People check it by eye
No one is free to count during the busiest hour. A value jotted down from memory after the peak is no basis for next week’s staffing.
saai queue · Queues & crowding
Why does the line only
grow long at this hour?
Use existing CCTV to count how many are waiting at checkout and how long the wait runs, together. The moment crowding starts and the route that caused it are both kept.
- Waiting · standing
- Passing · walking through
- Congestion zone
The camera watches the checkout and the aisle in front of it, together. Standing in line is waiting; walking past it is passing through.
- Waiting
- 4
- Expected wait
- 6 min
People standing at checkout
Time if you join the line now
Suggests opening another counter
Threshold 5 people · over 10 min · the store confirms
No identity is kept. Video is deleted after analysis and only the statistics remain. * Figures are examples.
01 Limits
By the time you notice the line grew, it already has
Only the paying customer remains
You know how long checkout took, but not how many were standing in front of it.
Anyone who looked at the line and walked out never appears in POS. The sales you lost live on the side with no data.
The cost that stayed invisible
ExampleDrop-off in two peak hours
runs about 4.32 million KRW a month
People who saw the line and turned away never appear in POS, so this figure is written in no ledger anywhere. It becomes manageable the moment it is measured.
saai queue keeps the headcount waiting and the expected wait by the minute, along with the moment the threshold was crossed and the staffing at that time.
- Layer 1Peak drop-off · 8 people × 2 peak hours16 people/day
- Layer 2Average ticket · 16 people × 9,000 KRW144,000 KRW/day
- Layer 3Over 30 days4,320,000 KRW
1 store · month
about 4.32M KRW
10 stores · month
about 43.2M KRW
Assumptions · 8 people lost per 2-hour peak · 9,000 KRW average ticket · 30 trading days a month
Drop-off headcount and average ticket are example values. Real numbers replace them once measured on site.
saai queue measurement
Before the line grows long,
it is already visible.
Headcount waiting, expected wait, and zone congestion, measured by the minute on the same camera.
02-1 Two signals
Waiting and crowding are different signals
Waiting is checkout’s score. Crowding is the space’s score.
Put the two together and every store lands in one of four places. Which place it lands in splits the fix between adding staff and changing the layout.
Low congestion · long wait
Checkout problem
People are spread out but the line is still long. Look at processing speed.
High congestion · long wait
Understaffed
When both signals rise together, decide when to add staff.
Low congestion · short wait
Normal
Use this layout and staffing as the baseline.
High congestion · short wait
Aisle problem
The line is short but the aisle is blocked. Look at the layout.
Neither POS nor a gate counter knows how long someone stood. That is why they cannot build the second axis.
What HQ receives is not a live screen. It is a table that ranks every store on the same two axes.
Sub-metrics behind the two signals
- Peak by hour
- Weekday variance
- Threshold crossings
These explain why the two signals came out the way they did. Figures are examples.
02-2 Simulation
Is it a staffing problem, or a layout problem?
The two signals do not say where the problem is. Run the measured inflow through two layouts and the cause splits apart.
Same inflow · same time slotExample
2 crossings
Layout A. The queue in front of one checkout counter reaches down to the main aisle and crosses the walking route in two places.- Peak queue
- 6
- Expected wait
- 11 min
- Route crossings
- 2
No crossings
Layout B. Checkout is split into payment and pickup, and the queue turns toward the wall so it never crosses the main aisle.- Peak queue
- 4
- Expected wait
- 6 min
- Route crossings
- 0
A · Reading the current layout
Crossings remain even with more staff
Once the queue passes 6 people, its tail reaches the main aisle and meets the walking route in two places. Faster checkout does not change when the aisle gets blocked.
The simulation runs on measured values. A simulation run without real measurement is not evidence. Measure waiting and crowding first, then compare layouts with those values. The store and HQ confirm.
03 Evidence
Evidence that it counts right, and evidence that the judgment changes
It counts right
94–96%
Accuracy compared against a manual headcount of the queue.
- Convenience store 2025 Q1
- 95.2%
- Café 2025 Q2
- 94.4%
- Food court 2025 Q3
- 96.1%
* Figures are examples. Real accuracy varies with camera angle, lighting, and how the queue forms.
The judgment changes
Where the judgment flipped
ExampleA time slot flagged for more staff turned out to be an aisle-layout problem
Checkout processing speed was within the brand average. The wait only spiked when the queue overlapped the walking route, so staff were not added; the queue direction was.
Deployment scale
Example42 stores · 9 months
Written as category and scale. The brand name is not stated.
Cumulative measurement
Example11.8 million records
Cumulative records of headcount waiting and expected wait, by the minute.
04 Privacy
Every measurement uses only anonymized spatial data.
Privacy by Design
Anonymization comes first. The SEAL engine
- No video stored
- Deleted after analysis, stats only
- No originals kept
- Source footage is never retained
- No re-identification
- Identity removed, flow only
CV reads the floor while SEAL removes identity.
05 Install and measure
Three steps on site, three steps in the app
Install · on siteOn the CCTV you already use
01
Check the frame
Just check that checkout and the aisle in front of it fit in one frame
02
Plug in the AI box
One outlet next to your existing recorder. No cabling work
03
Connect your phone
Set up from the phone browser. No keyboard or mouse needed
1 / 3
Measure · in the appSet the zones and it starts
04
Set zones
Mark the waiting zone and the walking zone on screen
05
Set thresholds
The store decides how many people, how many minutes, before an alert
06
Alerts and reports
Threshold alerts, reports by daypart and weekday, and CSV
1 / 3
No new cameras go up, so a rollout to 100 stores does not stretch the schedule by 100 times.
Read through sense, analyze, and think
One measurement reads three ways
care senses it. insight analyzes it. agent proposes what to do next and connects it to the action.
- saai careSense
What needs attention?
Real-time alert when a queue or crowding crosses its threshold
- saai insightAnalyze
Why did it happen?
Waiting-time peaks and crowding cycles
- saai agentThink
What should we do?
Staff reallocation, extra counters or self-checkout routing proposed
1 / 3
If you need what happens at the door
Who passed the door and who walked in is counted by saai count. queue picks up where the checkout line starts.
Where it leads