California interview study links AI purchasing difficulties to data and workforce capacity

Work with the city manager, procurement lead, CIO and department sponsor to identify an actual service bottleneck before discussing products. Ask who understands the data, who evaluates vendor claims and how much internal support the municipality can sustain. A bounded engagement could compare two approaches on one administrative task and produce a costed operating plan. The value hypothesis is a more supportable purchasing decision; these interviews do not prove savings. Avoid assuming smaller jurisdictions lack controls merely because public disclosures are absent. Clarify what existing shared services can supply and what responsibility must remain with the local service owner.

Treat build-versus-buy as a maintainability and data-boundary decision. Map source systems, permissions, model access, interfaces and recurring usage costs before prototyping. Prerequisites include an approved test dataset, reproducible tasks and staff able to inspect outputs. Compare products under the same quality and security criteria, then test export and replacement. For generated code or action-taking agents, use isolated repositories, limited credentials and human approval for consequential changes. Proposed proof of value should demonstrate the workflow on representative data and expose operating costs and failure recovery. Neither in-house development enthusiasm nor a shared platform establishes a secure cloud, on-premises or hybrid fit.

The department head should own the service outcome, with procurement, IT, privacy and workforce leads jointly responsible for their controls. Establish a pilot charter, train internal champions, track assistance requests and document how roles change. Dependencies include data cleanup, time for staff participation and authority to resolve cross-department disputes. Proposed acceptance requires a named support owner, tested recovery and export, completed supplier disclosures and documented staff competence on agreed scenarios. Assess quality and net workload before expansion; these are suggested criteria rather than reported achievements. Risks include dependence on one enthusiastic employee, unbudgeted usage costs and a pilot that cannot be maintained after vendor onboarding ends.