AI Procurement Rules Are Turning Buying Committees Into Risk Committees

Public agencies entered the AI cycle hoping to buy productivity. They are discovering that they are also buying a new category of institutional risk. The result is a procurement process that looks less like a software selection exercise and more like a risk committee review. Buyers still ask whether the tool works. They now ask, with equal seriousness, who owns the data, who can explain the output, who is liable when the workflow fails, and whether the agency can audit the system after the contract is signed.

The specification is doing more work

The most important AI procurement decisions are being made before the vendor presentations begin. A requirement for human review changes the field. A requirement for model logs changes the field. A rule against training on agency data changes the field. These are not legal decorations attached to a technical purchase. They determine which products can compete and which forms of automation the public body is willing to accept.

That shift is healthy when it forces agencies to describe the public value they want rather than the novelty they are chasing. It becomes dangerous when risk language is vague enough to block useful tools without creating safer ones. The procurement team needs enough technical fluency to know the difference between a real control and a reassuring phrase.

What a better process would ask

A better process would ask vendors to map the workflow, not merely demonstrate the interface. It would require evidence of data boundaries, failure escalation, audit records, model-change notification, and user training. It would also state where automation is not acceptable. Those exclusions matter because a public agency earns trust partly by naming the decisions it will not delegate.

AI procurement will not be solved by buying slower. It will be solved by buying more deliberately. The agencies that build sharper specifications now will not only reduce risk. They will also avoid the more expensive failure: adopting tools that make the work look modern while leaving the public unable to understand how decisions were made.