Events

Negotiation Transformed: How AI Agents Negotiate with Suppliers at Scale

Most of procurement's work never gets reached: suppliers go unengaged, savings leak, and teams stretch across dozens of systems. Pactum's Digital Workers change that, negotiating autonomously with suppliers across Indirect, Direct, and Supplier Management for enterprises such as Walmart, AstraZeneca, and Honeywell. This session shares what we have learned: how buyers set strategy and guardrails, how suppliers respond to an AI counterparty, and where human judgment still matters. William Gu will demonstrate a live supplier negotiation, and we will close with questions of trust, fairness, and the changing role of the negotiator.

From Contract Embellisher to Negotiation Machine: The Use of AI in Contracting

In my contribution, which is based on a post published on the Oxford Business Law Blog, I will argue that AI tools can be used as a “Contract Embellisher”: the parties to a contract deploy them to improve an agreement they have already concluded. Such tools could identify Pareto improvements, drawing on vast amounts of commercial data and reducing transaction costs. This is less risky and less complex than using AI to negotiate a contract in the first place because the parties already have an agreement to fall back on. As increasingly sophisticated and competitively priced AI Contract Embellishers emerge, they could also be used to assist in negotiating new contracts and, eventually, to automate the negotiation process entirely. Rather than each party deploying its own AI agent against the other, jointly appointed AI “honest brokers” could create more value and help divide it fairly. Whether this more cooperative model prevails may depend crucially on competition between AI providers and in the markets for the goods and services being negotiated.

Multimodal Interaction Models: Models that see, hear, speak and act in real time

As AI agents move from chat windows into phones, cars, and physical spaces, they must perceive, reason, and respond across
modalities in real time — listening, watching, and speaking in 200-millisecond micro-turns. This talk introduces the interaction
model: a single model with a shared token space across text, speech, images, video, and screens that streams continuously and handles
turn-taking, interruptions, and proactive behavior. I present two advances from our work: a low-bit tokenizer aligning speech tokens
one-to-one with text tokens, and a hybrid design where a real-time interaction model talks in text to deliberative agents behind it,
post-trained with supervised fine-tuning and reinforcement learning on task success and latency. The core insight: interaction is a timing
problem, not just a reasoning problem. Interaction models will unlock many new applications, including negotiation — either coaching humans in such skills or performing on their behalf.

Cicero: A Negotiation AI Agent in the game of Diplomacy

AI has big potential in negotiation. In this talk, I will cover how to build Cicero, the first human-level AI agent that can persuade, negotiate, collaborate and cooperate with human players in the game of Diplomacy. I will discuss how to understand other players’ intents, how to detect unfaithful messages, and how to generate dialogues grounded in the game.

Coming soon! 

Autonomous AI tools at Crosby: The AI-native Law Firm

Most discussion of AI in negotiation focuses on helping people work faster: drafting responses, flagging risks, and suggesting positions. But what does it take for AI to do the work itself? This talk traces Crosby’s shift from AI-assisted to fully autonomous legal work, the seemingly basic but critical evaluations that make autonomy possible, and includes a demo of how Crosby works in practice.

LegalBench2: Learning Expert Preferences in Law

Prior research on legal AI has focused on building and evaluating AI systems on tasks with objectively correct answers that experts largely agree upon. However, much of legal work is inherently subjective: there may be no single correct answer, and experts may disagree about which answers are better or worse. This talk introduces LegalBench2, an ongoing research effort to develop benchmarks that evaluate AI systems against expert preferences on subjective legal tasks. We argue that building AI systems for such tasks requires moving beyond one-size-fits-all notions of quality toward personalized systems that can learn and adapt to the preferences of individual attorneys.

Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender

People increasingly consult generative artificial intelligence (AI) while reasoning. As AI becomes embedded in daily thought, what becomes of human judgment? We introduce Tri-System Theory, extending dual-process accounts of reasoning by positing System 3: artificial cognition that operates outside the brain. System 3 can supplement or supplant internal processes, introducing novel cognitive pathways. A key prediction of the theory is “cognitive surrender”—adopting AI outputs with minimal scrutiny, overriding intuition (System 1) and deliberation (System 2). Across three preregistered experiments using an adapted Cognitive Reflection Test (N = 1,372; 9,593 trials), we randomized AI accuracy via hidden seed prompts. Participants chose to consult an AI assistant on a majority of trials (>50%). Relative to baseline (no System 3 access), accuracy significantly rose when AI was accurate and fell when it erred (+25/-15 percentage points; Study 1), the behavioral signature of cognitive surrender (AI-Accurate vs. AI-Faulty contrast; Cohen’s h = 0.81). Engaging System 3 also increased confidence, even following errors. Time pressure (Study 2) and per-item incentives and feedback (Study 3) shifted baseline performance but did not eliminate this pattern: when accurate, AI buffered time-pressure costs and amplified incentive gains; when faulty, it consistently reduced accuracy regardless of situational moderators. Across studies, participants with higher trust in AI and lower need for cognition and fluid intelligence showed greater surrender to System 3. Tri-System Theory thus characterizes a triadic cognitive ecology, revealing how System 3 reframes human reasoning and may reshape autonomy and accountability in the age of AI.

The Psychological and Organizational Consequences of Delegating to AI    

As AI shifts from a tool people use to an agent that coaches, communicates with, and manages them, its consequences reach well beyond efficiency. This talk explores how AI changes the psychological experience of receiving feedback, for whom those changes matter most, and the organizational implications of delegating human roles to machines.