Local governments have spent decades developing procurement systems designed to create competition, control costs and protect public dollars. AI introduces a technology category that can change faster than the contracts used to purchase it.

Pinellas County recognizes procurement as part of the work in its current AI Strategic Roadmap and Governance Framework initiative. That belongs in an AI roadmap because purchasing decisions can determine which models government uses, where public data goes and how easily an agency can change direction later.

Consider a typical multiyear technology purchase.

Government defines requirements, evaluates competing products, selects a vendor and negotiates a contract. The winning system may then operate for several years before the next procurement.

During a comparable period in AI, a vendor might replace its underlying model several times. Capabilities can improve substantially. Pricing can change. Features once available only through specialized products can become standard components of larger platforms. A vendor can change model providers without the government agency changing vendors at all.

Government could reach the second year of a contract using technology materially different from what procurement staff originally evaluated.

That requires a different purchasing discipline.

The first change should occur before an RFP is written. Government should define the operational result it wants rather than starting with an AI product.

If a permitting department wants to reduce the time employees spend checking applications for completeness, that becomes the procurement problem. Establish current processing time, accuracy requirements, security constraints and the improvement the department expects.

Vendors can then compete on their ability to produce that outcome.

This approach also gives government something concrete to measure after implementation. A system that performs impressive demonstrations but does not improve permit processing has failed the operational test.

Contracts need to anticipate change as well.

Data rights should remain clear regardless of which model sits behind the application. Government needs to know where its information goes, how it is stored, whether it can be used to train other systems and what happens to it when the contract ends.

Portability becomes increasingly valuable. Government should understand how difficult it would be to move its data, workflows and integrations to another provider or model. The cost of switching belongs in the original purchasing decision.

Performance requirements need to continue after launch. AI systems should be evaluated against defined measures for accuracy, reliability, security and operational results. A model update should not quietly change the quality of a public service.

Vendor-change requirements should establish when government must be informed about significant changes to underlying models, data practices, security architecture or other components that influenced the original evaluation.

These requirements make procurement part of AI governance.

They also create a different way to think about vendor lock-in. The concern extends beyond whether government can export its database. An agency can become dependent on proprietary prompts, workflows, integrations, evaluation methods and model-specific behavior that make changing providers expensive even when the underlying data remains portable.

Procurement officials will need greater technical support to evaluate those dependencies.

They will also need contracts with enough flexibility to respond when the market moves. A government should have mechanisms to evaluate significant technology changes, renegotiate where appropriate and exit when a product no longer meets its requirements.

The objective is not predicting which AI vendor or model will lead the market three years from now. Government is unlikely to know.

A better procurement system preserves options while protecting public data, performance and dollars.

That gives local government something particularly valuable in AI: the ability to change its mind.

A Practical Model for Civic AI Procurement

Procure outcomes. Define the operational problem, baseline performance and measurable improvement government expects before specifying technology.

Protect portability. Preserve control of public data and understand the cost of moving workflows, integrations and systems to another vendor or model.

Evaluate continuously. Establish ongoing measures for accuracy, reliability, security and operational performance rather than relying solely on pre-purchase evaluation.

Design for change. Require disclosure of significant technology changes and create contractual options to adapt, renegotiate or exit as capabilities, costs and risks change.

‍