AI footfall cameras count people crossing a line at the door, then add dwell time, queue length and busy zones, and some estimate age or gender. The privacy test is what leaves the device: numbers only, or video and faces. Before buying, check storage, signage, staff use and any biometric feature with a data-protection adviser.
Counting is not identifying. A camera that outputs "37 people entered between 2 and 3 pm" is a different product from one that stores faces to spot repeat visitors, yet many vendors sell both on the same hardware, so the contract and the settings decide which one you have. Decide the question the numbers must answer before you look at a single camera.
What a footfall camera counts, from mild to intrusive
Vendors name these levels differently, so ask to see the actual output at each one rather than relying on the product label.
| Level | What it counts | What the output looks like | Privacy weight |
|---|---|---|---|
| 1. Door count | People entering and leaving, by crossing a line | "Entries this week: 1,260" | Low, if only totals are kept |
| 2. Occupancy | People inside right now (entries minus exits) | "14 inside at 2:10 pm" | Low |
| 3. Capture rate | Passers-by who enter, against all passers-by outside | "9% of passers-by came in" | Medium: the sensor looks at a public pavement |
| 4. Zones and dwell time | Where people stand and for how long | A heat map; "4 minutes average at the display table" | Medium: follows anonymous movement paths |
| 5. Queue length | People waiting at the till or counter | "5 waiting, about 6 minutes" | Medium |
| 6. Demographic estimates | Age band or gender guessed from the face | "62% women, mostly 25 to 44" | High: processes facial features |
| 7. Identification | Recognising one person again: repeat visitors, staff, watchlists | "Visitor 4471 returned" | Highest: needs a face template or similar |
Levels one to five can be delivered as numbers with nothing recognisable stored, if the product is built and configured that way. Levels six and seven need the system to process facial features. That is the line between counting and identifying. Every level still involves a device that sees people, so the real question is what leaves the device and what is kept.
Match the device to the question you need answered
Start with the question, not the product. A week or two of hand counts at the door will show whether you need a camera at all, and it gives you a baseline to test any sensor against.
| Question | Least intrusive way to answer it | What to avoid |
|---|---|---|
| How many people visit each hour? | A door counter: a beam, a thermal sensor, or a camera set to count only | A camera recording the whole shop floor |
| Is the till queue too long on busy days? | A queue counter at the till that outputs a number, or a ticket system | Face recognition for "VIP" customers |
| Which display holds attention? | Zone dwell time from an overhead sensor, or a week of manual observation | Age and gender profiling of whoever stops |
| What share of passers-by come in? | A door counter plus a one-week hand count outside | A camera filming the pavement all day |
| Do the same people come back? | Loyalty-card or booking data that customers agreed to share | Face templates stored by a vendor |
Counters that listen for phones' Wi-Fi or Bluetooth signals use device identifiers rather than faces. That sounds harmless, but identifiers like these can count as personal data in some regimes, so put the same questions to that vendor. The tutorial on convenience store AI cameras covers the security side of cameras, including cost.
Counts or identification: three questions that settle it
Almost any product falls on one side or the other once you ask these:
- Where does the analysis happen? On the device, often called edge processing, means the footage is analysed inside the camera and only numbers leave it. In the vendor's cloud means video is streamed out first.
- What is stored, and for how long? Nothing; a short rolling buffer kept for troubleshooting; or full recording.
- Can the system recognise a person? Face templates, visitor IDs, "returning customer" flags, staff recognition and watchlists all count. If any of them is on, the system identifies.
Personal data means information about an identifiable person. Video in which people can be recognised counts in most data-protection regimes. A total such as "37 entries between 2 and 3 pm" is not about anyone, but the feed that produces it might be. If the answers are on the device, numbers only and no recognition, the privacy work is mostly signs, access control and vendor terms. If any answer differs, treat the installation like CCTV and take advice before you sign.
Two extras. Switch audio off unless you have a specific, advised reason, because recording sound is often treated as more intrusive than recording video. And remember that in many places people can ask for footage of themselves: a system that stores video needs a process for that, while one that stores none has nothing to hand over.
Turning visitor counts into a better Saturday rota
These figures are illustrative, for a 90-square-metre homeware shop open 10 to 5 every day with two people on the floor, so 14 staff-hours a day. A door counter gives visitors and the till gives sales. Conversion is sales divided by visitors; visitors per staff-hour is visitors divided by those 14 hours.
| Day | Visitors | Sales | Conversion | Visitors per staff-hour |
|---|---|---|---|---|
| Monday | 120 | 11 | 9.2% | 8.6 |
| Tuesday | 130 | 12 | 9.2% | 9.3 |
| Wednesday | 150 | 14 | 9.3% | 10.7 |
| Thursday | 160 | 14 | 8.8% | 11.4 |
| Friday | 210 | 18 | 8.6% | 15.0 |
| Saturday | 340 | 21 | 6.2% | 24.3 |
| Sunday | 150 | 12 | 8.0% | 10.7 |
| Week | 1,260 | 102 | 8.1% |
Weekday conversion sits between 8.6% and 9.3%, then drops to 6.2% on Saturday, when each staff hour faces 24 visitors. Friday, at 15, shows the first slip. On this made-up data, conversion holds until visitors per staff-hour reach the mid-teens.
The hourly profile shows where on Saturday. Visitors arrive as 25 from 10 to 11, then 55, 70, 65, 55, 40 and 30. The three hours from 11 to 2 carry 190 of the 340. A third person from 11 to 3 (four hours) would cover 245 visitors at their busiest. Pay for it by ending the second shift two hours early on Tuesday and Wednesday, which each see about 20 visitors after 3 pm in this example. Total hours stay the same. In the 11-to-2 window, visitors per staff-hour fall from 31.7 to 21.1.
If Saturday conversion rose from 6.2% to the weekday level of about 8.8%, that is roughly nine more sales, and at a $38 average basket about $340 a Saturday. Treat that as a hypothesis, not a promise. Run the new rota for four Saturdays and compare conversion and till queues with the four before. Saturday's low conversion could also reflect browsing families, which an extra person won't fix.
Two cautions. The raw count includes staff coming and going, delivery drivers and people sheltering from rain, so audit it before basing decisions on it. And keep the numbers at shop level: using them to rank individual staff turns a rota tool into monitoring, which changes your privacy position and your staff's trust. The tutorials on building a retail rota from hourly till sales and a weekly retail KPI sheet in Excel cover the spreadsheet side.
When a clicker beats a camera, and when it doesn't
The homeware example needed one number per hour for one door. A volunteer with a clicker on four Saturdays would have produced it, at the cost of four mornings. A camera earns its place when the count has to run every day, across several doors or long hours nobody can staff, or when you need queue and dwell figures that a person serving customers can't gather at the same time.
Ask each vendor for a three-year total covering the sensor, installation, dashboard subscription and support. Divide by 36 for a monthly figure, then set it against the staff hours the decision frees up. Say the total works out at $60 a month and a staff hour costs you $18 with on-costs: the system has to save or redirect about three and a third hours a month just to break even, before counting any extra sales. In the homeware example, moving four hours a week to the busiest slot is a rota change that costs nothing, so the camera has to give you something the clicker can't. Prices vary by product and by quote, so get the three-year figure in writing.
Check the count with a clicker before trusting it
Stand at the door with a hand clicker or a phone counter app for three 30-minute windows and note your entries against the sensor's for the same minutes. Choose a quiet window, a busy one, and one when groups arrive, such as a school run. Error is the sensor's count minus yours, divided by yours.
| Window | Your count | Sensor | Error | Likely cause |
|---|---|---|---|---|
| Quiet | 14 | 13 | -7% | Someone paused in the doorway |
| Busy | 46 | 39 | -15% | Groups of three or four counted as fewer |
| School-run groups | 31 | 36 | +16% | Prams, and a staff member going in and out |
My rule of thumb, not a standard: within about 10% is fine for rota decisions, and a steady bias is harmless for trends because the pattern still shows. An error that swings from minus 15% to plus 16%, as here, is not harmless. Ask the installer to adjust the mounting height, the counting line or the staff exclusion, then repeat the audit. Repeat it every quarter too, because a moved display or new lighting can shift accuracy.
Some mistakes show up in the hourly profile rather than the daily total. Imagine Monday reads 180 visitors against a usual 120. The hourly view shows 60 of them between 8 and 9 am, before opening, when staff arrive, unpack deliveries and prop the door open. The fix is to restrict counting to opening hours and to exclude the staff door. Look at the hourly profile for odd spikes before you trust the weekly totals.
Privacy checks in the order I would do them
- Write the purpose in one sentence. For example, "to set Saturday staffing and track conversion". Add what you won't use it for: judging individual staff, identifying customers, marketing. If you can't write the sentence, you're not ready to install.
- Get the data flow in writing. Where is video analysed, what leaves the device, what is stored, and who at the vendor can see it? A vendor that can't answer in a paragraph has not designed for privacy.
- Switch off what you don't need. Demographic estimates, face features, audio and cloud video are the usual extras. Screenshot the settings page on installation day so you can show later what was on.
- Set retention. Keep any video for the shortest period that lets you fix a fault or review an incident. Counts can be kept as long as they are useful. Ask your adviser what is reasonable where you trade.
- Put up signs. Wording and placement are covered in the next section.
- Tell staff before installation. Explain what the sensors do and don't do, that the numbers are for shop-level decisions, and who to ask. Where staff consultation rules apply, follow them.
- Check the vendor terms. You want a data-processing agreement (a contract saying the vendor handles your data only on your instructions), a list of sub-processors, no use of your footage to train its models, a promise to tell you about breaches, and deletion when you leave.
- Keep a one-page record and review it. Ask your data-protection adviser whether you need a written risk assessment, often called a data-protection impact assessment, and review the record whenever a setting or the vendor changes.
A filled-in record for the homeware shop, as an example of what to keep:
Footfall sensor record (homeware shop, example)
Purpose: set Saturday staffing and track conversion (sales divided by visitors).
Not used for: judging individual staff, identifying customers, marketing.
Devices: ceiling sensor at the front door; second sensor at the staff door (excluded from visitor totals).
Output: entries and exits per 15 minutes, as numbers.
Images stored: none (confirmed in vendor settings; screenshot saved [date]).
Switched off: age and gender estimates, face features, audio, cloud video.
Retention: counts kept 3 years; no video exists.
Logins: owner and shop manager, named accounts, two-step sign-in.
Vendor terms: data-processing agreement signed; no training on our data; deletion on exit.
Signs: both doors, wording checked with adviser [date], photographed [date].
Staff told: [date], at the team meeting; notes kept.
Review: every March, and whenever a setting or vendor changes.Signs that say what the sensor does
"CCTV in operation" says nothing about purpose or who to ask, and it is wrong if all you run is a counter. For counting-only sensors, something like this:
People-counting sensors operate in this shop. They count entries and exits as numbers only; no images are stored. We use the totals to plan staffing. Questions: [contact].
For cameras that record, the sign needs more:
CCTV is in operation in this shop to prevent and detect crime and to protect staff and customers. Images are kept for [period] and viewed only by [role]. Operated by [business name]. Contact: [contact].
Only write "no images are stored" if you have confirmed it in the settings and the contract. Put the sign at each entrance at eye level, and check the wording against the rules where you trade. The tutorial on AI-drafted CCTV privacy notices and signage shows how an installer approaches it.
What the AI Act says about emotion and biometric features
If you have customers in the EU, two parts of the AI Act bear on which features you switch on. Article 5 has banned AI that infers people's emotions in the workplace since 2 February 2025, except for medical or safety reasons, so a camera that reads staff mood is out whatever the brochure says. Article 50 adds that anyone deploying a system that recognises emotions or sorts people into categories from their biometric data must inform the people exposed to it, a duty that has applied since 2 August 2026. Article 5 also bans biometric categorisation used to infer sensitive traits such as race, political opinions or sexual orientation.
A plain counter does none of these. Whether estimating an age band or gender from a face counts as biometric categorisation is a question for your adviser, and the safest answer for a shop is to switch the feature off. None of this replaces data-protection law, which applies to any footage in which people can be recognised.
A bike shop, a toy shop and a shop with two doors
A bike shop with a Saturday repair queue
The problem is the wait at the repair desk, not the door count. Try a ticket system and a clipboard of drop-off times for a month first. If the queue still builds, a sensor that outputs "people waiting at the desk" answers it without faces. The privacy question is whether that camera also records the desk, where customers hand over bikes and give names and phone numbers.
A toy shop full of children
Counts only. Children are a large share of visitors, and children's data attracts extra protection in many places, so there is no case for age estimates or face features. A door counter and hourly totals are enough to staff the school-holiday weeks. Keep the sign's wording plain and put it at adult eye height by every entrance.
A shop with two doors
Count both, and exclude the staff door from the visitor total. Delivery drivers use whichever door is nearer, so audit each door separately. If the shop shares a lobby with other tenants, check what the sensor's view takes in: a counter angled across a shared entrance may be counting a neighbour's customers, and filming them. The tutorial on AI for retail store layout shows what to do with dwell-time findings once the counts are trustworthy.
What to ask a footfall vendor, and answers that should worry you
| Question | The answer you want | An answer that should worry you |
|---|---|---|
| Where is video analysed? | On the device; only counts leave it | Streamed to the vendor's cloud by default |
| Is any image or video stored, and for how long? | None, or a short buffer you control | Continuous recording with no set retention |
| Does it build face templates, visitor IDs or repeat-visit data? | No, and the feature can be disabled | "Optional", but on by default or needed for the dashboard |
| Does it produce age, gender or emotion estimates? | Off, or not offered | Always on and cannot be disabled |
| Who can see live and recorded video? | Named users you control; vendor staff only with your consent | Vendor staff can view at will |
| Is your footage used to train the vendor's models? | No, stated in writing | Silence, or "to improve our services" |
| What happens when the contract ends? | Deletion confirmed in writing | Data kept indefinitely |
| How do you prove the count is accurate? | A method you can repeat on site | A single accuracy percentage on a brochure |
Checking the privacy set-up after installation
- Compare the settings page with the screenshot from installation day. Demographics, face features and cloud video should still be off.
- Walk the shop and check each sensor's view. None should take in a fitting room, toilet, staff room or a neighbour's property; a counter never needs to.
- Log in as each user and confirm they see only what they should.
- Ask the vendor to delete one day's data and confirm it has gone.
- Photograph the signs in place and date the photo.
- Repeat the clicker audit after a month and then every quarter.
- Put the record's review date in the diary, and add a line to your AI policy so the next manager knows the camera exists.
Footfall camera questions shop owners ask
Do counters that store only numbers need a sign?
A sign is good practice either way, because the device still sees people even if nothing is kept. Whether the law requires one depends on where you trade and on whether any image is captured or stored, so ask a data-protection adviser. If you do display a sign, make sure it describes what the device really does.
Can footfall cameras leave staff out of the count?
Often, in one of three ways: staff wear a tag the sensor recognises, staff use a door or zone that is excluded from counting, or you subtract an estimate for staff movements. Tags and zones are less intrusive than any method that recognises staff by face. Ask the vendor which it offers, then test the result with a clicker audit.
Is age and gender estimation allowed in a shop?
It depends on where you trade and how the feature works, and it is the step where a counting camera starts processing faces. For staffing and conversion you don't need it. If you have customers in the EU, ask whether it counts as biometric categorisation under the AI Act, and switch it off unless an adviser agrees.
How accurate are AI footfall cameras?
Accuracy depends on mounting height, door width, lighting, groups and prams, so a brochure percentage says little about your door. Count by hand for three 30-minute windows, compare with the sensor, and ask the installer to adjust until the error is small and steady. Repeat the check each quarter.
Further reads
- AI Software for Retail Stores: Five Checks Before You Buy — Five checks to run on any retail AI product before buying.
- CCTV Installation Checklist: Commissioning, Signage and Handover With AI — What an installer should hand over: signs, settings, records.
- AI Customer Service Kiosk for a Shop Counter: Worth the Cost? — The other screen at the counter: is a kiosk worth it?
- Does a Five-Person Business Really Need an AI Policy? — Where a footfall camera fits in a one-page AI policy.
- Best AI for Shop Design: Floor Plan and Display Brief for Your Fitter — Use dwell-time findings when planning windows and displays.
- Customer Loyalty Programmes for Retail Stores: Can AI Design One? — A consent-based way to see who comes back.
- AI Mirror for a Clothing Store: Is a Smart Fitting Room Worth It? — A small-store payback sum for an AI smart mirror, the vendor claims to test, camera privacy questions and cheaper ways to get the benefit.
- AI Photo Booth for Events: Hire or Buy, Costs and Guest Privacy — Hire or buy an AI photo booth: the break-even sum, the vendor claims to check, and the consent signage and retention rules that protect your guests.
- AI Shoplifting Detection for a Small Shop: What It Sees and the Law — Loss sums, false-alarm numbers, privacy duties and a vendor question list for a small shop weighing AI shoplifting detection on its cameras.
- AI Camera Software for CCTV Installers: Analytics You Can Sell Clients — Human detection, line crossing, counting and plate recognition compared for installers, with a 14-day false-alert test, privacy checks and margin sums.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: the EU AI Act text, Article 5 (prohibited practices, applying since 2 February 2025) and Article 50 (transparency duties, applying since 2 August 2026), read in October 2026. General information, not legal advice.