Publication

Public AI Agents: Who Runs Them When the Vendor Leaves?

In September 2026 Gartner predicted that by 2028, 70% of enterprises will abandon agentic AI built by vendor forward-deployed engineering, citing soaring costs and enterprises being unable to change the systems on their own. The Korean government is moving from owning AI toward using it: the budget for OnAI, AI for civil servants' work, rises from 18.7 billion won in 2026 to 54.2 billion won in the 2027 government proposal, and 83.2% of the 2027 amount is fees for private collaboration software. Current contract rules center on source code and documents; the model settings, prompts, knowledge base index and execution logs needed to operate an AI agent are not listed as items, and there are no rules on the scope or completion criteria for handing over operating knowledge. The Board of Audit and Inspection found 574 contractor staff stationed at the National Information Resources Service for maintenance, 1.9 times its 305 employees, and that 336 of 491 general AI products and services in the public sector were commercial products. The United States, the United Kingdom and the European Union address the problem through contract terms on knowledge transfer, data and model portability, rights to code and models, and exit plans. The brief proposes five tasks: widening the scope of AI project deliverables and making handover at contract end a mandatory term, putting knowledge transfer into the scope of work and price, setting an accountable person and termination criteria for each project, submitting switching costs for subscription projects with the budget, and setting targets for in-house experts at operating institutions.