Aegis Vision

Evaluating AI Security Systems, a Guide for School Leaders

Vendor-neutral questions, red flags, pilot design, parent communication and a DPIA checklist for any AI camera analytics in a school.

This guide is for heads, bursars, governors, security managers and data-protection leads who are considering AI analytics on school cameras, from any vendor. It does not recommend a product. It sets out the questions that separate a well-designed system from a poorly designed one, how to run a pilot that produces evidence, and how to talk to parents and staff about it.

Start with the problem

Write down, in one paragraph, what you want to be true after the system is installed that is not true now. For most schools it is something like "a trained person is told within a short time when something serious appears on a camera, so they can go and look". That paragraph becomes the test. A system that does something impressive but different has not solved your problem.

Decide who will respond to alerts before you talk to any vendor. Without that person, the technology has nothing to attach to.

Questions to ask every vendor

Ask for written answers.

Data flow. Draw the path of a video frame from the camera to the alert. Which steps happen inside the school and which outside? What leaves the building, in what form, and to whom?

Where video goes. Is any video, image or clip stored or processed outside the school? If so, in which country, by which company, and under what contract? Can the system run with no internet connection, and what stops working if it does?

Who verifies alerts. Does a person review every alert before anything is escalated? Which person, employed by whom, and where are they? What can the system do on its own without a human confirming?

Accuracy evidence and test conditions. Show measured precision and recall per detection type, with confidence intervals, and describe the footage used. Was it from cameras not used in training? Who staged the scenes? Figures without conditions are not evidence.

False-alarm rates. How many false alerts per camera per day should my guards expect, measured where, over how long? What happens to a false alert, and does the system learn from dismissals?

Identification. Does the system identify, recognise or track individual people in any mode? Is facial recognition present in the software even if switched off? Who can switch it on?

Retention and deletion. How long are alert images and clips kept, where, and who sets the period? How is deletion carried out and confirmed? What happens to our data when the contract ends?

Sub-processors. Which other companies touch our data, for what purpose, and where are they? Will you tell us before adding one?

Incident response. If your software or your company is breached, how and when will you tell us? Who is accountable? Do you have a published vulnerability disclosure policy?

Exit and portability. If we leave, can we export our alert records in a usable format? Can we keep running the system without you for a transition period? What is deleted and when?

Updates. Who decides when the models in our building are updated, and are we told?

Red flags

  • Guarantees of complete accuracy. No detection system is correct every time. A vendor who says otherwise has either not measured it or not reported it.
  • Refusal to share methodology. You do not need the vendor's source code. You do need to know how they measured what they claim, on what footage, and under what conditions.
  • Cloud-only processing of video of minors without a data protection impact assessment. A vendor who has not considered the question is unlikely to have answered it.
  • Alerts routed through consumer messaging apps that your institution does not control.
  • Facial recognition or identity features bundled in by default when the problem you wrote down does not require them.
  • Pressure to sign before a pilot. The pilot is the evaluation.
  • Reference customers you cannot speak to.

How to run a pilot

A pilot should last long enough to include ordinary days and should produce numbers you can compare with the vendor's claims.

  1. Scope. Choose a small number of cameras that represent your real conditions, including at least one that is difficult. Agree in writing what is being tested and what counts as success, using the paragraph you wrote at the start.
  2. Baseline. Record how incidents are noticed and handled today.
  3. Staged tests. Have trained staff stage test scenes at agreed times, with safe props and with the knowledge of everyone on site, and record whether and when each was alerted. Do this in more than one condition and on more than one day. Do not test with pupils present.
  4. Live measurement. Count every alert your guards receive during normal operation and record whether they dismissed or confirmed it. This gives you the false-alarm rate on your own cameras.
  5. Operational fit. Ask the guards who used it. Did it help? Did they trust it? Were they still using it after the first week?
  6. Data check. Before the pilot ends, verify what data was created, where it sits, and that it is deleted as agreed if you do not proceed.
  7. Decide. Compare the results with the vendor's written answers. Any difference is a finding.

Communicating with parents and staff

Tell people before the system is installed, in plain language. Say what it looks for, what it does not do, who sees alerts, where the video is processed, how long anything is kept, and whom to ask. Say explicitly whether it identifies individuals. Offer a short written note and a way to ask questions, and record the questions you receive.

Staff should be told the same things as parents, plus what happens when an alert is raised, who responds, and what is expected of them. Security staff who will use the system need training on it, including the fact that it will produce false alerts.

Do not describe the system as a guarantee of safety. It is one part of a safeguarding approach that still depends on people.

A short DPIA checklist

A data protection impact assessment is required under the EU General Data Protection Regulation for processing likely to result in high risk, which includes systematic monitoring of a publicly accessible area (Regulation (EU) 2016/679, Article 35). In the United Arab Emirates, the Personal Data Protection Law (u.ae: Data protection laws) sets out the duties of controllers, and a documented assessment is good practice there as well. Wherever you are, the exercise is useful.

  • Purpose: what the processing is for, in one sentence, and why camera analytics is necessary for it.
  • Data: what is captured, about whom, including children, and what is derived from it.
  • Flow: where each type of data goes, including vendor and sub-processor locations.
  • Lawful basis and consultation: the basis relied on and who was consulted, including parents and staff.
  • Identification: whether any individual can be identified by the system, and if so, who, why and with what safeguards.
  • Human review: who reviews alerts, and what the system can do without them.
  • Retention and deletion: periods, responsibility, and confirmation.
  • Security: access control, encryption in transit and at rest, breach notification.
  • Rights: how a parent, pupil or staff member can ask what is held and request deletion.
  • Risks and mitigations: false alerts, misuse of footage, function creep, vendor failure, and what you have done about each.
  • Review date: when the assessment will be revisited.

Keep the vendor's written answers with the assessment. They are part of the record.

One note on neutrality. This guide is published by Aegis Vision, which sells a system of this kind. The questions above apply to it as much as to anyone else.

Updated 2026-11-16