Evidence, not verdicts
What designing offers taught me about where AI should stop.
In the Offer Insights prototype, a buyer offers $170 for a turntable listed at $219.99. The seller can accept, counter, or decline.
What should the product add to that decision?
“Seems like a good offer” sounds helpful. So does “Try a higher offer.” I explored guidance like this. It gives people a clear next step, but it also puts the marketplace in the middle of their negotiation. Calling an offer good tells the seller what to accept. Asking a buyer to offer more favors the seller.
I wanted to help both people keep the conversation going. That meant deciding where our guidance should stop.
A clear answer can hide a difficult judgment
An offer arrives as a number, but the decision behind it is personal. A seller may want to sell quickly, wait for a better price, or recognize a returning buyer. A buyer may have a firm budget or see something in the item’s condition that changes its value to them.
The product cannot know all of that.
Even a comparison with recent sales involves choices: which items count as similar, which conditions qualify, and how far back to look. Those choices affect the range we show.
AI can turn that incomplete picture into a confident sentence. My concern is that someone acts on it without understanding what supports it—or what is missing.
The trade-off I chose
With my product partner, I developed an Offer Insights vision around recent comparable sales and the history of the negotiation. We gave up the simplicity of telling someone what to do. In return, people could inspect the basis for their decision.
That choice shaped the interaction:
Shared price context. Buyers and sellers see the same comparable-sales information: the recent price range, how many items sold, and the last offer in their conversation.
Room to disagree. A buyer can offer below the range without a warning or a block. A range is context, not a minimum price.
No preferred outcome. Accept, Counter, and Decline receive equal visual weight.
A deliberate fallback. When comparable sales are insufficient, the price section disappears. The offer history remains useful on its own.
Showing evidence still influences a negotiation. A price range can become an anchor. That makes the choice of evidence—and the decision to withhold it—part of the design responsibility.
One unresolved example is listing age. It might help a seller decide whether to adjust their expectations. Showing it to a buyer could create pressure on the seller. My starting hypothesis is to keep that information with the seller, but I would test that choice before treating it as a rule.
What an experiment taught me
In related offers work, I designed Repeat Buyer: a small label showing that a buyer had purchased from the seller before. It gave the seller a relationship fact to consider without recommending a response.
In a US A/B test, offer acceptance increased by 3.3%.
That result supports the value of timely context. It does not validate a price range or prove that Offer Insights will work. Those are different signals, with different risks.
Offer Insights remains a concept, and the prototype uses sample prices. The first question I would test is whether the evidence helps more first offers become counteroffers. I would also check whether people understand the range’s limits, or read it as the marketplace telling them what an item is worth.
Where I draw the line
For AI guidance that affects someone’s money or choices, I want people to understand what supports a suggestion, recognize its limits, and have room to disagree.
Sometimes that means offering less guidance than we could generate. I consider that a product decision worth making deliberately.