What makes an agreement good?

Warmth changes the path to agreement.

The competition reported by Vaccaro and colleagues includes 182,812 simulated negotiations. Warmer agent prompts were associated with more deals and stronger aggregate outcomes. But among negotiations that reached agreement, warmer agents earned fewer points and claimed less value. Dominance helped value claiming while also carrying an impasse tradeoff. [S2]

This distinction matters for a procurement product. A system can appear successful by accepting too readily. It can also pursue a small additional concession until the supplier walks away. Neither a high agreement count nor an aggressive counteroffer is enough to judge performance.

Useful conversation and firm boundaries can coexist.

Question-asking, gratitude, and positive language were associated with warmth in the study’s transcript analysis. These findings motivate testing clear, respectful dialogue that uncovers constraints, rather than assuming adversarial language is necessary. They do not guarantee a better supplier price, and the study’s AI–AI scenarios do not substitute for field testing with procurement teams. [S2]

Imagine a supplier unable to reduce price but able to shorten lead time. A question about scheduling flexibility could reveal a useful option. This is our illustrative procurement example, not an observed result from the competition. Its value depends on the buyer’s actual demand, inventory policy, and service needs.

Measure several kinds of failure.

AgenticPay evaluates feasibility, efficiency, and welfare, extracting structured actions from natural-language dialogue. Its evaluated models differ in price-bound violations, timeouts, and buyer versus seller performance. Each task is executed once per model under the reported protocol, so those scores should not become a universal ranking for production procurement. [S3]

For our product, this points toward a pilot scorecard that separates commercial value from process effort and control failures. The following measures are proposed evaluation criteria; no Procurement Lead pilot result is being claimed.

Scroll the table to see all columns →

Separate laboratory evidence from customer evidence.

Waditwar’s procurement discussion offers useful architecture and adoption questions, but its business examples do not validate our product’s savings or cycle times. We will need a defined baseline, comparable cases, and recorded assumptions before reporting an improvement. [S5]