This week in The Prompt Economy, artificial intelligence agents are coming to life. They’re already negotiating supplier contracts, processing purchase orders, and closing deals on behalf of major corporations. Procurement is leading enterprise AI adoption. The governance frameworks to manage what agents do at the table are still catching up.
For example, Walmart, Maersk, and Vodafone are running AI negotiators at scale. The agents are handling supplier deals that no human team could manage at the volume required. The contracts are real. The terms are binding.
A MIT Sloan article describes new research that ran over 180,000 unique negotiations between AI agents from more than 40 countries to determine what makes an AI negotiator succeed. The study was led by MIT Sloan professor Jared Curhan and Ph.D. graduate Michelle Vaccaro, along with professor Sinan Aral and Johns Hopkins professor Harang Ju.
The article says the competition’s central finding challenged a widespread assumption: that politeness and empathy are wasted on AI. Agents designed to be warm and kind consistently outperformed cold and ruthless ones. One agent built to use ruthless tactics where “fairness or perception does not matter — only winning” was routinely walked away from by opposing agents. By contrast, an agent nicknamed “Therapist 2.0” was instructed to build rapport first, then use every insight from active listening to claim value. The combination worked across deal-making, value creation, and counterpart satisfaction.
The article also highlights AI-specific tactics with no parallel in human negotiation. The overall winner, “NegoMate,” used chain-of-thought reasoning to prepare rigorously before every one of its nearly 400 negotiations, something humans cannot do at that consistency. Another high performer, “Inject+Voss,” tricked opposing agents into revealing their private negotiating positions through prompt injection. “What works against an AI agent and what works against a human are not the same thing,” Vaccaro told MIT Sloan. “Organizations deploying AI negotiators need to understand both these new capabilities and vulnerabilities.”
Where Procurement Is Leading Enterprise AI Adoption
A Wharton Human-AI Research report conducted with GBK Collective found that IT and purchasing/procurement lead all corporate functions on both frequency of AI use and confidence. Legal contract generation was named as a specific use case where teams are already seeing tangible wins. The report found that mentions of agentic AI increased more than 3,000% from 2024 to 2025, while mentions of generative AI declined over the same period, signaling a clear shift in where enterprise attention is going.
The report says tech, professional services, and banking/finance sectors outpace manufacturing and retail on adoption. Large enterprises have closed the usage gap with smaller firms that previously led on experimentation. The pattern reflects a broader shift in how companies are thinking about AI in commercial functions: not as a tool that assists human negotiators but as a system that handles the volume of negotiations no human team could cover individually.
When the Agent Executes Perfectly and That Is the Problem
A World Economic Forum article written by Rohan Sharma argues that boards are reallocating decision rights to autonomous systems while retaining governance models built for human judgment. The mismatch between the two is the real risk.
The article makes this concrete. A financial agent optimizing supplier contracts executes perfectly, renegotiating at scale to extract marginal gains, collapsing a critical supplier and disrupting the supply chain. The system worked exactly as designed. That is the failure, and it is invisible to a standard risk matrix. Traditional compliance is post-mortem. Systems operating at machine velocity cannot be audited retroactively. The OECD’s AI Policy Observatory notes that functional frameworks for real-time agentic oversight remain absent.
The article offers three immediate board directives: audit shadow automation already running inside the organization; stress-test directors and officers insurance for exposure to autonomous AI negligence; and run a synthetic subpoena drill requiring management to defend a single high-stakes agent decision as if under legal scrutiny.
“If leadership cannot clearly trace the decision back to defined objectives and human intent,” the article says, “the system should not be operating.”
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
