
AI Agent for Libraries: Loans and Reservations With Clear Limits
How an AI agent can handle library reservations, loan renewals and catalog questions, while logging fines, losses and exceptions for staff to decide.
Want to apply this in your company?
Book a call with Juan →Superintelligence in a library, in practice, is not a system that replaces the staff. What you can run safely is an AI agent with a narrow scope, one that handles reservation, renewal and catalog requests within rules the library writes itself. Anything outside those rules, such as a disputed fine, a damaged book or a declared loss, should be logged and sent to a human decision.
Key takeaways
- Superintelligence is a term for systems far beyond what exists today. For a library, the useful question is which AI agent, with which permissions, runs each workflow.
- Reservations, renewals and catalog lookups are repetitive workflows and good candidates for automation, as long as the rules are written down.
- Fines, losses, damage and exemption requests should leave the agent and enter a decision queue with a complete record.
- Before any agent, the catalog, item statuses and loan policies need to be standardized.
- The pilot should cover one real workflow, with logs and a metric agreed in advance. Only then do you decide whether to expand, adjust or stop.
What superintelligence means in a library
The term comes up in debates about systems with capabilities far beyond human ones, and that debate has its place. For a library manager, though, the useful discussion is more concrete: which steps of service can be done without a person, which need a clear rule, and which need a librarian. If you want to separate the concept from the tool, read Is ChatGPT superintelligence? What it is and what it is not.
A good design starts with the list of decisions the library makes every day. Some are mechanical, like confirming that a title is available. Others depend on policy, like accepting a renewal with an outstanding balance. Others depend on judgment, like deciding whether a returned book with signs of use counts as damage. The agent handles the first kind, applies the second when the rule is written, and routes the third.
How to run reservations without promising what does not exist
The reservation queue usually creates more friction than volume. The reader wants to know when the book will be available, and staff want to keep the queue from turning into confusion. The agent can record each request in order of arrival, notify the reader when the copy is returned, and cancel the reservation after the pickup window the library defines.
The care point is not to state availability the system does not confirm. Copies in transit between branches, lost copies or copies not yet checked into the catalog produce wrong answers when the agent reads only the title and not the copy status. So the minimum rule is this: if the status is uncertain, the agent says the information is being checked and passes the case to staff.
Renewals: what the agent can decide and what it cannot
Renewing a loan looks simple, but it involves three checks: whether the copy is not reserved by someone else, whether the reader has no registered restriction, and whether the library policy allows renewal in that situation. If all three conditions are met, the agent can renew and confirm the new due date. If any one fails, the correct answer is the rule, not an improvised exception.
The library needs to define in advance what happens with an outstanding balance, when a reader asks to renew after the due date, or when a copy is rare. The agent should not invent criteria in those moments. It applies what is written and, when the policy does not cover the case, logs it and escalates.
Catalog lookups with minimal data
Searching the catalog is the safest workflow to start with, because it mostly involves public information: title, author, location, branch and availability. The care point here is the source. The agent should answer from the catalog, say when information may be out of date, and never suggest a book that is not in the collection.
Personal data is also involved, such as loan history and reader identification. Data protection law (LGPD) applies to that data, and the principle of minimization should guide the design. The agent asks only for what the service needs, does not expose one person's history to third parties, and does not keep conversations longer than the library can justify. The national data protection authority (ANPD) publishes official guidance on the topic.
Exceptions for staff: fines, losses and damage
This is the core of the proposal. Disputed fines, declared losses, physical damage, exemption requests and loans outside policy should not be resolved by the agent. It should collect the case clearly, record the facts and give the reader an honest answer about the next step. Decisions that significantly affect the reader deserve human review. The topic parallels the right to review under LGPD, covered in Automated decisions under LGPD: Article 20 and the right to review.
A useful exception record contains at least:
- date, time and channel of the message;
- identification of the copy and the related loan;
- objective description of the reason, with the reader's own words preserved;
- available evidence, such as system dates or photos sent by the reader;
- the policy rule that applies, when there is one;
- the person responsible for the decision and the date of the decision;
- the final response sent to the reader.
The agent records the case and explains the next step. The decision on the exception stays with staff.
An illustrative scenario makes the difference clear. Suppose a reader says they returned a book before the due date, but the system shows it as overdue. The agent does not change the fine or confirm the return. It records the copy number, the return date the reader gave, the original message and the return date in the system, then replies that staff will review the case within the timeframe the library states. The decision belongs to the librarian or circulation lead, with the evidence at hand.
How to run a safe pilot in a library
The safest way to start is a pilot with a single workflow, in a controlled setting, before opening the service to the public. In our way of working, the path is map, limit, pilot and hand over. Map the current process and the baseline. Limit the agent's data and actions. Pilot a real workflow with records and an agreed metric. Then hand over the operation and decide whether the agent should be expanded, adjusted or switched off. This sequence keeps the technology from arriving before the rule.
A practical pilot outline:
- Choose one workflow, preferably catalog lookups or reservations, because it is easier to audit.
- List the answers the agent may give on its own and the ones that need routing.
- Define who receives exceptions, through which channel and with what response time.
- Test with real cases and with constructed cases meant to provoke errors, such as renewal requests outside the rules.
- Track the records for a few weeks and compare them with the previous service, using a metric agreed beforehand.
- Decide, based on the records, whether the workflow advances, whether the rules need adjustment or whether automation should stop.
For the security layer, apply least privilege. The agent accesses only what the workflow requires, and no write action, such as changing a fine or checking in a copy, should be available without approval. The OWASP authorization guidance is a useful reference for setting these limits. Preparing data before the pilot is covered in Data ready for superintelligence: what to standardize first.
Frequently asked questions
Can AI decide on library fines? It should not. The agent can explain the fine policy and log a dispute with its evidence, but the decision to cancel, reduce or keep the fine stays with staff. This separation protects both the reader and the library, because every decision has a named owner.
Can the agent tell a reader that a book is on the shelf? Only when the system confirms the copy status. If the catalog is out of date or the copy is in transit, the correct answer is to say the information is being verified. Claiming availability without confirmation creates frustration and extra work at the desk.
Do we need to replace our library management system to start? Usually not. The first step is to map the process and standardize the rules and statuses, which can be done with the current system. Changing tools only makes sense once the design is clear and the team has identified a concrete limitation.
How do we know if the pilot worked? Before starting, the team should agree on a simple metric, such as time to first response or the share of cases resolved without misrouting, and review samples of the records. If the metric improves without more errors in exceptions, there is a basis to expand. If errors appear, the path is to adjust the rules or stop the workflow.