Superintelligence for Cinemas and Theaters: Groups, Sessions and Refunds

An AI agent for cinemas and theaters is an automated assistant that answers questions about sessions, rooms and accessibility, organizes group requests, and applies the refund rules the venue has already set. Anything that depends on a financial exception or on judgment goes to a person, with a summary of the conversation. The result depends less on the technology and more on three decisions: what the agent may answer, what it only records, and who approves each refund.

Key takeaways

• The agent answers official information: showtimes, duration, age rating, rooms, accessibility, and accepted payment methods.
• Group requests follow their own flow, with data collection, an availability check, and human approval for any condition outside the published table.
• Refunds and exchanges need written rules before the agent exists, with defined exceptions and a named owner.
• Every conversation is logged, and the handoff to a person includes a summary so the customer does not have to repeat everything.
• The pilot starts with one flow, such as session questions or group bookings, with an agreed metric defined before launch.

What the agent should answer and what it should pass on

Before configuring any message, we sort customer questions into three types. The first is public information: showtimes, duration, location, parking, age rating, and accessibility policies. This information should come from a single source, maintained by the box office team. The second type is actions, such as reserving seats, changing sessions, or requesting a refund. These require a connection to the sales system or a formal record for the team. The third type is judgment: exceptions, complimentary tickets, half-price eligibility in unclear cases, complaints about service, and requests for compensation. This third type stays with people.

Illustrative scenario: a moviegoer writes that they missed the session because of a health emergency and asks for a refund after the usual deadline. The agent neither denies nor approves the request. It confirms the ticket details, records the reason given, attaches the purchase number, and forwards the case to the box office with a summary. To the customer, it says the team will reply through the same channel, without promising an outcome.

How to organize group bookings

Group bookings involve the number of people, the date, the session, a block of seats, the payment method, and a responsible person who takes over the purchase. For that reason, they should not be handled like an ordinary chat. A practical flow follows this sequence:

• Collect the name of the responsible person, the desired date, the session, the number of people, and the type of group, such as a school, a company, or an association.
• Check availability in the box office's official source before confirming any seats.
• Apply only the discount bands already published. Any condition outside those bands goes to human approval.
• Send the responsible person a summary with the amounts, the payment deadline, the cancellation rule, and what happens if the group changes size.
• Keep the request pending until payment or signature of the agreement, without holding seats permanently on its own.

The most sensitive point is usually a change in group size after confirmation. The rule must state what can be adjusted without new approval and from which point a change requires the team.

Refund and cancellation rules written before the agent

An agent does not create policy. It applies a policy that already exists. For that reason, the first task of the project is to write, in plain language, what happens in each situation. A minimum decision list usually covers these cases:

• Session canceled by the venue: refund through the original purchase channel, without the customer having to justify anything.
• Unused ticket, with the session not yet started: the agent explains the official return path and records the request.
• Request for health, transport, or other emergency reasons: the agent records the case and passes it to a human decision.
• Purchase made through a third-party channel: the agent states that the return rule belongs to the sales channel and gives the correct contact.
• Online purchase: the withdrawal period and return conditions must follow Brazil's Consumer Protection Code (Código de Defesa do Consumidor). The company's legal team confirms how this applies to the specific case, and the text used by the agent must match what appears on the website and at the box office.
• Compensation request or formal complaint: immediate handoff to a responsible person.

Consistency matters as much as content. If the website says one thing, the box office says another, and the agent gives a third answer, the customer notices the contradiction and trust drops. For that reason, the rule should have a single reference text, reviewed by whoever is accountable for operations.

Sessions, rooms and accessibility

Most day-to-day messages are simple: which session starts earliest, whether a room has wheelchair access, whether a play has subtitles or audio description, or whether there is space for a person with a service dog. These answers are useful and fast, as long as they come from a table that is kept up to date with a cutoff time. After the cutoff, the agent should say that the information may change and indicate where to confirm it.

There are two extra precautions. The first is not promising anything that has not been confirmed, such as availability of an audio description session that does not yet have a date. The second is respecting the channel: if the conversation happens on WhatsApp, the team must know the rules for messages the business starts and the customer service window. This topic is covered in detail in WhatsApp 24 Hour Customer Service Window: How an AI Agent Should Act.

Personal data, security and LGPD

To buy a ticket or request a refund, the venue collects data such as name, phone number, email and, in some cases, a national tax ID (CPF). The principle that should guide the project is necessity: the agent asks only for what the purpose requires and does not store more than it needs. The team also has to define how long conversations are kept and who can access them. Brazil's National Data Protection Authority (ANPD) publishes guidance on its portal, and it is worth reviewing before defining the data flow.

Security is another point. The agent reads messages that may contain malicious instructions, such as someone asking it to ignore the rules and apply a discount. For that reason, the agent should not have permission to change prices or approve refunds on its own. The OWASP project documents defenses against this kind of instruction injection.

Handoff to people with context

The handoff is part of the product, not a failure. A good transfer to the team includes the reason for the escalation, a short history of the conversation, the data already confirmed, and the exact question that needs a decision. This way, the person handling the case does not start from zero, and the customer does not repeat the order number three times. The guidance to keep human checkpoints in flows with sensitive decisions also appears in Anthropic's documentation on agents, which favors simple, controllable patterns over systems that act without supervision.

How to pilot without promising more than the flow delivers

The method we use has four steps: map the current process, limit data and actions, pilot a real flow with logs and an agreed metric, and then hand over the operation and decide whether to expand, adjust, or stop. In practice, for a cinema or theater, the pilot usually starts with session questions, because the risk is low and the volume is high. Group bookings come later, once the discount table and the cancellation rule have been tested.

Metrics should be simple and agreed before launch: how many conversations were resolved without an information error, how many were passed on correctly, how long the team takes to reply to transferred cases, and how many complaints came up. Before putting the agent in front of customers, test the flow in a controlled environment, including the edge cases. The guide on mock tools for testing AI agents in a safe sandbox shows how to set up that testing.

Frequently asked questions

Can the agent approve a refund on its own? In most cases, it is not advisable. It can apply a rule that is already written and explain it to the customer, but exceptions, health reasons, and compensation requests should go to a person who records the decision.

Is this superintelligence for cinemas? No. An customer service agent for a box office is software that follows rules, checks defined sources, and passes complex cases to people. The term superintelligence refers to a research concept and does not describe a ticket sales tool. If you want to understand the difference, the article on whether ChatGPT is superintelligence, what it is and what it is not helps separate the concepts.

How long does it take to launch an agent like this? It depends on the ticketing system, the quality of the session table, and how much the venue already has its return rules written down. When those parts are ready, the pilot can start with a single flow. When they are not, the initial work is organizing the information, and that comes before any agent configuration.

Do we need to integrate the agent with the ticketing system? To answer general questions, an up-to-date information source is usually enough. To reserve seats, check availability in real time, or record refunds, integration is necessary, with permissions limited to the minimum the task requires.

Sources

• Anthropic: Building effective agents
• OWASP: Prompt injection prevention cheat sheet
• ANPD: Autoridade Nacional de Proteção de Dados