Yes, but a shortlist is not permission to discontinue
AI can help identify products that deserve a commercial review. It can bring together sales, returns, stock movements and support notes, then explain why a line appears expensive to keep. What it cannot reliably do from a sales export alone is decide what your customers will buy instead, whether a supplier agreement allows an exit, or whether a low-volume item protects a valuable account.
That distinction matters when assessing an AI supplier. A colourful ranking of your worst sellers is not the same as evidence that removing them will improve profit. Ask what the recommendation is based on, which costs would actually disappear, and what could be lost elsewhere in the business. If the tool cannot show the source records behind its answer, do not use that answer to remove products.
There is a genuine UK example of AI supporting range decisions. In its 1 July 2024 announcement, dunnhumby said Tesco was using its Assortment AI capabilities for range curation. The announcement describes forecasting what happens when products enter or leave a range, including customers switching to another product. This is dated supplier evidence of a named application, not an independent result or a promise that a small retailer will achieve similar benefits.
The buying question is therefore not whether AI can produce a recommendation. It is whether your business has enough trustworthy evidence to challenge it.
Would dropping the product actually save money?
Low revenue and low accounting profit are different problems. A line can look unprofitable after a share of rent and management salaries is allocated to it, yet still contribute cash towards costs that remain after it is removed. Conversely, an apparently healthy gross margin can disappear once delivery subsidies, returns and product-specific handling are included.
ACCA's relevant-cost guidance makes the important distinction: reallocating existing fixed costs does not itself create a cash-flow saving. Changes in future cash flows caused by the decision are what matter. Have your accountant check that distinction before trusting an AI recommendation to close a product line.
Here is hypothetical arithmetic, not a client result or a market benchmark. A product brings in £10,000 over a review period. Its purchase and other avoidable selling costs total £7,000, leaving £3,000 contribution. Your management report allocates £4,000 of shared overhead to it, so it appears to lose £1,000. If those overheads remain and no other sales replace it, discontinuing the product loses the £3,000 contribution. Removing an accounting loss has made the business worse.
The calculation changes if stopping the line releases scarce capacity that another product can use profitably, or avoids genuine storage, licence or staffing costs. Those benefits must be evidenced, not assumed. Keep cash savings separate from staff time freed up: spare minutes do not automatically reduce the wage bill. Ask the supplier to show a keep-versus-remove scenario using the same period, VAT treatment and assumptions.
What evidence makes an AI assessment worth trusting?
A credible assessment starts with consistent product identifiers, net sales, discounts, refunds and purchase costs. Add stock availability, order dates, delivery costs and support effort where those figures are dependable. Missing costs should be marked as unknown, not silently treated as zero. Keep estimates visibly separate from recorded transactions.
Availability is especially important. A line with few sales may have been out of stock, hidden on the website or temporarily unavailable from the supplier. AI cannot infer weak customer demand from transactions that customers never had the chance to make. New and seasonal lines also need appropriate comparisons; a winter product should not be condemned for quiet summer sales.
Ask for a recommendation that shows the review period, the exact product and variant, the source of each material number, the assumed alternative purchase and the information still missing. The evidence should distinguish between a measured fact, a forecast and a judgement. A confident paragraph is not a confidence measure, and a percentage generated by a chatbot is not a statistically validated forecast.
There is no universal minimum number of orders that makes a recommendation reliable. Sparse sales, unusual promotions and large individual customers can all distort an average. A useful supplier should explain when the evidence is too thin and recommend keeping the decision with a person. For customer-linked analysis, use only necessary information in an approved environment; names and contact details are often unnecessary for a product-level comparison.
Will customers buy an alternative or take their whole order elsewhere?
A slow-selling item can still be a reason that a good customer buys from you. Think of a specialist spare part, an unusual size or a low-margin accessory bought alongside a larger order. Removing it might transfer demand to your other products. It might instead send the whole order to a competitor. Sales history can suggest these relationships, but it does not establish what a customer will do after a change.
This is why dunnhumby's published example focuses on transfer of demand, not simply deleting the lowest-selling biscuit pack size or flavour. The principle is useful for a small business even if an enterprise assortment platform is far beyond its needs. Look at whole orders, customer requirements and realistic substitutes before treating individual product margin as the complete answer.
A supplier should be able to compare keeping the line, changing its price, reducing stocked variants, making it available to order and withdrawing it. Each option has different consequences. Special-order status may preserve an important relationship without tying up the same stock, but only if the lead time is honest and the supplier can fulfil it.
Speak to the people who handle sales and complaints. They may know that a product supports a contract, brings repeat customers back or creates disproportionate problems. Treat those observations as evidence to check rather than as an automatic veto. A narrowly scoped, reversible trial can be safer than a permanent removal, provided existing commitments and customer communications are handled first.
Should you buy AI, use a spreadsheet or ask your accountant?
For a small, stable range, a spreadsheet and an accountant's review may be the better choice. Microsoft Excel or LibreOffice Calc can expose sales, contribution and stock age without a custom AI build. Your existing stock system may already provide the reports you need. Start there rather than paying to reproduce a ranking that already exists.
| Option | Good fit | Main limitation |
|---|---|---|
| Spreadsheet and human review | Manageable range with clean records | Manual preparation and judgement still needed |
| Existing stock or reporting software | Regular product-level comparisons | May miss support notes and customer context |
| AI-assisted assessment | Large or changing range with useful unstructured evidence | Requires testing, checked calculations and accountable decisions |
AI becomes more plausible when important information sits in complaint notes, return reasons or scattered supplier communications. Even then, use deterministic calculations for the financial totals and AI to classify or explain the evidence. ChatGPT, Microsoft Copilot or another assistant should not become the sole source of your margin calculation merely because it can write an explanation.
Ask vendors to demonstrate the assessment on a representative sample, including products you deliberately retained and products you previously withdrew. Compare it with a straightforward baseline. Request separate prices in pounds for data preparation, licences, integration, support and ongoing review. There is no defensible universal project price or guaranteed saving for this question. A proportionate purchase depends on demonstrated usefulness beyond your current reporting, not on the promise of a cleverer dashboard.
When this is NOT right for you
Do not rely on AI-led discontinuation if product costs are unreliable, returns are not reconciled, important customer agreements are missing or nobody can validate the conclusion. Fix those gaps before buying a recommendation engine. A large range with poor records is not automatically a better AI opportunity; it can simply produce more persuasive mistakes.
Product safety concerns and contractual obligations are separate decisions, not factors to trade away for a margin improvement. A suspected unsafe product needs the appropriate safety process and specialist advice, not a wait for the next profitability review. Similarly, removing a listing does not erase obligations to customers who have already bought.
GOV.UK's returns guidance says distance-selling customers generally have 14 days after receipt to notify cancellation and a further 14 days to return the item, subject to exceptions such as certain personalised or perishable goods. It also makes clear that sale items carry the same refund rights as full-price items. Do not allow an AI-generated clearance or withdrawal plan to imply that discontinued goods are automatically excluded from returns. Consumer rules and business-to-business contracts are not interchangeable.
Keep final approval with a named commercial owner, supported by finance and whoever handles customer obligations. If you want help assessing whether AI adds anything to your current product review, a conversation about the evidence and decision is a sensible starting point. The useful outcome may be a better spreadsheet, not new software.
Is This Right for You?
This assessment suits retailers, distributors and product businesses considering whether AI can improve range decisions. It is useful when you have reliable product records and a commercial owner able to challenge recommendations.
It is not a substitute for an accountant, product-safety advice or checking customer contracts. If a spreadsheet already answers the question, keep the simpler approach.
