Prince Mario-Max Schaumburg-Lippe: AI Agents Learn to Negotiate, and Lawyers Are Setting the Rules

Here is a sign of how fast AI agents are growing up: Anthropic is talking to law firms about negotiation tactics. Not about contracts for itself – about how its agents should behave when they sit across a table, digital or otherwise, and bargain on a user’s behalf against other companies’ agents. The Financial Times reported the consultations on Sunday, citing two people familiar with the company’s thinking.

The questions on the table are wonderfully concrete. How much should an agent check back with its user mid-negotiation? What may it concede without asking? And how wary should it be of a rival agent’s tactics without becoming, in the FT’s memorable framing, paranoid?

From answering questions to making deals

Chatbots answered. Assistants drafted. Agents act – booking, buying, scheduling, and now, negotiating. Each step raises the stakes of a mistake. A wrong sentence costs nothing; a wrong concession costs money. That is why the norms being sketched in these consultations matter more than any single product launch this month. They are, in effect, the first draft of commercial etiquette for machines that deal.

Law firms are a shrewd choice of tutor. Negotiation is their native craft, and its rules – authority limits, walk-away points, what must never be disclosed – map almost perfectly onto the settings an agent needs. A good negotiator knows their mandate to the dollar. An agent, properly built, can know it to the cent and never get tired, flattered or rushed. The failure modes are different too, which is the point of asking lawyers before the failures happen in the wild.

The questions that decide whether this works

Three design choices will determine whether agent negotiation earns trust. First, the check-back threshold: interrupt the user too often and the agent is just an expensive messenger; too rarely and it is a liability. Second, disclosure discipline: an agent that reveals your reservation price has lost before it starts, so what it may share has to be engineered, not improvised. Third, counterparty skepticism: agents will meet agents designed to extract concessions, and a calibrated wariness – skeptical of pressure, immune to flattery – has to be tuned like any other parameter.

The trend line is consistent with everything else crossing the wire this week, from agents logging six times the hours of the humans beside them to security teams deploying supervised virtual workers. Autonomy is arriving in supervised, rule-bound form – and in negotiation, the rules are quite literally being written with lawyers in the room.

The bottom line

Anthropic declined to comment to the FT, so treat the details as informed reporting rather than a product announcement. But the direction is unmistakable, and it is a healthy one: the company building agents that will haggle for you is first asking how a good haggle should work. When your software starts negotiating your next deal, you will want it to have had exactly this education.

Trust is the real product

Step back, and the law-firm consultations reveal something larger about where agents are headed. The bottleneck for agent commerce is not capability – models can draft, compare and concede already. The bottleneck is delegation: the moment a user lets software spend their money or bind their name, the questions become legal and psychological, not technical. How much authority did I actually grant? What happens when two agents deadlock, or when one exploits the other’s politeness? Anthropic going to lawyers first suggests it understands that the winning agent will not be the boldest negotiator. It will be the one whose mandate, limits and paper trail a cautious person can inspect and trust. Expect rivals to make their own pilgrimages to the law firms, and expect a new professional specialty to bloom at the intersection: counsel for agents, fluent in both contract law and system prompts. The deals of 2028 may be struck by software. The rules they are struck under are being written, sensibly, by humans in conference rooms right now.

Supervised autonomy is the week’s theme in security too, where Axonius just handed teams an AI workforce of virtual workers.

Also published on: