How do you separate real AI from marketing in wholesale platforms?
Almost every B2B ordering platform now advertises AI. A buyer can test those claims with specific questions about data, behaviour, evidence and cost, and by asking for a demonstration on its own data.
KEY TAKEAWAYS Summary by the editors
- A real AI feature can be described by its input data, its output, how its quality is measured and what a human can override; a claim that cannot be described this way is probably marketing.
- The US Federal Trade Commission stated in 2024 that there is no AI exemption from the laws on the books, and brought five actions against companies that overstated AI capabilities or sold AI tools used to deceive.
- Gartner commentary reported by CIO Dive notes that ROI is harder to calculate when many people use a tool and workflows are not clearly defined, so ask vendors for results in named workflows, not general claims.
- McKinsey's 2026 survey found that 37% of respondents attribute some EBIT impact to AI and that about 20% say operating costs, including tokens, have constrained use, so ask how pricing behaves as usage grows.
- The most reliable test is a pilot on your own order history, product data and customers, with a pre-agreed success measure and a comparison to the current method.
You separate real AI from marketing in a wholesale platform by asking what the feature takes as input, what it produces, how its quality is measured and what happens when it is wrong, and then testing it on your own order and product data. A genuine feature can answer these questions specifically. A marketing claim usually answers with adjectives. Wholesale is a useful test case because its data is structured and its outcomes (re-orders, margins, returns, sell-through) are measurable, which means vendors have fewer excuses for vagueness.
Why is AI language so common in platform marketing?
Buyers and investors expect AI, so vendors label features accordingly. Some of those features are conventional rules, filters or statistical methods that have been on the market for years. Calling them AI is not necessarily dishonest, but it can mislead a buyer into expecting adaptability the feature does not have. Regulators have noticed. In September 2024 the US Federal Trade Commission announced Operation AI Comply, five enforcement actions against companies that overstated AI capabilities or sold AI tools used to deceive, and its then Chair said there is no AI exemption from the laws on the books. Those cases involved consumer schemes and tools, not wholesale software, but the principle that AI claims must be substantiated applies more widely.
What questions reveal whether a feature is real?
- What data does it use, from our systems and from others, and how fresh is it?
- What exactly does it output: a score, a suggestion, a draft, an automatic action?
- Is it a rules engine, a statistical model, a machine learning model or a large language model, and who built it?
- How is quality measured, on what data, and can we see the results for a customer like us?
- What can a user override, and is the override logged?
- What happens when data is missing, such as a new retailer with no order history?
- What does it cost, and how does the price change as usage grows?
What do typical wholesale AI claims mean in practice?
| Claim | What it should mean | Test |
|---|---|---|
| Smart re-order suggestions | A model proposes quantities from past orders, stock and sell-out where available | Run on last season's data; compare with what the account actually ordered |
| AI product recommendations for buyers | Suggestions for retailers based on their history and similar accounts | Check relevance on a sample of real accounts, and see how new products are handled |
| AI-generated product descriptions | A language model drafts text from your attributes | Review a sample for factual errors in sizes, materials and care labels |
| Automated order entry from emails or PDFs | Extraction of order lines into structured form | Measure the share of lines extracted correctly on your real documents |
| Assistant for sales reps | A chat interface to account and order data | Ask ten real questions, and check each answer against the source |
| Demand forecasting | Statistical or machine learning forecasts of sell-in or sell-out | Back-test against held-out periods; compare with the current method |
Why does your own data matter in a test?
Demonstrations on a vendor's sample data show what the feature can do under good conditions. Your data shows what it does with your product codes, size scales, seasonal drops and irregular retailers. Ask for a pilot or a back-test on your history, with a success measure agreed in advance and a comparison to what you do today. A feature that cannot be tested on your data is, at best, unproven for your case.
Be wary of results quoted without a baseline or sample size, and of case studies that describe satisfaction rather than outcomes. Gartner commentary reported by CIO Dive notes that ROI is harder to calculate when many people use a tool and workflows are not clearly defined. A vendor that can name the workflow, the metric and the comparison is more credible than one that cites an overall percentage.
What risks should a wholesale buyer weigh?
- Data rights: whether the vendor may use your order and product data to train models that serve others.
- Accuracy and accountability: who is responsible if a suggested order or generated text is wrong.
- Dependence: how hard it would be to move your data and configuration elsewhere.
- Cost: usage-based pricing can increase with adoption; McKinsey's 2026 survey found about 20% of respondents say operating costs, including tokens, have constrained AI use.
- Regulation: transparency duties for chatbots and generated content under Article 50 of the EU AI Act apply from 2 August 2026, with a grace period to 2 December 2026 for some systems already on the market, according to a published summary of the Act.
What does a fair evaluation process look like?
Write down the three outcomes you want, such as fewer manual order lines or better re-order accuracy. Ask each shortlisted vendor to show how each feature contributes, using the questions above. Run a time-boxed pilot on real data with a defined success threshold. Involve the sales reps who will use the tool: if they do not trust the suggestions, adoption will fail whatever the technology. Finally, check references with customers of similar size and product mix, and ask them what the feature actually changed.
McKinsey and the Business of Fashion's State of Fashion 2026 states that leaders should move beyond small pilots toward a fundamental rethink of how organisations work. In wholesale, that means asking not only whether a feature is real AI but whether it changes the ordering process, because the value is in the changed process, and a feature that only adds a screen to an unchanged process rarely moves the numbers.
Frequently asked questions
How can I tell if a B2B platform's AI is real?
Ask what data it uses, what it outputs, how quality is measured and what a user can override. Then test it on your own order and product data against your current method.
Is a rules-based feature marketed as AI misleading?
Not necessarily, but it may set the wrong expectation. A rules engine does not learn or adapt unless it is rebuilt, so ask the vendor to say precisely which parts use machine learning or language models.
What should a pilot of AI re-order suggestions measure?
Compare suggested quantities with actual orders on historical data, and then with outcomes such as sell-through and stock levels where available. Agree the success threshold in advance.
Do regulators act against exaggerated AI claims?
Yes. In 2024 the US Federal Trade Commission announced five actions against companies that overstated AI capabilities or sold AI tools used to deceive. The cases concerned consumer schemes, but the principle that claims must be substantiated is general.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
SOURCES
- US Federal Trade Commission: Operation AI Comply press release
- CIO Dive: Ballooning AI budgets and ROI pressure (Gartner)
- McKinsey: The state of AI in 2026, On the road to ROI
- McKinsey and Business of Fashion: The State of Fashion 2026, When the rules change
- Future of Life Institute: EU AI Act high-level summary
