AI Persona, Service Consumption, and User Intent Entropy: Field Experimental Evidence from an LLM Platform

Key Findings

• — Relational versus non-relational persona: service sessions +8.1%, cumulative duration +10.6%, chat rounds +24.2%, user intent entropy +5.8%, files produced +12.3%, distinct goals +12.1% over the sample period (4,865 treatment versus 4,721 control users; 16,511 service sessions)
• — First session by entry intent: no significant effect for task-execution users; socialisation users +16.0% chat rounds and +9.2% intent entropy with no increase in outputs; knowledge-exploration users +15.9% chat rounds with +18.7% files and +18.9% distinct goals
• — Subsequent sessions: session count +12.2% for task-execution users and +49.4% for socialisation users (not significant for knowledge exploration); effects on session count and intent entropy strengthened over the 18 days while output effects stayed flat
• — Process evidence: the relational persona raised intent-transition entropy for socialisation users (+11.8%) and intent-continuation probability for knowledge-exploration users (+12.8%)
• — The authors frame persona as an operational lever that consumes compute without proportionate output for socialising users; no wellbeing, dependency or harm outcomes were measured

Methodology Notes

Two-arm randomized field experiment on newly registered users of Agnes AI (public launch July 2025); six-day pre-treatment window 6 to 12 December 2025, assignment 12 December 2025, 18-day treatment; one persistent conversation thread per user; sessions split at 30 minutes of inactivity; entry intent classified from the first message (before persona exposure) into task execution, socialisation and knowledge exploration; intent labels and entropy derived from classified chat rounds. Authors listed alphabetically; one co-author is affiliated with the platform operator (Singapore Sapiens Technology). Operations-management framing (service consumption and outputs), no user-wellbeing outcomes, single platform and model. Date is the arXiv v1 submission (20 September 2026, cs.HC, CC BY). Verified at the arXiv abstract page by the curator; the arXiv beat read the PDF (affiliations, arm sizes, design).

Sources