AI Check-In Kiosks: Privacy Limits to Set Before Superintelligence
Keep check-in kiosks away from raw face images, full ID scans and idle camera footage. Set retention and notice rules before any smarter AI is added.

An AI check-in kiosk should collect only what the visit needs: a booking match, a code entered by the guest, and a logged decision that does not carry the underlying identity data. It should never keep raw face images, full identity document scans or continuous lobby camera footage, and every retained record should have a stated purpose, a deletion date and a notice the visitor reads before using the screen. These limits matter more as the system gets smarter, because a capable agent can connect data that a simple kiosk never could.
Key takeaways
- A check-in kiosk needs a booking match, not a copy of the guest's identity. Keep the decision and drop the raw data.
- Do not collect raw face images, full identity document scans or continuous camera footage for the kiosk's own memory.
- Give every retained record one stated purpose, one owner and one deletion date.
- Tell visitors what the kiosk does, what it keeps and how to reach a person, before they use it.
- Add intelligence only after the collection limits are written, tested and reviewed.
Why a smarter kiosk raises the privacy stakes
A basic kiosk reads a code, prints a badge and forgets the visit. An agent-backed kiosk does more. It can interpret a spoken request, look up a reservation in several systems, recognize a returning guest and write notes that a later agent will read. Each capability is useful, and each one creates another place where identity data can accumulate. The risk is rarely one dramatic leak. It is a slow drift: a camera added for a fault check starts feeding a visitor profile, or a note about a late arrival quietly becomes a record of who came in with whom.
A more capable system will also find more uses for data than the people who configured it had planned. That is why the privacy limits have to be decided by people before the capability is switched on, and written as refusals rather than hopes. The NIST AI Risk Management Framework treats privacy as part of trustworthy design from the start. A kiosk is a good place to apply that discipline, because the data is personal, the setting is public and the visitor has little time to object.
Identity data a check-in kiosk should never collect
Start by listing what the kiosk does not need. For most check-in flows, the decision is whether this person matches a booking, a delivery, an appointment or a permitted visitor. That decision rarely requires a copy of an identity document, a biometric template or a face image kept after the decision is made.
Do not collect or store:
- Raw face images or video frames of the visitor, including thumbnails generated for a dashboard or a log.
- Biometric templates or face embeddings, unless a specific legal basis, a documented necessity test and a separate approval exist. Treat that as a different project, not a setting.
- Full identity document scans, including the photo page of a passport or driver's license. If verification is required, keep only the result and the method used.
- Full national identification numbers, tax numbers or dates of birth when a booking code or a staff confirmation would do.
- Free-text notes about appearance, mood, health or apparent relationships between visitors. A kiosk agent should not infer these at all.
- Location trails from the visitor's phone, unless the visitor opted in for a specific purpose and can switch it off.
Pair this with a field-level review. Each field on the kiosk form should have a named reason. If nobody can name the reason, the field goes. Minimization is easier at the start than deletion later. Our guide on data minimization for AI agents walks through the field-by-field method.
Camera rules: keep footage out of the agent's memory
Cameras are where kiosk privacy most often goes wrong. A camera installed for a legitimate reason, such as safety, queue measurement or equipment monitoring, can quietly become an identity source. Set three rules before anything is connected.
First, the agent does not receive the live feed. If a vision function must run for a specific purpose, such as counting people in a queue, it runs on the device or in a service that returns only the event, not the frame. Second, no frame is stored by default. If an incident needs review, a named person opens a time-limited clip under a logged reason, and the clip is removed when the review closes. Third, the camera's purpose is written on a card beside the device. A camera with no stated purpose is switched off.
Logs need the same discipline. Transcripts of what a visitor said, and records of what a sensor reported, can contain names and sensitive details. The PII redaction guide for agent logs explains how to keep audit detail without copying personal data into every record.
Retention and deletion rules, step by step
Retention is where a smarter system tends to keep more, because memory is useful. Decide the periods before the first visitor arrives, using this sequence:
- Name each record type the kiosk creates: check-in result, visit note, transcript, event log, incident clip.
- Write one purpose and one owner for each type. A purpose such as 'improve the service' is too broad and needs a narrower statement.
- Set a retention period that matches the purpose, with the shortest periods for anything containing personal details.
- Separate operational logs from personal records, so debugging never requires keeping identity data.
- Build deletion into the pipeline rather than a manual monthly cleanup. Test that a deleted record disappears from search, summaries and backups on the schedule you declared.
- Record who approved each period and when it will next be reviewed.
Memory needs the same rules. If the kiosk agent keeps a summary of recurring visitors to speed up check-in, that summary is personal data. The guide on deleting agent memory on request covers how to find a record, remove it and verify the removal.
Notice and what the visitor sees
A visitor should know what the kiosk does before they use it. The notice should be short, placed at the screen and available in the languages your guests speak. It should say what the kiosk checks, what it keeps and for how long, who can see it, and how to ask for a person instead.
Avoid notices that are hard to read or that bury the important part behind a link. A useful notice names the actual behavior: the booking code is matched against today's reservation list, no face image is stored, transcripts are deleted after a stated period, and a staff member can complete check-in at the desk. If a feature is not active, do not describe it as active. If a conversation is recorded, the visitor needs to know before they speak. Our guide on telling customers they are talking to an AI agent covers the disclosure questions that apply at this moment too.
Illustrative scenario: A coworking space adds an AI assistant to its entrance tablet. The first design only matches bookings. Later, a manager asks the assistant to flag members who often arrive late so the front desk can prepare. The team stops before building it, because that flag would create a long-term attendance profile with no notice and no retention rule. Instead, they add a same-day view that shows booked and actual arrival times for the current day only, which meets the front desk's need without a standing record.
How Be Human sets these limits before an agent runs
Be Human treats a kiosk as an operating context, not only a device. In a first engagement, we define the single check-in role, the fields it may read, the records it may write, the camera functions it may use (often none), the retention period for each record type and the notice text shown at the screen. Those decisions become configuration the agent cannot widen on its own, and any change to the collection list goes through the client's named approver. Managed operations then add recurring reviews of what the kiosk has stored, permission checks on every tool it can reach and incident response if a record appears where it should not. The boundary is written down, tested and reviewed on a schedule, so a more capable model does not enlarge it quietly. If you are deciding where a check-in agent could help without collecting more than it should, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Can a check-in kiosk verify identity without storing an ID scan? Often yes. The kiosk can check a booking code, a one-time code sent to the guest or a staff confirmation, and store only the result and the method used. If a document check is legally required, it should run in a process you can describe, with the scan handled by the verifying service rather than copied into your kiosk records.
Is face recognition ever appropriate for a check-in kiosk? It is a separate decision with a higher bar. Biometric processing usually needs a clear legal basis, a documented necessity test, a genuinely available alternative for the visitor and a retention rule. Treat it as a new project with its own review, and do not enable it as a feature inside a general kiosk setup.
Should the AI agent keep notes about returning visitors? Only if you can state the purpose, the content, the retention period and how a visitor can ask for removal. Most check-in needs are met by same-day operational views rather than long-term profiles. If you do keep notes, they belong in the same inventory as any other personal data, with the same review and deletion rules.
How do we know the limits still hold after launch? Review stored records on a schedule, compare them with the written inventory, test a deletion request from start to finish and confirm that the kiosk's tools still match the permissions you approved. Any new field, camera function or memory feature should trigger the same review as the original design.
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.


