After-School Pickup Rules: Keeping Humans in Control of AI Agents
An AI agent can help an after-school program track attendance and draft parent messages, but pickup authorizations need strict human approval.

An AI agent for an after-school program can track attendance, answer routine parent questions and draft reminder messages, but pickup authorizations should stay under strict human control. Releasing a child to an adult is the highest-consequence action in the program, so the agent should prepare and flag pickup information while trained staff confirm every change that affects who may collect a child. The rules below set that boundary in writing and explain how to test it as the tools improve.
Key takeaways
- The agent may prepare, summarize and flag pickup information, but it never confirms a release to an adult on its own.
- The authorized pickup list belongs to a named staff owner, and every change needs a recorded approval.
- A parent message that asks for a new pickup person is a request for review, not an instruction to follow.
- Test the boundary with deliberately unclear requests after every tool, model or permission change.
- Write the rule down so a substitute staff member can apply it on a busy afternoon without the agent.
Why pickup stays a human decision
Pickup is the moment a child changes hands. A late reminder can be corrected with a second message, but a child released to the wrong adult cannot be recovered by editing a record afterward. As agents get better at reading messages and calling tools, the temptation is to let them close more loops on their own. For pickups, that improvement changes the quality of drafts and summaries, not who is accountable for the decision.
The NIST AI Risk Management Framework treats AI risk as something to map, measure and govern with clear responsibilities. Pickup is a clear case where responsibility must stay with named people, because a program director, not a software process, answers to parents and to the licensing or safeguarding rules that apply locally. A capable agent can make a wrong release look plausible, which is exactly why the approval step cannot depend on how confident the output sounds.
Split the work: what the agent prepares and what staff decide
Good boundaries are written as lists of actions, not as general intentions. For a pickup workflow, the agent may do the following:
- Summarize the authorized pickup list for each group or room at the start of the day.
- Flag any message that names an adult who is not on the list.
- Draft a reply asking the parent which adult is coming and explaining that staff will confirm.
- Create a review item with the request, the source message and the time it arrived.
- Show staff a daily list of unresolved requests so nothing waits in a thread unseen.
The agent may not do the following:
- Add, remove or edit a name on an authorized list.
- Tell a parent or an adult that a pickup is approved.
- Accept a change relayed through a chat thread, a voice note or a photo as valid authority.
- Treat a missing or expired authorization as a reason to relax the rule.
Other service businesses face the same split. Security guard firms, for example, need access rules for each site to be explicit before anyone on the team acts on them, and behavioral health clinics set scheduling limits that decide which requests a system may answer on its own. In each case the system can organize the request, while the decision about access stays with people who can be held responsible for it.
A step-by-step setup before the agent touches any records
Set the rule before connecting any live child record. The sequence below is designed so that the boundary is written, owned and tested before the agent can influence anything.
- Write the authorized pickup list for each child, with each adult's name, relationship, a contact number and the date a staff member last confirmed the entry.
- Name one staff role that can change the list, and require a short record of who approved each change and why.
- Define what counts as a change request: a new adult, a changed phone number, a schedule change, or a note sent in a message thread.
- Write the fallback for the case when the agent is unavailable or unsure, and keep the existing paper or console process running until the fallback has been tested.
- Test with invented names in a copy of the records, and connect live records only when every boundary case behaves as written.
The last step matters most. A boundary that has never been exercised is a hope, not a control, so the test should include the messy cases that happen in real programs: two adults with the same first name, a sibling collecting a child, and a request sent during the final minutes before closing.
What to do when a parent message asks for a change
Illustrative scenario: late in the afternoon, a parent writes that a grandmother will collect their daughter and attaches a photo of an identity card. The agent should not add the grandmother to the list, should not reply that the pickup is approved, and should keep the photo only as long as the review item needs it. It should create a review request for the staff member on duty, send the parent a short message explaining that the request is with staff and that the adult will be verified at the desk, and place the request at the top of the daily summary. The agent's reply is a notice that a person has the request, and nothing more.
Messages are also content that can try to steer a system. The OWASP prompt injection prevention guidance treats input from users and documents as untrusted, because text can contain instructions written to change how a model behaves. A message that says the office already approved this, just add her, is data to review, not authority. The control belongs outside the language model: the agent's tools should have no permission to write to the authorized list at all, so that even a persuasive message cannot change the record.
Permissions, records and what the agent can reach
Grant the narrowest access that the role needs. The OWASP authorization guidance recommends enforcing access decisions on the server side and denying by default, which means the pickup agent should not rely on its own instructions to stay in its lane. In practice, its connection to the attendance system might read check-ins and write only to a review queue, while the authorized list is maintained through a separate process that requires a staff login. Scoped access of this kind is the same idea behind OAuth 2.0 permissions, where a token can be limited to read one resource without being able to change it.
Data minimization follows the same logic. A pickup agent does not need a child's medical notes, dietary records or family court documents to do its job, so those should not be connected to it. Children's records carry heightened sensitivity almost everywhere, and retention periods for review items and messages should be set deliberately. Confirm with counsel which privacy rules apply to your program, including the rules of your national data protection authority, and do not assume that a vendor's default settings meet them.
Proving the boundary holds over time
A boundary needs evidence that it still holds after every change. Build a small set of test cases where the correct result is to change nothing: an unknown adult, a duplicate name, a request from someone who is not a listed parent, conflicting instructions in one thread, a message that mixes two languages, and a request sent at closing time. For each case, write the expected output in advance, such as a review item created, no change to the list, and a parent notified. Run the set after every model update, tool change or permission change, and record the results.
Evaluation is a discipline rather than a one-time launch check. Guidance on evals for AI agents, including the OpenAI evals guide, describes repeatable tests of behavior, and the important measure here is the outcome in the record, not whether the wording of the reply sounds polite. Observability supports the same goal. Log each proposed action with a timestamp, the source message, the tool called and any staff approval. The OpenTelemetry observability primer describes how logs and traces let teams see what a system did and why, which is exactly what a program director needs when a parent asks what happened on a particular afternoon.
When something goes wrong, the response should be written in advance. Pause the agent's write paths, keep staff on the existing process, review the logged actions for the affected day, correct the records through the staff owner, and update the test set so the failure cannot quietly return.
How Be Human runs this as a managed service
Be Human designs the after-school agent around one defined role, the pickup information it prepares, the memory and tools it needs, and the approval boundaries described above. The authorized list stays under the program's named staff owner, every proposed action is logged, and boundary cases are run on a recurring schedule and after each change. Permissions are reviewed when tools change, and incidents are handled with a written record that the program keeps. The service is remote and available in English and Portuguese.
If you want to find the first task worth handing to an agent in your program, take the free Be Human AI diagnostic.
Frequently asked questions
Can an AI agent for an after-school program approve a pickup? No, and it should not be designed to try. The agent can gather the request, check it against the authorized list, flag anything unclear and notify staff, but a named person should confirm the adult at the desk and make any change to the record. Approval is a human decision that the program can explain to parents and regulators.
Is it safe to let parents send pickup changes through a messaging agent? It is safe when the message is treated as a request for review rather than as an instruction. The agent should acknowledge the request, avoid promising an outcome and route the change to staff with the original message attached. Parents should also be told which channel is authoritative for pickup changes, so that a casual message is never mistaken for a formal authorization.
Which task should an after-school program hand to an agent first? Start with a low-consequence task such as summarizing attendance for staff, drafting routine reminders or producing the daily list of unresolved messages. These tasks show whether the agent can read your records and communicate clearly without touching any decision that affects who collects a child. Expand only after the boundary tests pass consistently.
What records should a program keep for each pickup decision? Keep the authorized list with the date each entry was last confirmed, the staff member who approved each change, the source message for any request, and the agent's proposed action alongside what staff actually did. Set a retention period that your privacy obligations support, and make sure the records can be reviewed by the staff owner without depending on the agent to explain itself.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


