Year: 2026
Type: dissertation
Abstract: Artificial intelligence (AI) and large language models (LLMs) are shaping how organizations work. Enterprise contracts and model-provider safeguards reduce risk, but companies must still trust providers to honor their data-processing commitments. Organizations can run self-hostable open-weight models on their own servers, keeping confidential data processing under their control. Tasks using only public data may use any approved model. Most model routers optimize quality, latency, or cost. I examined whether data sensitivity could instead determine model eligibility: operations involving public data may use any approved model, while confidential operations are restricted to self-hostable open-weight models. I built and evaluated a proof of concept for an enterprise AI application that separates public, organization-managed, and user-scoped information. Evaluation included a requirements-based implementation review, a 28-call demonstration, and a seven-week public-data pilot at Siili Solutions with 187 calls. The benchmark contained 75 runs across five models; 71 included all four judge dimensions. Of the 215 traced calls, 213 involved public data. Both archived internal-mode calls used eligible models. A requirements-based implementation review found validation on dedicated internal paths, while one mixed-mode path did not repeat the check when an internal tool was called. On complex tasks, DeepSeek V3.2 scored 90.8 and Gemini Flash 91.5. Three tested open-weight models were within 2.1 points of Gemini Flash; simple-task scores ranged from 93.3 to 95.2. The results suggest that sensitivity-based routing is feasible in the evaluated application. Although the tested open-weight models scored close to Gemini Flash, the benchmark does not show that open-weight models generally match proprietary models in quality. Stronger proprietary models were excluded because of project cost limits, and the evaluation covered neither confidential documents nor self-hosted inference.
Source: Aaltodoc (Aalto University)
Author: Nurmela, Elias
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7203728974