What was breaking before?
Manual stock takes were the problem: slow, error prone, and blind between counts. A stock take meant staff walking aisles with clipboards or spreadsheets, reading labels by eye, and typing counts in later. Every retype was a chance for a wrong digit, a skipped shelf, or a duplicated line.
Because counting took so long, counts happened rarely, and the business ran between them on stale numbers. No real time visibility meant reordering, promotions and shrinkage checks all leaned on figures that were already out of date the day they were captured. Errors surfaced late, usually as a mismatch between what the shelf held and what the record claimed, and tracing the mismatch back to its source was guesswork. The team did not lack effort. The process itself burned hours and still produced numbers nobody fully trusted.
What did we build?
We built an OCR workflow: photos in, structured stock data out, with a human spot checking edge cases. Staff photograph shelves, bin labels and delivery paperwork on an ordinary phone. The images upload to a processing pipeline where OCR reads the text and an AI layer turns raw characters into structured line items: product, code, quantity, location.
Each extraction carries a confidence score. High confidence lines write straight into the stock record. Low confidence lines, unusual quantities and unreadable labels queue for a person, who confirms or corrects them in a simple review screen. The correction feeds back, so the same edge case gets easier over time. The whole flow bolts onto the retail operator's existing stock system rather than replacing it, and every write is logged, so any figure can be traced back to the photo it came from.
Client identity and specific commercial metrics stay under NDA. What we publish is the shape of the system and the outcomes the client has approved. On a call we walk through live systems, not slides.
How does the system decide what to do?
The deciding rule is simple: the system only trusts itself when the evidence is clear. Every OCR read produces both a value and a confidence score, and that score drives the routing. A crisp label, a clean match against the product catalogue and a plausible quantity mean the line posts automatically.
Anything else escalates. Blurred text, a code that matches nothing in the catalogue, a count far outside the normal range for that product, all of these land in the human review queue instead of the stock record. Nothing ambiguous writes itself into the numbers. A person works through the queue from a phone or desktop, and each decision is logged with who made it and why. This split keeps the maths honest: the machine handles the repetitive bulk it can prove, people handle the judgement calls, and the audit trail shows which was which for every line.
What changed for the team?
Two things changed: stock take time was substantially cut, and real time stock visibility was unlocked. Those are the outcomes the client has approved for publication, and both follow directly from the mechanics. Counting became photographing, so the slow part of a stock take, the walking, squinting and retyping, collapsed into pointing a camera.
Visibility changed shape as well. Before, the stock picture was a snapshot that aged from the moment it was taken. Now data lands in the record as photos are processed, so the picture stays current instead of waiting for the next full count. The human role shifted from doing every count by hand to spot checking the edge cases the system flags. We publish no percentages or rand amounts for this build. Under NDA, on a call, we can walk through what the numbers look like on a live system.
How would this look in your business?
The same pattern fits any business where paper or shelves hold data that a system needs: photos or scans in, structured records out, humans checking only the exceptions. Retail stock is one version. Delivery notes, invoices, meter readings, job cards and warehouse labels follow the identical shape.
The build order stays constant. First we map where the numbers currently come from and where they need to land. Then OCR and an AI extraction layer read the documents or images, a confidence rule decides what posts automatically and what queues for review, and the results write into the system already in place, whether that is a stock platform, an ERP or an accounting package. A tightly scoped pilot on one document type or one store proves the accuracy before anything scales. Tell us which count or capture step eats the most hours, and we will give an honest read on whether this pattern fits.
Related capabilities. The same parts, your business.
Keep reading. Pages close to this one.
Tell us what runs slow. We build what fixes it.
Send one message describing the count, capture or paperwork step that eats the most hours. We reply with an honest read on what an OCR workflow can fix, what it will take, and whether it is worth building at all.
