Procurement as AI policy: six clauses a state buyer can demand in an AI contract

A state buyer can write six obligations into a contract with a frontier-AI vendor: model access for independent evaluation, information sharing, funding for reliability research, state-of-the-art risk-management requirements, termination and exit provisions, and support for domestic or allied capability. That is the core of a new RAND framework released today by researcher Afek Shamir, which calls the approach "procurement as governance."

The framework is aimed at what RAND calls "AI middle powers": states with significant economic, diplomatic, security, or regulatory capabilities that nonetheless depend technologically on a small number of US- and China-headquartered frontier-AI firms. It is the report's own coinage and travels with a built-in admission. These states do not build frontier AI, so they must govern it through what they buy.

That governance begins with three risks the report says fall disproportionately on state buyers: vendor lock-in and the threat of service withdrawal, limited visibility into how frontier systems are built, and data-sovereignty and information-security exposure. A state agency running sensitive workloads through a foreign-hosted model, for example, can find its own data governance subordinate to the vendor's home-jurisdiction national-security law. The six obligations are the contractual counterweights.

In the framework, the six obligations are layered. Model access for independent evaluation means the state can run its own safety and capability tests rather than rely on the vendor's published results. Information sharing requires the vendor to disclose meaningful details about training data, evaluation, and incidents. Funding for reliability research commits the vendor to a research budget the state can direct. State-of-the-art risk-management requirements are a moving target: the contract should demand the vendor apply the strongest current methods, not a fixed 2024 checklist. Termination and exit provisions let the state walk away, and crucially, take its data and fine-tuned weights with it. The sixth obligation, support for domestic and allied capability, treats the contract itself as industrial policy, with a portion of spend building in-house evaluation, secure hosting, or fine-tuning talent inside the buying state or its partners.

RAND's framework splits the playbook into a 12-month track and a 1–5 year track. In the first year, RAND recommends a shared cross-government risk assessment, a choice of institutional vehicle (the report names the G7, the EU, NATO, and the International Network of AI Safety Institutes as candidates), and demand pooling so smaller middle powers can negotiate as a block. Over the longer term, the framework calls for selective investment in domestic and allied compute, secure hosting for sensitive workloads, multi-model optionality so no single vendor becomes load-bearing, and in-house fine-tuning talent. Procurement coordination, in RAND's view, should eventually extend into joint incident reporting and resilience exercises.

The same report is explicit about where the lever breaks. Government procurement is still a small share of frontier-AI revenue, so leading developers can refuse onerous terms and lose little. Contracts with foreign providers may be subordinate to the vendor's home-state national-security law, which the contract cannot bind. State buyers face persistent information asymmetries: even with disclosure obligations, they struggle to evaluate what the vendor discloses. And the strongest obligations, like termination and exit, are the hardest to use in practice.

RAND's central recommendation is therefore modest: AI middle powers should assess whether collective coordination on procurement requirements can strengthen their bargaining position. The framework is a tool, not a forecast. The next concrete milestone to watch is whether any G7 or EU working group, or the International Network of AI Safety Institutes, picks up the six obligations as a negotiating starting point, and which vendors agree to be first.