Home › AI & decision support
Topic 3
AI and decision support
Which decisions a small shop should hand to a system, which it should keep for itself, and how to tell a useful tool from an expensive subscription with the word "AI" on the box.
Four questions, in order
Almost everything sold as "business intelligence" or "AI" answers one of four questions. They build on each other, and skipping a step does not work.
Scroll sideways to see the whole diagram.
Level 1: Descriptive Analytics – "What happened?"
Summaries of data in language. At Lerato’s a simple example is a sales report that shows R15 000 worth of stock sold yesterday.
Level 2: Diagnostic Analytics – "Why did it happen?"
Looks at transaction details to find reasons. At Lerato’s this means digging into sales to see that cement sales jumped because a local building contractor bought twenty bags on credit.
Level 3: Predictive Analytics – "What will happen?"
Uses patterns to guess trends. For Lerato’s this is estimating that cement demand will rise by 25 % this month because of building activity.
Level 4: Prescriptive Analytics – "What should we do?"
Gives advice based on models. At Lerato’s this could be a computer instruction to order fifty cement bags.
Core Dependency – A prediction or advice is only as good as the data below it. If Lerato’s handwritten pointofsale logs contain mistakes or missing amounts even the advanced algorithms will give results that look convincing but are wrong.
Honest Recommendation– For a shop like Lerato’s Person 3 should first master Level 1 and then develop Level 2 habits. Clean and consistent data on sales and stock will bring benefits. Justify the cost of the system. Person 3 may wait on Level 3 until two or three years of trading history are available.
Vendor Reality– Vendors sell Level 4 features because they seem advanced and cost more. In reality they often fail for shops because there is not data. Before Person 3 buys the vendor must prove: "Can your system give advice with the history of a singlestore hardware shop like ours?”
Which decisions should a system make?
Not all decisions are the same shape. The useful split is how repeatable they are.
| Type | What it means | At Lerato's | Give it to a system? |
|---|---|---|---|
| Structured | The rule is clear, the inputs are known, the answer is the same every time | When to reorder cement. How much VAT on an invoice. Whether this customer is over their credit limit. | Yes — automate it |
| Semi-structured | Part of it can be calculated; part needs judgement | Which slow line to mark down. Whether to extend a contractor's terms. What to stock for winter. | System advises, person decides |
| Unstructured | No rule, no reliable inputs, consequences are strategic | Whether to open a second branch. Whether to take on a partner. Whether to sell up. | Keep it — but bring better data to it |
Core Dependency – A prediction or advice is only as good as the data below it. If Lerato’s handwritten pointofsale logs contain mistakes or missing amounts even the advanced algorithms will give results that look convincing but are wrong.
Honest Recommendation– For a shop like Lerato’s Person 3 should first master Level 1 and then develop Level 2 habits. Clean and consistent data on sales and stock will bring benefits. Justify the cost of the system. Person 3 may wait on Level 3 until two or three years of trading history are available.
Vendor Reality– Vendors sell Level 4 features because they seem advanced and cost more. In reality they often fail for shops because there is not data. Before Person 3 buys the vendor must prove: "Can your system give advice with the history of a singlestore hardware shop like ours?”
What is actually useful to a shop this size
Honest assessment, in order of what returns the most for the least.
Worth it now
- Reorder suggestions — sales rate against supplier lead time. Arithmetic, not AI, and it is the single highest-value calculation in the whole system.
- Exception lists — what is below reorder, who is over 60 days, which records are failing validation.
- Margin by product and by customer — reveals which trade accounts are busy rather than profitable.
- Dead stock and shrinkage — both are the owner's money, and neither is visible on paper.
Later, with clean data
- Seasonal demand forecasting — genuinely useful for roof paint, geyser elements and building materials, but needs two or three years of trustworthy history.
- Basket analysis — what sells with what, which drives layout and add-on selling.
- Customer segmentation — separating contractors from DIY walk-ins properly rather than by impression.
- Reading supplier invoices automatically — the technology works; the question is whether the volume justifies it.
Be sceptical
- "AI-powered" anything with no explanation of what it does. Ask what it predicts, from what data, and how often it is wrong.
- Automated pricing — in a market where relationships and a handshake price matter, this breaks things quietly.
- Automatic ordering with no human approval. One bad forecast becomes a yard full of stock.
- Customer chatbots for a shop with one phone line and loyal regulars who would rather speak to Sipho.
Check against POPIA first
- Anything that profiles customers by behaviour.
- Anything that uses purchase history for marketing without specific consent.
- Anything that sends customer data to a third-party service to be analysed — including a general-purpose AI assistant.
- Automated decisions that affect a person, such as refusing credit. Verify what s.71 requires here and cite it.
For a retailer like Lerato, the way to keep operations smooth is to do math all the time. Counting stock correctly and knowing sales numbers are more useful than machine learning. Stating these limits makes Lerato more trustworthy. It stops Lerato from spending money on technology.
A statistical forecast is an estimate that has uncertainty. Showing a sales number without telling the error margin makes Leratos think the number is certain which can lead to stock choices. For example saying Leratos will sell one hundred units but not showing how much that number could change might leave Leratos with stock if people buy less.
Machine learning models need a lot of data, deep history and clean data to work well. A small shop that has eighteen months of spreadsheet data does not give Leratos data for a useful machine learning model.
Algorithmic Bias in Retail: AI models copy patterns in data. If Leratos never stocked power tools the data will show zero sales for that category. A machine learning model that looks at this data will think there is no demand for power tools and will suggest Leratos never buy them. That mixes market demand with data.
Explainability Standard: Small business owners need to see why a system suggests something. If a system tells Leratos to double a bolt order but gives no logic Leratos owner will ignore it. A useful recommendation must show calculations.
Three Practical Questions for AI Vendors:
- "Can you show me the formula or logic your system used to calculate this recommendation?"
- "How does this software perform when our store data is missing entry fields or contains formatting errors?"
- "What volume and duration of sales history must we feed into the system before its suggestions become reliable?"
What the system must never be allowed to decide
A system should decide things that are repeatable. A person should decide things that are consequential and rare. The failure mode of a small business is not having too little automation — it is automating a judgement call and then not noticing when the circumstances changed.
Core strategic decisions within a store must stay under Leratos control. Never be fully automated. Examples include:
Extending credit to a trade customer experiencing shortterm cash flow difficulties.
Adjusting product pricing for a neighboring business owner.
Continuing to stock a lowmargin hardware line because a key account customer relies on it.
These choices involve dynamics, community trust and longterm customer relationships—things that cannot be captured fully in numbers in a database.
Letting an automated system make these decisions creates risks. If an algorithm quietly denies Leratos credit based on payment patterns it makes a choice that hurts a customer’s life without Leratos knowing without Leratos being accountable and with no chance to appeal.
When an automated system makes decisions it should act as a helper. It must show Leratos facts—such as balances and buying trends—record the action along with the responsible manager’s name and make Leratos owner keep Leratos accountable for all client relationships.
Enterprise Systems & Integration (pages/enterprise-systems.html)