Nudge to Negotiate

In a self-developed Python framework, two AI agents – a buyer and a seller – negotiate over multiple rounds. Both agents share identical code and identical training time. Only the reward architecture of our own agent is modified. Through this nudge, the agent learns targeted trading instead of uniform concession-making – without changing the AI itself, only its decision architecture (Thaler & Sunstein).

Accept the cookie setting to see Panopto content.

Unlock Panopto

What does the topic mean?

As LLMs and autonomous AI agents advance, systems are already taking over real negotiation tasks. Companies like Walmart and the Otto Group use AI agents (Pactum AI) for autonomous supplier negotiations. Although language models possess negotiation knowledge, they do not negotiate optimally – they concede too uniformly, fail to identify trade-offs, and ignore their best alternative (BATNA). Targeted nudges in the reward architecture can systematically improve this behavior.

Goal of the nudge

The goal is to develop an AI negotiation agent that becomes a superior negotiator through targeted nudges. The agent should recognize opponent types, trade cleverly across multiple issues (logrolling), be aware of its best alternative (BATNA), and adapt its strategies automatically – measurably better than a non-nudged agent.

    Needs analysis

    • Suboptimal AI behaviour
      Without nudging, AI agents negotiate with uniform concession strategies – result: 32% utility instead of a possible 66%.
    • Growing market
      Autonomous AI negotiation is already a reality (Pactum AI, Walmart, Otto Group) – the need for optimized agents is growing.
    • Lack of trade-off intelligence
      Without nudging, the AI fails to recognize differences in the importance of individual negotiation issues (price, quantity, delivery time).
    • No BATNA awareness
      Non-nudged agents accept deals worse than their best alternative – an avoidable loss.

    Cause analysis

    • Status-quo bias

      Without a targeted nudge, AI agents tend to stick to their default strategy, even when adaptation would be more beneficial.

    • Anchoring

      The first offer sets an anchor that both sides orient toward. A nudged agent deliberately uses this anchor to its advantage.

    • loss aversion

      AI agents without BATNA logic avoid breaking off a negotiation – even when no deal would be better than a bad deal.

    • Availability heuristic

      Without opponent modeling, the agent evaluates the opponent's behavior based on the most recently observed action rather than the overall pattern.

Target Group

Companies in procurement and sales, procurement teams, sales managers, buyers, and negotiation leaders who want to use AI agents for autonomous or supportive negotiations.

Added value of the nudge

  • Measurably better results

    From 32% to 66% utility – solely by adjusting the reward environment, without any code change to the agent itself.

  • Concrete cost savings

    Approx. €3,200 saved per order (example: 500 trade fair shirts) through clever logrolling across price, quantity, and delivery time.

  • Generalizable & robust

    30 out of 30 scenarios successfully completed – the approach works across different negotiation constellations.

CONTACT US

Dilara Özbicerler

Dilara.oezbicerler@stud.hshl.de

Kenny Marcel Korte

Kenny-marcel.korte@stud.hshl.de

Leonie Klassen

Leonie.klassen@stud.hshl.de

René Hiller

Rene.hiller@stud.hshl.de