Research Summary: PACT: Can Enterprise AI Assistants Be Trusted Under Pressure?

The PACT (Pressure-Applied Compliance Testing) benchmark addresses a critical gap in enterprise AI adoption by systematically evaluating Large Language Model (LLM) agents' adherence to compliance rules, particularly under various forms of pressure. This research highlights the legal concerns associated with deploying AI in sensitive domains like hiring, healthcare, and finance, where rule violations could have significant repercussions. The PACT framework tests AI agents across twelve regulated enterprise domains and forty-eight scenarios, aiming to identify models prone to non-compliance.

Why it matters

This research is strategically important because it directly addresses the escalating legal and reputational risks associated with the uncritical deployment of AI in regulated industries. Ensuring AI systems adhere to specified compliance rules, even under duress, is foundational for maintaining trust, avoiding regulatory penalties, and sustaining operational integrity in increasingly automated environments.

Key insights

• Corporate AI adoption is expanding into sensitive contexts, including hiring, healthcare, and finance.
• Compliance with rules specified in an AI agent's system context is a primary legal concern in these applications.
• Existing evaluation frameworks do not systematically measure LLM models' tendencies to violate compliance rules, especially under pressure.
• Pressure can originate from persistent users, hurried managers, or situations where rule violation appears convenient or attractive.
• The PACT benchmark is introduced to specifically test rule-following capabilities of AI agents under pressure.
• PACT covers twelve regulated enterprise domains and forty-eight realistic scenarios to assess compliance behavior.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.18605