Project Swap: AI can bargain, but does it understand what you want?

Tasking an AI to find the "best value subscription" sounds simple. But for one person, value means the lowest price, for another, it's Ukrainian language support, and for a third, it's offline work. If the assistant misses the priority, it might diligently solve a completely different task.

An experiment with books instead of predictions

In Project Swap, released by Anthropic on September 24, 2026, 201 employees from six offices handed over a book for exchange. After a short conversation about reading preferences, Claude agents negotiated among themselves. People separately ranked ten books so that researchers had something to compare the agents' ratings against.

According to the authors' report, after a five-minute conversation, the agent's ranking matched the human's in 61% of book pairs. This is not a customer satisfaction percentage. One of the main limitations of the result is incomplete knowledge of the owner's preferences. In repeat runs, the choice of model influenced the negotiations more than the specified behavior style. A sample of employees from a single AI company and a book exchange does not represent all markets. Design, results, and limitations of Project Swap →

What to check before assigning "choose for me"

Try to formulate the conditions so that another person could also make a choice. "I want a normal AI service" defines almost nothing. "I need a tool for Ukrainian subtitles, ten short videos per month, and export without a watermark" already sets a direction.

Next, separate the mandatory from the desirable. If Ukrainian subtitles are mandatory, a lower price does not compensate for their absence. If integration with a specific editor is only a convenience, it can be weighed against time savings. This distinction helps explain the trade-off in plain language.

Hypothetical example: two plans for a video creator

Let the first hypothetical plan cost 300 UAH and provide five exports, and the second — 500 UAH and twenty. For two videos, the cheaper option might be sufficient. For twelve, it no longer meets the task without additional conditions. These numbers are an educational example, not the prices of real services.

Ask the agent to also show the billing period, the price for additional usage, and what it could not find out. If a card says "from 300", that is not yet the full cost of your script. It is better to leave missing information unknown than to inadvertently turn it into zero.

Brief for an assistant choosing offers

Below is our comparison template. It is also suitable for a regular chat if you provide the current terms of the offers yourself. Without access to sources, the assistant cannot confirm the current price.

Help me compare the offers I will provide. Goal: [what I want to achieve]. Mandatory conditions: [what makes an option unsuitable]. Desired conditions: [what can be compromised]. Total budget and period: [amount, currency, month/year/one-time purchase]. First, ask up to three questions that could change the choice. For each offer, show: compliance with conditions, total known cost, unknown data, source. Note any assumptions. If no option fits, say so. Prepare a recommendation for my decision; do not buy or send anything without a separate instruction.

A check you should do yourself

Before reading the recommendation, write down one reason why you would decline the purchase. For example, a mandatory annual payment or the lack of a required export. Then, check if the agent took this condition into account and provided confirmation.

A second useful step is to ask what one new circumstance would change its recommendation. A good answer can be very specific: "if you do more than twenty exports" or "if the team requires shared access." Check this exact circumstance in your own work.

Next, you can take our AI subscription test on three tasks. And if you are creating your own agent for product selection, start with transparent comparison and clear user confirmation. The quality of a decision should be evaluated by the fulfilled need, not by the persuasiveness of the message.

More from this collection

• Claude and formal verification: what was proven for 123 libraries
• Why AI forgets important things: JAM and memory for a specific task