Automated inventory: where manual tracking breaks down
Manual recounting breaks down not during the count itself, but earlier — during movements. A pallet was moved from receiving to the storage area, part was shipped, the remainder was returned — and in the records this is either not reflected or reflected retroactively.
Every unrecorded movement per shift adds its own line to stock discrepancies, and by the end of the week they no longer reconcile with any of the reconciliations.
A separate source of errors is double entry. The storekeeper counts on a paper sheet, the operator transfers the numbers into the accounting system, and a discrepancy appears out of nowhere: a mixed-up line, a lost sheet, a handwritten item.
Mis-sorting amplifies the effect — goods with similar SKUs and identical labels get swapped, and the error surfaces only during the next inventory.
The third layer is time. A full walkthrough of the warehouse area and piece-by-piece counting stop shipping: to get an accurate picture, the warehouse has to be closed. The wider the product range, the longer the cycle and the less of the shift remains for operational work.
The result is the same: manual recounting gives a picture with delay and error, while purchasing and shipping decisions are made blindly. Below we look at how warehouse stock accounting is transferred to automated inventory — via cameras, without stopping warehouse operations.
What machine stock control delivers in practice
What becomes measurable
Cameras on the shelf provide a video stream that is converted into events: item picked, item placed, zone emptied. The events are then matched against the product catalog and sent to the accounting system — 1C or its equivalent — via standard exchange, without manual entry. Real-time stock is obtained not from a recount, but from the stream of these events.
The speed of recounting changes qualitatively: a walkthrough of the hall that used to take a shift and required stopping work now runs in the background on a schedule and is completed in minutes.
The share of recognized items is the number-one metric, and we calculate it for each zone separately: a distant shelf, glare from display cases, and dense placement degrade differently; an overall percentage explains nothing here.
The second effect is a class of disappearing errors. Mis-sorting, shortages, and receiving errors stop accumulating until inventory: a discrepancy is visible at the moment it occurs and is resolved immediately.
Hence the reduction in losses — not by “tightening control,” but because an error has a short path to correction.
This directly affects product availability. An item that previously went out of stock for a week until the next count returns to the shelf the same day, and accounting accuracy stops depending on how diligently the walkthrough is done.
Data capture methods: cameras, drones, mobile stations
The choice of capture method is determined not by optics alone, but by compatibility with the warehouse's physical reality: rack height, aisle width, lighting conditions, and packaging type.
An option that reliably reads matte boxes in one storage zone fails on transparent and shiny film in the next. Therefore, we first describe the layout and product placement, then select the monitoring setup.
| Option | When it fits | Limitations |
|---|---|---|
| Stationary cameras | Continuous control of storage zones, even light, stable pallet placement | Blind spots behind racks and deep in shelves, glare from film, rearrangement breaks the scene |
| Rail-mounted and PTZ cameras | Aisles of the same width, repeatable walkthrough of the same rows | Requires installation of guide rails; when layout changes, the trajectory is recalculated |
| Drone | High racks and scheduled capture of the entire warehouse | Requires a free flight corridor; ceilings and safety nets interfere |
| Mobile inventory stations | Rack walkthroughs on equipment, large storage zones, moving flow | Accuracy drops as speed increases and camera tilts; route marking is required |
How the options combine
In practice, one setup rarely covers the entire warehouse. Stationary cameras take on receiving and shipping zones, the drone handles upper tiers where optics cannot reach, and mobile inventory stations pick up the remainder in large storage zones during normal equipment runs.
Items that are not read automatically — thin packaging, crushed boxes, goods without a code — go to a manual rack walkthrough, and its share shows exactly how well the scheme was chosen.
Stages of recognition implementation: from scheme to handover
We fix the work order before the start: each stage ends with an artifact that can be checked on site, so recognition quality is not adjusted during a working shift. Below is how automated inventory implementation proceeds.
-
Warehouse audit. We walk the site, capture aisle geometry, analyze the product range: dimensions, similar packaging, share of goods without readable marking. The output is a list of areas where cameras deliver results and where additional control is needed.
-
Field-of-view zone scheme. We calculate installation points, angles, and overlaps so that a shelf falls into at least two views, and blind spots are explicitly listed.
We agree on the scheme before installation: moving cameras after commissioning is more expensive than correcting the drawing.
-
Installation and calibration. We set views for the shift's lighting and link them to storage locations. We check that the night shift and local light sources do not break the image.
-
Recognition training. We label reference samples of your specific goods: new items, repackaging, seasonal batches. We train on real footage from the site — otherwise the share of false positives grows on similar packaging.
-
Testing on real scenarios. We run through mis-sorting, partially closed boxes, pallets in the aisle, and item substitution. Discrepancies are checked against an actual recount; disputed frames go to the operator with justification.
-
System acceptance. Control recounts on your product range: accuracy by area, full walkthrough time, list of exceptions. We sign a protocol with figures.
-
Handover. We provide access credentials, a procedure for handling discrepancies, and instructions for storekeepers. Separately, we agree on how recognition is updated when new product items appear.
What you keep: zone schemes, access rules, procedures
We hand over not a presentation, but a working kit: the shift uses it to complete the walkthrough, the supervisor resolves discrepancies, and your engineer edits zones without our help.
The system documentation is assembled during the project — you see its drafts at every stage and make edits before final delivery, not after.
- Zone schemes: floor plans with the boundaries of each capture zone, camera installation points, and a list of SKUs stored in the zone.
- Recognition configurations: rule sets for different rack and packaging types, trigger thresholds, and exceptions for glare, closed boxes, and obstructed shelves.
- Access matrix: who sees discrepancies, who confirms a recount, who changes zone settings — roles of shift operator, supervisor, and administrator.
- Recount procedure: the sequence from system signal to final report, with steps divided between the shift and the warehouse owner.
- Shift instructions: one- to two-page memos — starting a walkthrough, actions if a camera fails, sending a disputed item for manual check.
- Backup procedure: where configuration copies are stored and how to return to the previous rule set if zoning or assortment changes.
- Change log: every edit to the scheme or rules is recorded with date and author — six months later you can see why a threshold was set exactly that way.
We also hand over the files themselves, not just descriptions: the schemes and configurations are suitable for editing by your team. If you need to expand the set — for example, add a procedure for a second site — tell us in advance; we will preserve a unified structure.
Case study: accounting accuracy increased, warehouse losses decreased
What it was like before implementation
A warehouse with several thousand active SKUs was counted manually: a full inventory took two shifts and stopped shipping, while discrepancies surfaced only at the moment of recount — a month after they appeared.
Shortages held at 1.7% of turnover, and it was impossible to understand exactly where goods were being lost.
Cameras were installed in rack zones and receiving, and stream processing was moved to local NVIDIA accelerators — frames do not leave the warehouse perimeter.
Recognized items are checked against the accounting system, and a discrepancy immediately becomes a task for the storekeeper, not a line in a report a month later. The pilot launch in one zone took 4 weeks, and the full setup went live in 3 months without stopping the warehouse.
Metrics before and after
| Metric | Before | After |
|---|---|---|
| Share of recognized items | 84% | 99.3% |
| Discrepancies caught in the first 24 hours | 20% | 94% |
| Goods losses | 1.7% of turnover | 0.4% |
| Full recount time | two shifts | 5 hours |
A side effect: storekeepers stopped spending shifts on recounting and focused on shipping. The limits of applicability were also documented — recognition makes mistakes on crumpled and tightly closed packaging, so 5% of items were left for selective manual control.
How does automated inventory differ from scanners?
A data collection terminal and image recognition solve different tasks, and confusing them is the most common mistake when choosing accounting technology.
A scanner works on the principle of “a person brings the tag to the reader”: coverage equals the walkthrough route, and accuracy equals the employee's attentiveness.
A camera captures what falls into the frame, so accounting completeness depends on the placement of observation points, not on staff discipline.
Radio tags (RFID) eliminate manual presentation: a reader at the gate sees the entire flow of containers and boxes. The approach has its own bottlenecks — metal and liquid shield the signal, and a tag must be attached to every unit already at entry.
In practice, a combination wins: cameras provide continuous zone control, the terminal covers disputed items, and radio tags cover entry and exit.
The choice of accounting technology is determined not by fashion, but by what percentage of discrepancies is acceptable and how many manual operations you are willing to keep.
Common technical doubts before system launch
Before the start, almost every customer asks the same five questions — and these are the right questions. Below are engineering answers without promising “100% accuracy”: where the system makes mistakes and how that class of errors is closed.
How does the system distinguish similar packaging?
Identical containers with different contents are the main source of misses.
The solution is built not on a single image, but on a combination of features: reference samples for each SKU, control of multiple frame zones, and mandatory reconciliation of the result with the accounting system. Disputed frames go to the operator for confirmation, not automatically into accounting.
What about blind spots and rack rearrangement?
We design the camera placement scheme based on actual goods movement routes, not the floor plan — this covers zones obstructed by posts and equipment.
When racks are rearranged, no reinstallation is required: recognition zones are re-described in the settings, and the system continues to work with the previous cameras.
Will the system handle lighting changes?
Natural light from windows, lights turned off at night, glare from packaging — all of this affects accuracy. Therefore, calibration is performed separately for working lighting modes, and decisions are made based on an accumulated set of frames, not a single one. If lighting changes unpredictably, we add illumination at capture points.
How are partial shipments counted?
A shipment rarely leaves in full: part of a pallet goes out, part remains. The system records a change in the zone's state, not the fact that something “passed by the camera,” so short loading and the remainder are reflected separately. A quantity discrepancy is visible immediately, not at the next inventory.
What about returns and mis-sorting?
A return is goods crossing the same zone again in the opposite direction, and it is easy to confuse with a new shipment. We label the flow direction and operation status so that a return follows a separate scenario and does not distort stock. For each such case, a record with a frame remains — something to show when resolving discrepancies.
Let's discuss the task: how we'll start and what you'll get
To start, little is needed: a list of product groups and storage zones, access to an archive of camera recordings from several typical shifts, and a person who will confirm the accounting rules on your side.
This is enough to calculate labor intensity based on real data, not a description. The first stage ends not with a presentation, but with a solution scheme, work order, and timelines that can already be used to make a decision.
- Project assessment: number of zones and cameras, daily event flow, scope of exchange with the accounting system, and bottlenecks that determine labor intensity.
- Solution scheme: camera angles and placement, what is recognized on site and what goes to the server, how the result is reconciled with stock.
- Rules for disputed situations: obstructed view, goods without marking, pallet rearrangement — and what the system flags as requiring human review.
- Stage-by-stage implementation plan: zone preparation, recognition setup, reconciliation with accounting, staff training on handling discrepancies.
- Work timelines: duration of each stage and what affects it — lighting, placement density, archive quality.
- Pilot launch on one product group or zone with success criteria in figures: share of recognized items and number of false positives.
- Team composition and points of contact: who leads the project on our side and who approves changes on your side.
Describe the task — zones, product groups, and what is currently hindering accounting. We will come back with a project assessment, solution scheme, and work order; then the decision is yours.
