There is no special AI procurement authority. AI is bought like other professional services and IT, and the difficulties are the ordinary ones made sharper by genuine technical uncertainty.
The common paths
GSA Multiple Award Schedule. The default route for commercial IT services and products. Broad, well-understood, and the path most agencies reach for first.
Agency-wide IDIQs and GWACs. Pre-competed pools with task-order competition. Faster than open competition once you are on the vehicle, and effectively closed if you are not.
Blanket Purchase Agreements. Efficient for recurring needs against a known set of vendors.
Simplified acquisition. Below the simplified acquisition threshold, procedures are considerably lighter — a practical route for a discovery or assessment engagement that de-risks the larger buy.
Small business set-asides. Agencies carry small-business goals including socioeconomic categories. A Service-Disabled Veteran-Owned Small Business set-aside restricts competition to certified SDVOSBs, which materially changes the competitive field. Guardian Robotics is SDVOSB with UEI EU9ZEGK4C4U4 and CAGE 9QAQ2 — see our vendor profile.
Sole source, with justification, where only one vendor can meet the requirement. Scrutinized, and rarely the clean answer people hope for.
SBIR/STTR for genuine research and development rather than implementation of known techniques.
Thresholds and procedures change; verify current figures against the FAR and agency guidance rather than any summary.
Structuring the requirement
This is where AI acquisitions go wrong, in both directions.
Do not over-specify the solution. A statement of work mandating a particular model, framework, or architecture locks in a choice that will be outdated before award and prevents the vendor from using what actually works on your data.
Do specify the outcome measurably. "Classify incoming correspondence into the existing 14-category taxonomy with at least 90% accuracy, measured against 500 records labelled by agency staff and withheld from the vendor." That is evaluable, defensible, and vendor-neutral.
Sequence the buy. A small assessment task order — data review, feasibility, accuracy estimate on real records — before a large build. This is the single most effective risk reduction available, and it is cheap. It also produces the evidence needed to justify the larger action.
Require the evaluation set as a deliverable. It is government property and the most durable asset from the engagement. Without it the agency cannot verify performance or manage a successor vendor.
Address data rights explicitly. Who owns prompts, fine-tuned weights, evaluation sets, and derived artifacts. Ambiguity here creates real problems at recompete.
Put the authorization boundary in the requirement. If the solution must operate inside an existing ATO at a given impact level, say so — it eliminates non-viable offerors before evaluation rather than after award.
The most common federal AI failure is a large fixed-price build awarded before anyone confirmed the data could support the requirement. A $40,000 assessment preventing a failed $2,000,000 build is the best value in this entire process.
Evaluating proposals
Weight past performance on comparable data problems over general AI claims. Ask specifically:
- What accuracy is committed, and measured how, on whose data?
- What happens to the cases the system cannot handle?
- Who owns the code, the prompts, and the eval set?
- What does year-two sustainment cost?
- Which components are FedRAMP-authorized, at what impact level, and in what scope?
Frequently asked questions
Is there a specific federal contract vehicle for AI?
No single dedicated vehicle. AI services are bought through general IT and professional services vehicles, with GSA MAS being the most common. Some agencies have established AI-focused BPAs or IDIQs, so check the specific agency's landscape.
What does SDVOSB status change for an agency buyer?
It allows the agency to restrict competition to certified service-disabled veteran-owned small businesses, which supports statutory small-business goals and typically shortens the acquisition timeline relative to full and open competition. Verification is through SAM.gov registration and the relevant certification.
How should an agency handle AI performance uncertainty in a contract?
Sequence it. A small firm-fixed-price assessment establishes achievable accuracy on real agency data, and the results inform a properly-scoped build. Attempting to fix price and performance on unexamined data either produces a large risk premium or a vendor incentive to cut the parts you cannot see.
Can agencies buy AI as a subscription?
Yes, where the service is appropriately authorized and the subscription model fits appropriations rules. Verify the FedRAMP authorization covers the specific offering and impact level, and confirm that data handling terms meet agency requirements — commercial terms of service frequently do not.
What should an agency insist on owning?
The evaluation dataset, the prompts and configuration, documentation sufficient for another vendor to take over, and rights to any fine-tuned artifacts created with agency data. These determine whether you have bought a capability or a dependency.