What we learned from analysing 100 GB of small-retailer sales data

We spent six weeks earlier this year working through point-of-sale data from twelve independent UK retailers. Pharmacies, off-licences, a couple of grocers, two clothing shops, one homewares store, one florist. About 100 GB of raw transaction data in total. Anonymised, consented, and now deleted — we kept only the aggregated findings.

The exercise was not academic. We were trying to answer a single question: where does the money actually go in a small retail business? Not on paper, not in the P&L the accountant produces in March. In the day-to-day, line-by-line truth.

This is what we found. None of it is groundbreaking. Most of it is at odds with the conventional retail-advice industry.

Finding 1 — Discounts cost more than theft. Much more.

Across the twelve shops, shrinkage (theft, damage, miscount) accounted for between 0.4% and 1.1% of revenue. The shop owners almost universally believed it was higher. The number that surprised them was the next one.

Unmodelled discounting — that is, price reductions applied at the till or in promotions without anyone calculating what they would cost — accounted for between 2.8% and 5.6% of revenue. Four to ten times the loss from theft, depending on the shop. Yet every single one of the twelve owners had spent more money in the last year on security and stock-control than on any tool that touched pricing.

This is not because the owners are foolish. It is because theft feels like an attack and discounting feels like a strategy. Emotionally they are very different. Financially they are the same — money you do not collect.

Finding 2 — The 80/20 of decision waste is in two categories

Across the dataset, two categories accounted for roughly 80% of avoidable margin loss:

1. Over-ordering on slow-turning stock. A surprisingly small number of SKUs — typically 6% to 12% of the catalogue — tied up 30% to 50% of working capital and contributed almost nothing to sell-through. These were not bad products. They were products ordered at twice the rate they actually moved.

1. Late markdown on dead stock. The average dead line in the dataset sat at full price for 11 weeks past its useful life before being marked down. By that point the discount needed to clear it was so deep that it would have been cheaper to mark down at week three and reinvest the cash.

Both are decision problems, not data problems. The data was already there. The decision moment had no friction in front of it.

Finding 3 — Tuesdays and Thursdays are where money leaves

We did not expect this one. When we plotted avoidable margin loss against day-of-week, two days stood out: Tuesday and Thursday. Not the weekend, when revenue is highest. Not Monday, when staff are sharpest. Tuesday and Thursday.

The hypothesis we landed on, after talking to the owners, was that these are the decision days — the days when most ordering, markdown and discount calls get made. Decisions made under mid-week fatigue, without a structured second voice, are the decisions that cost.

The implication, for a software business: a tool that runs the decision loop is most valuable on Tuesday at 3pm, not on Saturday at noon.

Finding 4 — The best operator is not the most data-literate one

The single best-performing shop in the dataset — measured by margin retention against the category average — was run by a man in his sixties who has never opened a spreadsheet. What he had instead was a notebook and a habit. Every Monday morning, before the shop opened, he sat in the back room with a coffee and wrote down what he expected to happen that week. Every Saturday evening he wrote down what had actually happened. The gap was the lesson.

He had built, by hand, the entire Decision Loop. He just did not call it that.

The other eleven owners, all of whom had more sophisticated tooling, all underperformed him on margin. This is the most important finding in the study, and the one that is least possible to sell.

Finding 5 — Software is a forcing function, not an oracle

The thing the notebook did for the man in the back room was not oracle work. The notebook did not predict anything. The notebook just made it harder to skip the loop. It was a forcing function.

Most retail software fails because it tries to be the oracle. It tells the owner what to do. Owners, correctly, do not trust it — because the software does not know about the wedding next door or the council parking change. The owner has the context.

The role of software, we now believe, is not to replace the owner's judgement. It is to be a notebook that survives a busy Saturday. That is the entire pitch.

What we did with the findings

We deleted the data. We kept the aggregated patterns. We rebuilt the Decision OS around five principles that came out of the exercise:

• Friction at commit, not after. The Decision OS runs before the action, not after.
• The owner is the oracle, the software is the notebook. We never tell you what to do. We make it harder to skip the loop.
• Quantified expectations. Every decision recorded with a 30-day forecast attached. The forecast is the point of comparison later.
• Adversarial review, not consensus. A second voice that disagrees, drawn from the shop's own history.
• Signed receipts. Tamper-proof record of every decision, for the owner and, if it ever matters, an auditor or buyer.

None of this is a feature list. It is a discipline. The software is just the part that keeps the discipline alive when Tuesday gets busy.