions the agent could not answer show what is missing from the knowledge base. The ones it answered badly show where a rule is missing.

Is it right for my company?

The report "A Future That Works", from the McKinsey Global Institute (2017), estimates that fewer than 5% of occupations could be automated entirely, while about 60% have at least 30% of activities that can be automated. Customer service is a good picture of that: a large part is repetitive and a smaller, more important part needs people.

An agent makes sense when three things are true: there is a volume of similar messages, the information needed to answer exists and is organized, and there is a team to receive the cases that get handed off. If the information is scattered or the service process is unclear, start there. We cover that in map the process before you automate it.

Want to know where your customer service stands today? The Operational Maturity Assessment evaluates this in four minutes. And if you want to see how delpra™ builds agents, see our services.

Written by the delpra™ teamProcesses, automation and AI so operations scale efficiently.

How mature are your operations?

The Operational Maturity Assessment takes 4 minutes and shows where to start.

Take the free assessment

Read next