01 / AI DOES NOT UNDERSTAND THE BUSINESS BY DEFAULT
A model can summarise sales, classify products or draft a product page in seconds. That does not mean it knows which products deserve priority, what a brand can promise or which outcome matters to the business.
AI works with the data and instructions it receives. It can find patterns and produce convincing answers, but it does not automatically know the trade-offs behind a decision: margin, availability, returns, positioning, regulation or customer trust.
Imagine a product with many visits and few sales. A model might flag its product page as an opportunity for improvement. But perhaps the most popular sizes are out of stock, the price has just changed or a campaign is sending it the wrong audience. The signal is real; explaining it requires context.
Using AI in ecommerce is therefore not about asking what can be automated. It is about identifying which work it can help us do better and who should evaluate the result.
02 / WHERE IT CAN HELP
Catalogue analysis
A large catalogue accumulates inconsistencies that are difficult to spot by hand: attributes written in different formats, incorrect categories, variants with too little information or similar products described in incompatible ways.
A model can group listings that look like duplicates, suggest missing attributes or flag that some trainers include material and recommended use while others in the same category do not. The team can use that list to review cases and correct the structure.
The suggestion should not become a bulk change automatically. An attribute that appears to be missing may not matter for that product; two similar products may have important differences; a suggested category may not match how customers search. AI helps surface candidates for review. The taxonomy remains a decision about the catalogue and its users.
Product content
AI can prepare drafts from reliable product information: dimensions, materials, compatibility, instructions and differences between variants. It can also help find listings that omit information available in internal product data.
For example, if the manufacturer says an accessory works only with certain versions of a device, that compatibility should be clear. A tool can suggest simpler wording, but someone must check that it has not turned “compatible with some models” into “compatible with every model”.
Publishing generated descriptions without checking them against the source creates a specific risk: adding features the product does not have or presenting a model’s inference as a benefit. On a product page, an invented detail is more than a writing mistake; it can lead to a wrong purchase, a return or a complaint.
Sales analysis and anomalies
Models can help summarise data and highlight changes that deserve attention: a sharp drop in conversion, an unusual increase in cancellations or a difference between markets, devices and channels.
That does not prove why the change happened. If mobile conversion falls on a Tuesday, the data alone does not show that the new interface caused it. Traffic may have changed, a variant may have gone out of stock, a payment integration may have failed or a different campaign may have started.
A useful output identifies which segment changed, when it changed and which data was used for comparison. The team can then check the context before deciding. An alert without that review may speed up diagnosis, but it can also speed up a mistaken conclusion.
Customer service
In customer service, AI can classify enquiries, summarise conversations or suggest replies based on approved policies and documentation. This is especially useful when a team receives repeated questions about delivery, returns or product features.
Quality depends on the reference information being up to date. If the returns period changes and the answer base does not, a model may confidently repeat an outdated policy. If an enquiry combines several circumstances, a standard reply may miss the real problem.
A practical approach is to let the system draft a reply and have a person review sensitive, ambiguous or out-of-policy cases. Automating a repetitive response can free up time; making a customer argue with a wrong answer can consume even more.
Automation
AI can also prepare tasks that established rules then carry out: assign an enquiry to a team, suggest a label, extract details from a message or create a draft incident report.
It is useful to distinguish between recommending an action and executing it. Labelling an enquiry may be reversible and cheap to correct. Cancelling an order, changing a price or promising compensation has greater consequences. The greater the impact, the clearer the rules, controls and correction path need to be.
03 / WHEN A PLAUSIBLE ANSWER IS WRONG
An answer can be well written and based on a real pattern yet still be wrong for that business.
Part of the catalogue may be missing, attribute names may be inconsistent or data from different periods may have been combined. A promotion changes normal behaviour; an out-of-stock item changes conversion; a seasonal shift makes comparison with the previous week misleading. A model cannot make up for information it never received or that has been misinterpreted.
What we ask it to optimise matters too. If the goal is to increase sales, the result could favour discounts that reduce margin. If customer service is measured by speed, the number of quick replies that do not solve the problem may rise. A metric summarises one part of the business; it does not replace the decision about which consequences are acceptable.
Context is not an extra to add at the end. It is part of the question. “Which products sell less?” is a different question from “Which products with available stock, comparable traffic and positive margin have lost conversion since the last price change?”
04 / HUMAN JUDGEMENT AS PART OF THE SYSTEM
Keeping human judgement in the process does not mean manually reviewing every result forever. It means deciding what to delegate, what to verify and what should not run without evaluation.
For a content suggestion, review can check that claims match the product data. For a sales alert, it can confirm the period and rule out changes in stock or campaigns. For customer service, it can check policies, exceptions and tone before a sensitive reply is sent.
That work also helps uncover repeated failures. If a model keeps misclassifying a product family, perhaps the instructions are insufficient, the data is incomplete or the catalogue category does not represent the decision it needs to make. The useful fix may be in the information and process, not in asking the model for another answer.
Responsibility does not disappear when a decision is automated. If the system changes a product page, recommends a discount or replies to a customer, the business remains accountable for the effect.
05 / HOW TO START WITHOUT AUTOMATING THE ERROR
A good first use has a limited scope and an outcome someone can check. For example, detect listings with inconsistent attributes in one category, compare the suggestions with the catalogue and record how many were correct before expanding the process.
Before putting it into operation, define:
- the task to improve;
- what information the model needs and where it comes from;
- what a correct recommendation should look like;
- which cases need human review;
- how an error will be corrected and who is responsible.
Then compare the result with the current process. If the model reduces review time but increases compatibility or content errors, the trade-off is poor. If suggestions help find inconsistencies that previously went unnoticed, the team can expand their use gradually.
There is no need to start with a high-impact decision. Start by preparing drafts, grouping incidents or flagging anomalies for someone to investigate. Expand the scope when there is evidence that the system helps and a clear way to control its mistakes.
06 / THE QUESTION TO ASK
AI can reduce repetitive work and help reveal patterns in a catalogue or complex operation. Its output still depends on data quality, how the task is defined and the decisions made by the people putting it to use.
Before adding it to an ecommerce process, ask:
Which decision do we want to improve, what context does it need and who will take responsibility if the recommendation is wrong?
If the team can answer all three questions, AI has a specific role. If it cannot, automating the process will only make uncertainty move faster.
At La Oveja Negra, technology makes sense when it helps us understand behaviour, improve the experience and make decisions with more context. Artificial intelligence can be part of that work; the judgement to interpret its results can too.