Intent Drift at SME Scale: Deployment Practice, Not Model Capability, Determines Agentic Compliance

arXiv:2609.05975 · cs.CY, cs.AI · Submitted 2026-09-05 · Read on arXiv

cs.CY, cs.AI

Submitted: 2026-09-05

Updated: 2026-09-05

Comments: 22 pages, 3 figures, 4 tables. Code, data and all run logs: https://github.com/IliaHatesCoding/intent-drift-sme

Code: https://github.com/IliaHatesCoding/intent-drift-sme

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Terminology

• Ablation — 1× in this paper · explained in 1 episode
• Agentic AI — 1× in this paper · explained in 1 episode

Related papers

• Reasoning Enhances Robustness to Prompt Injection in LLM-Based Consensus
• Generative AI Purpose-built for Social and Mental Health: A Real-World Pilot
• PersonaMem-v3: Toward Omni-Platform Personal Intelligence for Holistic User Understanding, Recommendation, and Agentic Tasks
• What is an intelligent system?
• AI University: An LLM-Powered Learning Assistant for Engineering---A Finite Element Method Case Study
• Generative AI Use in Entrepreneurship: An Integrative Review and an Empowerment-Entrapment Framework