How to Use AI to Decide Which Products to Cut

Every product line has a tail: items that sell three units a month, occupy shelf space, tie up cash, and survive because nobody has done the arithmetic. Deciding which products to discontinue starts with that arithmetic, and AI is genuinely good at it and genuinely bad at the decision that follows. The split matters, because the failure mode here is expensive in a way most AI mistakes are not. Discontinue the wrong item and you find out ninety days later when a regular customer stops coming.

Here is a process that uses the model for the part it is good at.

Get the five numbers first

Before any prompt, pull five columns per product for the last twelve months. If you cannot produce these, the analysis is not blocked by AI, it is blocked by bookkeeping.

• Units sold, by month, so seasonality is visible

• Gross margin per unit, not price

• Inventory holding, average units on hand and how long they sit

• Attachment, how often the item appears in a basket with something else

• Returns or waste, whatever your equivalent is

Four of these live in your point of sale export. Attachment is the one people skip and it is the one that prevents the expensive mistake, because a low-margin item that pulls a high-margin basket is not a low-margin item.

How to ask AI which products to discontinue

Give it the table, not a description of the table. A CSV of 200 rows pastes fine and the model can compute across it, which is the technique in prompting AI to analyse a spreadsheet.

A prompt shaped like this produces something usable:

Here is 12 months of product data. Columns: sku, name, units_sold_by_month
(12 values), gross_margin_per_unit, avg_units_on_hand, days_on_hand,
basket_attach_rate, return_rate.

Do four things.

1. Compute annual gross profit and inventory turns per SKU.
2. Group SKUs into: clear keep, clear cut, and needs judgement.
3. For every SKU you put in "clear cut", state the single number
   that drove it and what the annual gross profit loss would be.
4. For every SKU in "needs judgement", say exactly what extra
   information would resolve it.

Do not recommend anything for a SKU whose 12 month unit trend is
rising, even if the total is small. Flag those separately.

Two things in that prompt do the work. Asking for the driving number per recommendation makes the reasoning checkable in seconds instead of taking the answer on faith. Asking what information would resolve the uncertain cases turns a vague middle group into a to-do list.

The last constraint is the one worth copying. A rising trend on a small base is a product that is working, caught early, and pure margin analysis will kill it. Models will not add that rule for you.

The four questions the data cannot answer

Take the model's "clear cut" list and run every item through these before acting. This is the part that stays human.

Does it anchor a customer relationship? The one item a regular buyer comes in for. If they can only get it from you, cutting it costs you the whole basket, not the margin on that line.

Does it complete a set? Ranges that look incomplete convert worse. The slowest size, the odd colour, the small pack that makes the large pack look reasonable, these can be doing pricing work rather than volume work.

Is it a signal or a service? The item that says you are a serious supplier. The gluten-free option. The part that costs you money and buys you being the place that has it.

What replaces the space? Cutting is only a gain if the shelf, the cash, or the attention goes somewhere better. If it goes nowhere, you have reduced revenue and kept the overheads.

A model can prompt you with these questions. It cannot answer them, because the answers live in things it has never seen: who your customers are, what they said last month, what your competitor down the road stocks.

Run the discontinue list as a trial

The lowest-risk version of this is not a discontinue list, it is a stop-reorder list.

Pick the items where the analysis and your judgement agree. Stop reordering, sell through existing stock, and watch three signals for a quarter: total basket size for customers who used to buy the item, direct requests for it, and whether the freed capital or space produced anything.

You keep the option to reverse, which a discontinuation announcement does not. Cash comes out of the tail either way. And you generate real evidence for the next round, which is the same discipline as measuring AI ROI for a small business applied to inventory.

Where it pairs with forecasting

This analysis is backward looking by design. Pair it with a forward view before a seasonal category, since an item that looks dead in August may be a December product. The method for that is in using AI to forecast demand, and the two together are much stronger than either alone.

If your stock data is a mess, fix that first. AI tools for inventory management covers getting to a clean export, which is the actual prerequisite here.

FAQ

How much data do I need for this to work?

Twelve months, so seasonality is visible. Six months works if the category is not seasonal, but expect the model to mistake a summer dip for decline.

Can AI decide which products to cut on its own?

No, and the reason is not caution. It lacks the inputs. Attachment effects, customer relationships, and competitive positioning are not in your sales export, and a recommendation made without them is confident and uninformed.

What if margin data is unreliable?

Then start there. Ranking on revenue rather than margin systematically protects high-revenue low-margin items, which are frequently the ones that should go.

How often should this run?

Twice a year is plenty for most small businesses, once before your main buying season and once after. Monthly turns a decision into a habit of churn.

Does this apply to services, not just products?

Yes, with the same columns renamed. Hours delivered, margin per hour, and whether the service pulls other work. The tail of an agency's service list behaves like the tail of a shop's shelf, and the same logic that governs sunsetting a feature applies. For the wider toolkit, see AI for small business.