Writing an RFP for Government AI Systems

AI acquisitions can fail when a solicitation describes a broad goal but leaves evaluation, data rights, or ongoing performance undefined. A useful RFP starts with the public service need and the workflow the system will support. It then specifies relevant user groups, operating conditions, success measures, prohibited uses, security and privacy needs, accessibility, required integrations, and how a human will review uncertain or harmful results. Vendors should be asked for evidence on representative tasks rather than generic accuracy claims. Agencies can define acceptance criteria, logging, incident reporting, model-change notice, monitoring, and an exit plan that preserves data and supports portability. Procurement staff should assess lifecycle costs, including integration, evaluation, support, retraining, and transition. Intellectual property and data-use terms need explicit treatment, including whether agency information can be retained or used to train other models. Requirements should be proportional to the system’s risk and public impact, and should comply with applicable acquisition law and agency policy. GAO’s 2026 review of federal AI acquisitions notes challenges with evaluating technical proposals and AI-related costs, and highlights lessons learned and contract terms as practical issues. The report is oversight analysis, not a complete solicitation template or binding rule. Agencies should involve procurement, program, privacy, security, legal, accessibility, and technical experts early. A well-scoped RFP makes vendor claims testable and sets expectations for monitoring after deployment; it cannot guarantee a suitable product without implementation oversight.