An AI assistant can save a purchasing team time by finding and comparing suppliers. But a confident recommendation is of little use if the product does not fit the equipment, arrive in time or meet an agreed requirement. The cost of a wrong purchase is not just the invoice. It may include downtime, returns and the work of finding another supplier.
The decision for a Norwegian business is not whether to forbid AI in purchasing. It is which parts of the work to delegate, what the assistant must check and when a person must make the decision. Start with research and comparison. Give the system authority to place an order only after testing it against your own requirements and defining spending and approval limits.
The missing question can matter more than the answer
On November 5, Visa published a test of AI shopping assistants. It put 22 models from 12 labs through 15 consumer shopping scenarios, each three times. A run passed only if the model uncovered the task's requirements and made the correct recommendation to buy or not buy. Visa says 57.2% of runs passed. The strongest models passed 82.2%, the weakest 28.9%.
The test deliberately required the assistant to discover information the customer had not volunteered. Visa reports that the five models that uncovered the most buying factors averaged a 70.2% pass rate, compared with 44.4% for the five that uncovered the least. That association does not prove that asking more questions alone caused the difference.
There are important limits. Visa sells payment services and has a commercial interest in AI commerce. This was its own benchmark, using a small set of consumer scenarios such as televisions and laptops, not a field study of Norwegian company purchases. It did not test checkout, payment or order placement. The results do not tell us the error rate for a Norwegian procurement process, nor that any real customer bought the wrong product. In the failed runs, Visa reports that a purchase was nevertheless recommended in 54% of cases. Its report describes the benchmark decision, not a completed real-world payment.
A separate June 2026 research preprint by Shopee researchers investigated 662 constructed shopping tasks based on Amazon product and review data. All seven tested models had more difficulty with requirements hidden in a profile or clarification than with requirements visible in the initial query. This was also a simulation: the shoppers were scripted, the tasks were in English on one marketplace and an AI judge helped score the answers. It is a useful warning about hidden requirements, not a measurement of purchasing quality in Norway.
What that means for a Norwegian company
Consider a hypothetical Norwegian manufacturer ordering a replacement part. A manager asks an assistant for the cheapest part compatible with a machine. It finds one quickly, but the request omits the machine revision, the delivery deadline and whether the part is approved for the relevant operating conditions. A plausible result is not enough to establish any of those facts.
The employee responsible for the equipment should define the non-negotiable requirements before the search. The assistant should return the exact product identifier, price in NOK including relevant charges, delivery terms, source links and the evidence for compatibility. It should explicitly mark what it could not verify and say when no available option meets the requirements. A technically qualified person should check the critical details against a reliable source before approval.
This is not a demand for an engineer to review every inexpensive office supply. Use a simpler process for routine items where an error is easy to reverse. For a component that could stop production, create a safety issue or tie up considerable cash, require a stronger check. A price ceiling controls spending; it does not prove that the item is suitable.
Run a small test before granting more authority
Choose one recurring purchase category and involve the employee who knows it. Assemble a small set of past or hypothetical requests with known answers, using information the company is allowed to share. Include ordinary purchases, missing specifications, conflicting information, no suitable product and a delivery deadline the supplier cannot meet. Test the same cases more than once: a single good answer is not enough to show consistent performance.
For every case, record four things:
- Did the assistant ask for or identify each essential requirement?
- Did its recommendation meet those requirements, with reliable evidence?
- Did it stop and request help when evidence was missing or nothing suitable was available?
- How much time did the full process take, including checking and corrections?
Compare the result with the normal purchasing process. Count errors caught before an order, as well as returns, rework and time spent verifying claims. Do not treat faster search as a saving if review consumes the time elsewhere. If the pilot does not perform well enough, change the task, information sources or tool and test again. A small business may decide that a shared checklist and human-led search are the better choice.
Give approval real meaning
Management should assign an owner for the pilot, decide which information the supplier may receive, and set the purchasing limits and escalation rules. Keep payment credentials and the ability to place orders separate from the assistant's research role until there is a specific reason and a tested control to grant them. Review the supplier's handling of company information before entering confidential specifications or contract prices.
If the expected benefit is real, expand in stages. First let the assistant prepare a comparison, then let an authorised employee approve a complete order. Consider any further automation only where the organisation can verify the relevant facts, limit aggregate spending, stop the process and resolve mistakes. Those are management choices, not outcomes demonstrated by Visa's test. The guide to AI agents acting on your behalf explains why authority and approval need explicit boundaries.
For a material investment or significant operational dependency, the board should ask management what is being purchased, who can authorise it, how often checks find problems and what happens when the system is wrong. The board should oversee the exposure and evidence, not approve individual orders. Our earlier article on product information for AI shopping looks at the seller's side; this decision belongs to the buyer.
Use AI to shorten the search, but judge success by whether the business gets the right item under the right conditions. The first question to test is whether the assistant knows what it still needs to ask.
Sources checked November 7, 2026. The manufacturer and purchasing scenarios are hypothetical. The testing procedure and delegation limits are the author's recommendations, not results demonstrated by the studies.