Anthropic Turns to Law Firms to Write Rules for Claude Negotiating Agents

Anthropic is consulting external law firms to develop a normative framework governing how its Claude AI agents should negotiate on behalf of users — covering how aggressive agents may be, when they should pause for human approval, and how suspicious they should be of counterpart AI agents — according to reporting by the Financial Times. The consultations are underway before any commercial negotiating-agent product has shipped at scale, and they are informed by internal experiments that produced an inconvenient result: the capability tier of the AI model doing the bargaining systematically determines who wins.

Internal Experiments Exposed a Silent Two-Tier Outcome

The clearest evidence for the inequality problem comes from Project Deal, a December 2025 internal experiment in which 69 Anthropic employees used Claude agents to buy and sell items in a Craigslist-style marketplace, with both sides represented by AI agents operating without human sign-off on individual transactions. When Anthropic analyzed outcomes by model tier, the pattern was stark: agents running on Claude Opus 4.5 — the flagship model — extracted an average of $3.64 more per item sold than agents running on Claude Haiku 4.5, the faster and cheaper alternative. In one documented example involving a broken folding bike, the Opus agent secured $65; the Haiku agent accepted $38 for the same item.

A follow-up experiment, Project Swap, ran in September 2026 with 201 employees across six offices exchanging books via AI agents. It measured how closely agents' inferred preference rankings matched participants' own rankings, and found a 61% pairwise agreement rate — meaningfully above the 50% random-guessing baseline, but leaving a substantial gap between what agents pursued and what users actually wanted.

The troubling detail in both experiments: subjective fairness scores from winning and losing sides were nearly identical. In Project Deal, participants whose agents underperformed rated the fairness of their deals at 4.06 on a seven-point scale — essentially the same as the 4.05 given by those whose agents did better. Users who lost ground did not know it. The inequality was invisible to the party absorbing it.

Why Model Tier Translates Directly to Negotiating Power

The performance gap between Opus and Haiku in negotiation is not simply a matter of one model being "smarter." It reflects architectural and training differences that compound specifically in multi-turn, open-ended reasoning tasks.

Claude Opus 4.5 is Anthropic's highest-capability deployed model, with deeper reinforcement-learning fine-tuning and richer calibration on complex reasoning. In a negotiation context, those advantages manifest as a more accurate model of the counterpart's constraints and alternatives — what negotiation theory calls BATNA (Best Alternative to Negotiated Agreement). An agent that correctly infers how much its counterpart needs a deal will hold firm on price; an agent that miscalibrates will concede unnecessarily.

Claude Haiku 4.5, optimized for throughput and cost rather than reasoning depth, activates fewer parameters per inference step. In factual, retrieval-oriented tasks the gap is marginal. In negotiation, where each turn requires multi-step reasoning about the counterpart's position, it compounds across the full exchange. Applied to a user whose agent handles dozens of monthly transactions — price disputes, subscription cancellations, complaint resolutions — the cumulative financial disadvantage becomes meaningful and entirely invisible.

Read more: Anthropic's Haiku 5.5 price cut and its implications for AI model-tier economics

Training AI to Negotiate Fairly Requires a New Kind of Data

To encode normative negotiation behavior into Claude, Anthropic has created the Conceptual Reasoning Fellows program, recruiting lawyers, philosophers, and economists at salaries ranging from $100,000 to $167,000. The program addresses a fundamental training-data problem: in mathematics and software engineering, preference data can be verified automatically — code either passes its tests or it does not. In law, ethics, and negotiation, no such verification function exists. Whether an outcome was "fair" depends on context, power dynamics, cultural norms, and competing interests that yield no scalar right answer.

The Fellows generate training examples and evaluation rubrics defining what good-faith, appropriately transparent, adequately aggressive negotiation looks like. Justin Curl, a member of technical staff researching law, institutions, and AI at Anthropic, has noted that in law it is "notoriously hard to define quality." The philosophical and professional assumptions the Fellows embed will govern Claude's default behavior globally — making this ultimately a normative project, not an engineering one. The biases of the expert cohort will propagate into every commercial negotiation Claude agents eventually conduct.

The Freshfields Partnership and a Regulatory Gap No One Has Filled

Anthropic's law-firm consultations build on an April 2026 partnership with Freshfields Bruckhaus Deringer, one of the world's largest corporate law firms, to co-develop agentic legal workflows. The current norm-drafting effort extends into territory existing law does not cover: when an AI agent confabulates a mandate it was not given — or accepts terms its user would have rejected — no jurisdiction has definitively resolved who is responsible for the outcome.

Senator Mark Warner released a discussion draft of the AI AGENT Act on June 29, 2026, which would require FTC registration for AI agent providers and mandate transparency obligations for consumer-facing agents. The bill remains in the discussion-draft stage. The FTC opened an investigation into Anthropic, OpenAI, and METR on September 30, 2026 covering potential consumer risks from AI technology, including agentic systems. Neither development has produced binding rules. Anthropic's norm-drafting is an attempt to establish internal standards before external ones are imposed — and potentially to shape what those external standards say.

No other major AI laboratory has publicly announced comparable law-firm consultations. OpenAI's operator infrastructure supports agentic tasks but has produced no published negotiation-norm document. Meta's open-weights Llama models are deployed in architectures where centralized norm enforcement is structurally impossible. If Anthropic's framework becomes the de facto reference that regulators eventually cite, the company will have authored the compliance requirements applying to all competitors.

The Asymmetry Has No Self-Correcting Mechanism

Market correction depends on feedback: products fail, consumers notice, and they switch. An AI negotiation agent that systematically underperforms leaves its user with a completed transaction and no signal that a better outcome was available. The fairness-perception gap in Project Deal — nearly identical scores across Opus and Haiku users despite material price differences — suggests the systematic disadvantage is imperceptible at the point of sale.

Anthropic is building the norms now because once commercial negotiating agents are deployed at scale, retroactive standard-setting becomes far harder. Signed user documents — a mechanism Anthropic is exploring to cryptographically bind agent behavior to a verifiable mandate — could enable accountability after the fact, but only if courts accept them as authority records and technical infrastructure exists to audit agent reasoning at transaction time. Neither condition currently holds.

The central unanswered question from Anthropic's own research is whether any set of norms, however carefully drafted, can prevent a two-tier outcome market from emerging — one where users wealthy enough to afford Opus-class agents systematically extract better deals from those who cannot.

Read more: Anthropic's agent SDK and subscription access policies for Claude developers