AI Speaking for the Founder: What to Automate and What Not To

AI Speaking for the Founder: What to Automate and What Not To

Learn which messages an AI agent can send in your name as the founder, and which decisions, apologies and sensitive conversations must stay with you.

•

Want to apply this in your company?

Book a call with Juan →

AI can answer in the founder's name for routine messages, such as confirming an appointment, sending an address, reporting the status of an order, or repeating a policy that has already been approved. It should not decide exceptions, promise compensation, or apologize for serious failures, because those points require personal accountability. The practical rule is simple: known information can be automated, while decisions, commitments and apologies need a person with a name and the authority to back them.

Key takeaways

  • Routine messages with a defined answer can be automated, as long as they have a clear source and written limits.
  • Exceptions, discounts, non-standard deadlines and promises are decisions for the owner or for whoever holds the authority to decide.
  • Apologies for failures, losses or conflicts should come from a person, because acknowledging the mistake is part of what the message must say.
  • The agent needs to know when to stop, hand the conversation to a human, and record why it did so.
  • Before putting an agent in charge of speaking for you, map the process, limit data and actions, pilot a real flow, and define who reviews it.

Why the founder's voice cannot be treated as generic text

When a customer writes to a company, they are not just receiving an answer. They are receiving a signal about who stands behind what was said. A time confirmation sent by a system does not create an expectation of personal commitment. A sentence like "I take responsibility for the error and will resolve it" does, and the customer is right to hold the person who wrote it accountable.

For that reason, the first boundary is one of identity. The agent should not present itself as the owner, and it should not imply that it is the owner when someone asks directly. If a customer wants to know whether they are talking to a person or a system, the honest answer must be clear. This care also connects with what the LGPD (Brazil's data protection law) says about transparency in data use and about decisions based on automated processing. For the legal details of your situation, review the topic with your lawyer, starting with LGPD automated decisions: Article 20 and the right to review.

Which messages can be automated

Automation works well when the answer already exists, has been approved by the owner, and does not depend on judgment about the customer's specific situation. In general, these fall into that category:

  • Appointment confirmations and reminders, within the cancellation rules already defined.
  • Order status, service order or protocol updates, read from a reliable system.
  • Sending the address, opening hours, standard documents and usage instructions.
  • Answers to frequent questions, using approved texts and without new promises.
  • Collecting data for triage, with a notice that a person will review the information.

A useful test is to ask whether the answer would change depending on the company's relationship with that customer. If it changes, it is no longer routine. It has become a decision. For customer service in general, it helps to compare this with AI customer support: what to automate and what to keep human.

What the owner decides and does not delegate

Some topics should not be answered by the agent, even when it has access to the information needed. The list below is a starting point for your business:

  • Exceptions to a policy, such as an extended deadline, a discount or an exchange outside the rule.
  • Contract cancellation, refunds outside the standard and changes to conditions already agreed.
  • Any position on a dispute, a formal complaint or a threat of legal action.
  • Refusing service for a sensitive reason, such as an old unpaid balance or a customer's conduct.
  • Commitments to results, deadlines that depend on third parties, and guarantees not written into the contract.
  • Conversations about employees, internal incidents and information about third parties.

In these cases, the agent can log the request, gather the history and let the customer know that someone responsible will get in touch. It should not try to resolve the case or suggest a solution just to buy time. An improvised answer usually costs more than a delay of a few hours.

Apologies and explanations that must come from you

A scripted apology is almost always read as a way of avoiding the problem. Phrases like "we regret the inconvenience" do not say what happened, who was affected, or what will change. When there is a real failure, the message must name the error, accept responsibility, and state the next step with a date and a responsible person.

Whoever made the mistake has to say what happened and what will change.

The agent can help draft a message using facts that are already confirmed, but the final text should be reviewed and sent by the owner or by someone who can stand behind what is promised. If the explanation depends on something still under investigation, the correct message says exactly that, without filling the gaps with assumptions.

Illustrative scenario: delivery delay and a compensation request

Illustrative scenario: a custom furniture store tells a customer, through the agent, the new delivery date confirmed by production. The customer asks whether there will be any discount or compensation for the delay. The agent does not answer about compensation, does not suggest amounts, and does not say the request will be approved. It logs the question, attaches the order history, informs the customer that a manager will reply, and flags the case as a priority for the owner. A person decides on the compensation and writes to the customer. The agent handled its part: correct information, a complete record, and a handoff without losing context.

How to set the boundary in five steps

The line between what the agent answers and what the owner takes on should not live only in the heads of the people handling customers. It needs to be written down, tested and reviewed. A practical path:

  • List the messages that repeat most often in recent months and mark which ones already have an approved answer.
  • For each message type, define what the agent may state, what it must not promise, and which signals should trigger a handoff.
  • Write the handoff sentence and name who takes over each type of case, with a realistic return time or channel.
  • Test with simulated conversations, including pressure questions, attempts to obtain exceptions and confusing messages, before connecting the agent to your main number.
  • Review the decision logs every week during the first few weeks, and adjust the rules wherever the agent answered beyond what it should.

This is the method we apply in agent projects: map the process and its baseline, limit data and actions, pilot a real flow with records and an agreed success metric defined in advance, then hand over the operation and decide whether to expand, adjust or stop. The founder's-voice boundary belongs in the limits stage, and the pilot is where it shows up in practice.

How to record and review what the agent says

Every response sent on behalf of the company must be traceable. That includes the message, the source used, the rule applied and, when relevant, the handoff to a person. Without that record, there is no way to know whether the agent crossed the boundary or how to correct it. Specific risks also need attention in systems that use natural language, such as a customer trying to persuade the agent to promise something outside the rules. The OWASP list of risks for applications built on language models is a useful reference for discussing these points with your technical team. For governance more broadly, the NIST AI Risk Management Framework offers a structure for thinking about risk, accountability and human oversight.

In daily practice, the review can stay simple: a weekly reading of the transferred conversations, the replies that triggered complaints, and the questions the agent did not handle well. Each item found becomes a new rule, an approved text, or a clearer handoff.

Frequently asked questions

Can the agent present itself as the owner of the company? No. The AI should identify itself as an assistant when a customer asks, and it should not suggest that it is the owner or that the owner is replying personally. This prevents customers from forming expectations of a personal commitment that no one has made. The identity also needs to match the company's customer service policy.

Can AI write the apology I send? It can prepare a draft using facts that are already confirmed, but the final version should be reviewed and sent by a responsible person. An apology has to say what happened and what will change, and that depends on investigation and authority that a system does not have. When the investigation is still ongoing, the message should say so honestly.

How do I tell whether a message is routine or a decision? A message is routine when the answer is the same for any customer under the same conditions and has already been approved in writing. If the answer depends on an exception, a prior relationship, a dispute or a new commitment, it is a decision. When in doubt, the agent should hand the case over, because a handoff costs less than promising something the company cannot deliver.

Do I need to tell the customer they are talking to an AI? In general, yes, and transparency is the safest stance, both for customer trust and for compliance with the LGPD. The notice can be short, given at the start of the conversation, and repeated whenever the customer asks directly. For your company's specific case, confirm the wording of the notice with a legal professional.

Sources