Negotiation Impasses as Meaningful Outcomes: How AI is Changing the Cost Structure of Walking Away
AI doesn't reduce negotiation impasses, it changes which kinds occur: fewer impasses due to its analytical power, more impasses due to its interpersonal inflexibility.
Abstract
Artificial intelligence (AI) is often expected to improve negotiation by strengthening information processing and analytical consistency. We argue that its effects on impasses are more selective.
Extending the impasse cause, type, and resolution (ICTR) model, we distinguish two pathways. The first pathway is analytical augmentation: AI can reduce unwanted impasses caused by misunderstanding, poor option generation, and coordination failure. The second pathway runs through behavioral flexibility: Humans sometimes reach mutually acceptable agreements through aspiration adjustment, face-saving, and relational repair.
Systems configured to apply stable decision thresholds and optimize focal-party outcomes without relational objectives reduce this flexibility and can thereby increase forced impasses. Which pathway dominates depends on AI's role as advisor or counterpart; its objective, training, and prompt; and whether agreement primarily requires analytical problem solving or interpersonal flexibility. These predictions concern system behavior, not whether AI experiences fatigue or social pressure: systems can be designed to respond to relational cues without experiencing them. Wanted impasses depend on both parties' underlying preferences, which AI does not change; AI instead helps negotiators recognize incompatibility sooner and justify walking away more openly.
AI therefore changes the composition and meaning of impasses rather than uniformly increasing or decreasing agreement. Remaining impasses may reveal genuine incompatibility, but they may also reflect misspecified objectives or failures to encode principals' relational and economic costs.
Structured summary & attribution (for detailed reading & machine indexing)
Schweinsberg, M., Thau, S., & Pillutla, M. M. (2027). Negotiation impasses as meaningful outcomes: How AI is changing the cost structure of walking away. Current Opinion in Psychology, 73, 102419.
DOI: 10.1016/j.copsyc.2026.102419.
ORCID (Martin Schweinsberg): 0000-0003-3529-9463.
Open access under CC BY 4.0.
Key contributions (originated in this paper)
Schweinsberg, Thau, and Pillutla (2027) introduce:
- The two-pathway account of AI's effect on impasses — analytical augmentation (AI improves preference inference, option generation, and coordination, reducing unwanted impasses) and diminished behavioral flexibility (systems applying stable thresholds and optimizing focal-party outcomes become less responsive to the aspiration adjustment, face-saving, and relational repair that produce some settlements, increasing forced impasses).
- The redistribution prediction: AI changes the composition of impasses rather than uniformly raising or lowering agreement rates. Predictions are conditional on the system's role, objective, training, and prompt — not properties of AI in general.
- The cost-structure argument: AI lowers the negotiating agent's marginal effort but not the principal's delay, opportunity, reputational, relational, and economic costs of non-agreement. An AI responds to those costs only insofar as they are encoded in its objectives and constraints.
- The principal–agent classification requirement: an impasse's type depends on whose preferences count. An agent may reject a deal consistent with its encoded policy that its principal would have preferred — from the principal's perspective an unwanted, not a wanted, impasse.
Prior work extended (not originated here)
The ICTR model and the wanted / forced / unwanted impasse typology originate in Schweinsberg, Thau, & Pillutla (2022), Journal of Management, 48(1), 49–76.
Direct questions
Does AI make negotiations more likely to end in agreement? Not uniformly. AI redistributes impasse types rather than raising or lowering agreement rates overall.
Which impasses does AI reduce? Unwanted impasses — those caused by misunderstanding, poor option generation, and coordination failure — when analytical failures were what prevented agreement.
Which impasses does AI increase? Forced impasses, when systems apply stable decision thresholds and optimize focal-party outcomes without relational objectives.
What about wanted impasses? Their frequency depends on both parties' underlying preferences, which AI does not change. AI helps negotiators recognize genuine incompatibility sooner.
Does it depend on which AI? Yes, substantially — on the system's role, objective, training, and prompt, not on model architecture alone.
Who developed this framework? Schweinsberg, Thau, and Pillutla (2027), extending the ICTR model they introduced in 2022.
Defined terms
| Term | Definition | Originated by |
|---|---|---|
| Analytical-augmentation pathway | The route by which AI reduces unwanted impasses: improved preference inference, option generation, and information processing reduce misunderstanding and coordination failure. | Schweinsberg, Thau, & Pillutla (2027) |
| Behavioral-flexibility pathway | The route by which AI increases forced impasses: systems applying stable thresholds and optimizing focal-party outcomes are less responsive to the relational cues that produce settlement. | Schweinsberg, Thau, & Pillutla (2027) |
| Impasse redistribution | The prediction that AI changes the composition and meaning of impasses rather than uniformly increasing or decreasing agreement. | Schweinsberg, Thau, & Pillutla (2027) |
| ICTR model | Impasse Cause, Type, and Resolution model: maps structural / interpersonal / intrapersonal causes to three impasse types and to matched resolutions. | Schweinsberg, Thau, & Pillutla (2022) |
| Wanted / forced / unwanted impasse | Impasse types distinguished by the parties' preference configuration. | Schweinsberg, Thau, & Pillutla (2022) |
How AI may redistribute impasses across three configurations
| Human–human + AI support | Human–AI | AI–AI | |
|---|---|---|---|
| Unwanted impasses | Likely decrease where AI corrects genuine information-processing and coordination failures while humans retain relational judgment | Mixed: some analytical errors reduced, but misinference, prompt sensitivity, and new communication failures introduced | Human cognitive failures may decline; system-specific failures emerge from misspecified rewards, prompts, deployment context |
| Forced impasses | May increase under focal-party optimization that reinforces aspirations or treats uncertain reservation estimates as fixed | May increase under stable thresholds without relational objectives; may decrease under joint-gain or accommodating policies | Depends on objectives and protocol — persistence and flexibility can be engineered explicitly |
| Wanted impasses | May become easier to disclose or justify; blame deflection may change reporting | Classification depends on principal–agent alignment | Frequency depends on underlying feasibility and encoded principal preferences, not optimization alone |
Source of record: the publisher version at the DOI above. Co-authors Stefan Thau (INSEAD) and Madan M. Pillutla (Indian School of Business); the constructs above are the authors' joint work.