
Superintelligence for Social Media Customer Service: Reply or Hand Off
Learn how to set rules for when an AI agent answers social media comments and direct messages, when it hands off to your team, and how to log every case.
Want to apply this in your company?
Book a call with Juan →An AI agent should answer comments and direct messages when the question is recurring, the answer has already been approved by the company, and the topic does not involve money, a complaint, sensitive personal data, or a commitment that only a person can make. In every other case, it should acknowledge the message, summarize the context, and pass the conversation to the team with the full history. Every reply, every handoff, and every refusal needs to be logged, because that record is what lets you revise the rules later.
Key takeaways
- Let the agent answer alone only what already has an approved answer: opening hours, address, order status, required documents, and frequently asked questions.
- Send to a person anything involving refunds, prices outside the published table, complaints, cancellation threats, sensitive data, or promises about deadlines.
- Set a retry limit: if the agent does not resolve the issue in two or three exchanges, the conversation should move to the team.
- Log each case with the reason, the reply sent, the channel, the time, and who took over, so you can measure whether the rule works.
- Start with one channel and one type of question, validate the log, and only then widen the scope.
Where the agent acts and where it should not reply
The word superintelligence shows up often in content about social media, but for anyone running customer service, the useful question is a different one: what is the agent allowed to do, and what is it not allowed to do? A support agent is a narrow system with a defined scope, written instructions, and limited access to information. The difference between that kind of tool and the broader idea is explained in Is ChatGPT Superintelligence? What It Is and What It Is Not, which is worth reading before you discuss expectations with the board.
In practice, the decision to reply or hand off can be made with a simple guide. Use the list below as a starting point and adjust it with the people who already handle the inbox today:
- Reply: questions with a fixed, documented answer, such as opening hours, location, contact options, and the status of an order already on record.
- Reply and confirm: information requests that depend on a system record, such as an appointment or order status, as long as the lookup goes through an authorized integration.
- Hand off with a summary: discount requests, negotiation, technical questions outside the script, explicit dissatisfaction, or messages with an urgent tone.
- Hand off without responding to the content: reports of wrongdoing, mentions of health, detailed financial situations, third-party data, or any sign of a scam.
- Neither reply nor hand off automatically: messages from accounts or profiles the company does not recognize, offensive content, and public comments with crisis potential. In these cases the rule is to alert the team immediately.
The hardest category is usually public comments. A public reply is visible to everyone and can become evidence in a dispute. That is why many operations start by letting the agent draft replies for approval, and only later allow automatic replies on a short list of topics.
How to write the handoff rule, step by step
The rule does not need to be sophisticated to work, but it has to be written down. A rollout usually follows this sequence:
- First, pull recent conversations from the channel and sort the most frequent topics by hand. Do not rely on intuition about what arrives.
- Second, for each topic, write the approved answer, the data it uses, and who is responsible if the answer is not enough.
- Third, define the transfer triggers: words or situations that require a person, such as a cancellation request, a price question, a formal complaint, or a message repeated without a solution.
- Fourth, set the exchange limit and the maximum time without a human reply, and tell the customer honestly that the conversation has been passed to the team.
- Fifth, configure the handoff with a summary of the case, so the person does not have to ask the customer to repeat everything.
- Sixth, test with real cases and with invented cases that try to push the agent outside its scope. OWASP's guidance on prompt injection describes exactly this kind of attack, in which the user tries to make the model ignore its instructions.
One point that is often underestimated is the reply window of each channel. On WhatsApp, for example, the type of message you can send depends on the timing of the conversation, and those rules determine when the agent can reply freely and when it must use an approved template. This topic is covered in WhatsApp 24 Hour Customer Service Window: How an AI Agent Should Act.
What to log for every case
The log is what separates a controlled agent from a channel nobody can audit. For each conversation, the useful minimum includes:
- Channel, date and time, conversation identifier, and whether the message was public or direct.
- The topic the agent identified and the rule that was applied.
- The reply sent, with the information source used, when there was one.
- The reason for the handoff, when there was one, and the summary delivered to the team.
- Who took over, how long it took, and how the conversation ended.
- Team feedback when the agent's reply was wrong or incomplete.
That record serves three decisions. The first is adjusting the list of topics the agent answers. The second is finding content gaps, because many repeated questions exist only because information is unclear on the website or in the price list. The third is reviewing cases that may involve personal data, which requires extra care with retention and access. The log itself must respect data minimization: keep what you need to operate and review, not the customer's entire personal history.
Illustrative scenario: a furniture store posts a video of a built-in kitchen on Instagram. One follower comments asking how much an identical kitchen costs, and another asks for a discount to close the deal that week. The agent, configured for comments, replies to the first one saying the price depends on measurements and finishes, and invites the person to send a direct message with rough measurements. For the second, it does not address the discount at all. It publishes one neutral sentence, says the sales team will reach out, and creates a log entry with the comment, the profile, the topic, and the time.
The sales team receives the record and contacts the person during business hours. If, after several weeks, the store notices that dozens of comments ask the same thing about measurements, the problem is not the agent but missing clear content. The next rule may be publishing a reference table of measurements and finishes, and the agent then answers using it. Adjustments like this only become visible with the log.
How to validate before widening the audience
The safest way to roll out is to pick one channel, one topic type, and one success criterion that the team agrees on before starting. The criterion can be simple, such as the share of conversations resolved without handoff within approved topics, or the time to the first human reply in handed-off cases. What matters is that the number is measured the same way throughout the pilot.
Our approach follows four steps: map the current process and the baseline, limit data and actions, pilot a real flow with logs and an agreed metric, and then hand over the operation and decide whether the scope should grow, be adjusted, or stop. This is not a promise of results, but a sequence of decisions that avoids expanding before you understand what the agent does. To test rules and replies without exposing customers, a simulation environment helps, as described in Mock tools for testing AI agents in a safe sandbox.
It is also advisable to restrict the agent's access by permission. It should query only what a given topic needs and have no access to full customer records. The principle of least privilege is a good test to review these permissions with your technology team.
Privacy and LGPD in direct messages
Direct messages often contain personal data, sometimes sensitive, such as addresses, health information in clinics, or financial details. Under the LGPD, the company needs a clear purpose for processing, must apply minimization, and must define who can access the record. If the agent makes a decision that significantly affects the customer, such as refusing a service based on an automated criterion, the topic also requires attention to the right to review, covered in LGPD Automated Decision-Making: The Right to Review (Art. 20).
In practice, the safest rule is that the agent does not make final decisions about people. It can inform, collect the data needed to serve the request, classify, and hand off. The decision to grant, deny, or change something should stay with a person, with the record available for review. If the operation involves sending data to model providers outside Brazil, the review must also cover what the LGPD requires in those cases, covered in International Data Transfer Under LGPD When Using AI Providers. The National Data Protection Authority (ANPD) is the official reference for guidance on the subject.
Finally, the NIST AI Risk Management Framework offers a useful structure for mapping risks before the agent goes live, including who is accountable for errors and how they are detected. The question of when to delegate to an agent and when to keep the decision with leadership is discussed in Superintelligence for Founders: What to Decide Yourself and What Not To.
Frequently asked questions
Can the agent reply to public comments without approval? It can, as long as the list of topics is small, the replies are pre-approved, and there is an alert for risky cases such as complaints or mentions of fraud. To start more cautiously, many teams let the agent draft the reply and approve it before publishing. The right choice depends on volume and on how much public error the company can tolerate.
How do you know the exact moment to hand off to a person? Use clear triggers: an explicit request to speak with someone, a mention of price or refund, a complaint, a message left unresolved after a few exchanges, or any sign of urgency. Also set a time limit without a human reply and inform the customer that the conversation has been passed on. Review the triggers using the log from the first few weeks.
What should each conversation record contain? The record should include the channel, time, topic identified, rule applied, reply sent, reason for handoff, summary given to the team, who took over, and the outcome. These fields let you measure quality and adjust the rules. Keep only what is necessary, in line with the purpose and minimization principles of the LGPD.
How long does it take to expand the agent to other channels or topics? The timing depends on volume, integration complexity, and how disciplined the pilot is, so it should not be set by a fixed calendar. Expansion makes sense when the pilot meets the agreed criterion and the team confirms that handoffs are appropriate. If results are not satisfactory, the right decision may be to adjust the scope or stop.