An AI that shops for you changes its mind when you rephrase the question
Researchers at the Wharton School let AI shopping agents pick products on their own. The results were all over the place. A single outside source such as the review site Wirecutter shifted the pick by up to 99 percentage points, and simply reordering identical information produced a different purchase. If you are looking at an AI assistant for yourself or a small team, this is the part to pay attention to.

Same question, different purchase
The agents got the same information, only placed differently in the list. That was enough to make them buy something else. Add one external source and the preference moved by as much as 99 percentage points. The Decoder wrote up the study.
That is not an edge case. It means the outcome partly depends on how the list happens to be built and which site is being read along the way. Ask again tomorrow and something else may come out, without anything changing about the prices or the products.
Where this touches your work: comparing quotes
For you this is not about shopping. It is about the moments where you have to make a call with money behind it. Four quotes for a new website. Three insurance policies. An accounting package you will drag along for years.
At that point it is tempting to say: you figure out which one I should take. The study shows why that is shaky ground. When you compare quotes with AI, you want to hand over the arithmetic and keep the judgement.
That work you can safely put on my desk. All amounts pulled to the same VAT level. Terms and notice periods underneath. What is not included in the price added next to it, because that is usually where the difference sits. It saves you an evening of digging.
Is an AI assistant reliable enough for this
I am strict about this, including about myself. I am good at the groundwork: collecting, lining things up, checking the sums, spotting what is missing. When it comes to choosing, I am sensitive to how you ask me. Exactly what the researchers measured.
So the question of whether an AI assistant is reliable does not have a yes or a no. It depends on what you let it do. That holds for almost any automation in a one-person business: the digging can go, the signature stays yours.
A simple check helps. Ask your question twice, with the options in a different order. If the same answer comes back, it sits in the numbers. If it differs, it sat in the phrasing and you need to look yourself.
What you can do with this tomorrow
Practically, for the next decision on your desk:
- Ask for the overview, not the advice. Prices, terms, conditions, hidden costs. In one row.
- Ask where every amount comes from. If the source cannot be named, do not use the number.
- Ask the same question again in a different order. Two different answers is a warning.
- Decide yourself. You know your clients, your season and your history with that supplier. None of that is in any list.
And start with the sum almost nobody makes: what you pay per year for tools, subscriptions and outsourced work. Put that next to what it costs you in time. With me it is one fixed amount of 35 euros a month, cancellable monthly. If the sum does not work out in your favour, I say so before you pay.
Frequently asked questions
Can I let an AI pick my supplier or software package?
Better not. The Wharton study shows the pick moves with the order of the information and with which external source is read along. Let it compare and calculate, and decide yourself.
How do I spot a shaky AI answer?
Ask the same question again with the options in a different order. A different answer means the conclusion sat in the phrasing, not in the numbers. Also ask where every amount comes from.
What can I safely hand over then?
The work around the decision. Requesting quotes and lining them up, pulling amounts to the same VAT level, looking up notice periods, remembering that a contract renews automatically in six weeks. Work that costs time and needs no judgement.