Buy Outcomes, Not Technology

What changed in the 2026 guidance

The procurement environment has moved beyond general calls for responsible technology. NIST finalized two directly usable system and supplier resources in summer 2026. SP 800-18 Rev. 2 treats security, privacy, and cybersecurity supply-chain plans as connected system plans and emphasizes purpose, data flows, components, roles, controls, and lifecycle decisions. SP 1326 provides a quick-start structure for supplier due diligence. [3-4]

NIST also released an initial public draft of the TEVV-Athlon framework in August 2026. It is designed to structure test, evaluation, verification, and validation of real-world AI systems and is open for comment through October 6. The existing AI Risk Management Framework remains available, but NIST states that AI RMF 1.0 is being revised. [6-7]

Michigan is scaling capability while statewide rules remain incomplete

Michigan's ALPR proposals remain pending rather than enacted statewide law. HB 5492 and HB 5493 were introduced on January 29, 2026, and referred to the House Judiciary Committee. SB 1131 was introduced on July 29 and referred to the Senate Civil Rights, Judiciary, and Public Safety Committee. The bills take different approaches to governmental and private systems, retention, access, reporting, remedies, and permitted use. [9-11]

Contemporaneous reporting in April 2026 identified ALPR use across more than 125 Michigan cities and counties. Because statewide legislation remains pending, local policy and contract terms continue to do much of the practical governance work. [15]

The same implementation question is emerging in advanced air mobility. In August 2026, Michigan awarded $2 million for five pilots pairing companies with EGLE, DNR, MDOT, MSP, and MDARD for environmental monitoring, mapping and wildfire support, roadway intelligence, drone detection, and pesticide-drift investigation. The state's 2025 AAM report described more than $42 million in public and private investment and nine active test sites. On September 17, 2026, the state announced another $2 million for 26 mobility, defense, aerospace, and manufacturing companies and reported more than $5 million awarded to 69 firms across three rounds. [12-14]

These investments are not proof that any specific system should be purchased. They increase the need for repeatable supplier, system, contract, and outcome controls as pilots become operational programs.

Federal ALPR policy now adds another governance layer

H.R. 10366, the No Federal License-Plate Observation and Camera Keeping Act, was introduced in the U.S. House on September 14, 2026 and referred to the Committee on Transportation and Infrastructure. The proposal would condition apportioned surface-transportation funding on states adopting specified restrictions for automated license-plate-reader use. It remains pending legislation and does not create a current federal mandate for Michigan agencies. [16]

Implementation consideration: Agencies evaluating procurement resilience can examine whether retention, permitted-use, sharing and deletion rules can be changed without replacing the entire system if federal or state requirements later tighten.

Four conditions that must be met before purchase

A high numerical score cannot cure a failed gate. This is especially important for biometric, persistent-surveillance, integrated-data, and consequential AI systems, where one missing control can dominate the risk.

Eight steps from problem definition to exit

1. Define the problem and baseline. State the public-safety outcome, current performance, affected population or geography, and credible non-technology alternatives.
2. Set the risk tier. Consider data sensitivity, persistence, automation, rights impact, integration, scale, and dependence on a supplier.
3. Build the system plan. Document purpose, system boundary, components, data flows, roles, controls, and lifecycle using SP 800-18 Rev. 2 as a practical model.
4. Review evidence and supplier. Separate vendor claims from independent evidence and conduct proportionate supplier due diligence using SP 1326.
5. Test the real use case. Establish test data, operating conditions, success criteria, error analysis, human review, and comparison points; use TEVV concepts for AI-enabled functions.
6. Contract for performance and control. Resolve data rights, cybersecurity, audit, change notice, evaluation, continuity, portability, termination, and deletion before deployment.
7. Monitor deployment drift. Track model, software, data-source, integration, workflow, staffing, and use-policy changes after award.
8. Make renewal affirmative. Report outcomes, total cost, incidents, changes, risks, and alternatives on a fixed schedule, then continue, modify, narrow, expand, or exit.

NIJ's 2025 adoption guide reinforces the same basic order: problem analysis first, followed by technical, operational, and governance review. [1]

Match review intensity to risk

Risk tiering should affect process, not predetermine the decision. A Tier 3 system may be appropriate when authority, need, evidence, controls, and outcomes support it. A Tier 1 tool can still fail if the vendor, contract, or security posture is weak.

Use a scorecard to document judgment—not replace it

Review the supplier, not only the product

NIST SP 1326 organizes supplier due diligence around five components: foreign ownership, control, or influence; provenance; resilience; foundational cybersecurity practices; and supply-chain tiers. The guide is scoped to information and communications technology suppliers, which fits many connected sensors, cloud platforms, analytics products, and AI-enabled systems used in public safety. [4]

Create a living system plan, not a one-time form

NIST SP 800-18 Rev. 2 links security, privacy, and supply-chain planning. It calls for documentation of the system purpose, components, environments, data flows, controls, roles, and information lifecycle. It also emphasizes machine-readable formats and continuous risk information rather than static point-in-time paperwork. [3]

• System boundary and intended tasks, including integrations and external services.
• Data created, collected, inferred, received, shared, stored, exported, and deleted.
• Named owners for operations, security, privacy, legal compliance, records, evaluation, and contract management.
• Controls in place, planned controls, known gaps, exceptions, and target dates.
• Supplier and subcontractor dependencies, hosting, model providers, and critical support services.
• Version, change, incident, complaint, and audit history throughout the contract.

The federal Privacy Impact Assessment model is also useful procedurally even when it is not legally required of a local agency. It asks how identifiable information is collected, stored, protected, shared, and managed across the lifecycle. [8]

Control model and capability changes after award

AI features can appear in ALPR search, video analytics, weapons detection, identity matching, report generation, investigative search, prioritization, and decision support. The version demonstrated during procurement may not be the version operating months later. Contracts and system plans should distinguish routine maintenance from a material capability change.

NIST's TEVV-Athlon draft offers a structured way to think about real-world AI testing and evaluation. It is a draft, not a binding standard, but procurement teams can already use its core lesson: test the system in the actual task, context, and operating environment rather than relying only on a general benchmark. [7]

NIST Privacy Framework 1.1 should also be cited accurately. As of this update, the final 1.1 version is still listed as forthcoming; the initial public draft remains the available update. [5]

Use one lifecycle record across procurement, operation and exit

Procurement and exit should not become separate governance silos. For consequential systems, the same case record should follow the system from need definition through purchase, deployment, material changes, complaints or incidents, evaluation, renewal and termination.

• Nine questions stay constant: Purpose, Authority, Procurement, Data, Decision, Human Control, Evaluation, Accountability and Exit.
• Add standard lifecycle fields: solicitation and vendor details, contract controls, renewal dates, termination rights, data export/deletion requirements and a dated change timeline.
• Keep incident tracking analytical: maintain a verified internal research log, but publish patterns and recurring control failures rather than a running public blacklist of agencies or vendors.

That same method can then be applied across three research tracks without multiplying product lines: Surveillance & Sensor Governance; Automated Decisions; and, when a concrete project exists, Digital Infrastructure.

Minimum contract controls

Name who recommends, approves and monitors the system

Cross-functional review should not diffuse accountability. Each required decision needs a named owner, a written recommendation, and an identified approving official.

A 90-day implementation path for Michigan agencies

1. Adopt a two-page Public Safety Technology Impact Assessment before an RFP, pilot, grant commitment, or sole-source purchase.
2. Classify systems into routine, significant, and high-impact tiers and publish the triggers for each tier.
3. Create a system-plan template aligned to SP 800-18 Rev. 2 and a supplier checklist aligned to SP 1326.
4. Publish model clauses for data control, security, audit, model change, records, continuity, portability, deletion, and exit.
5. Require a time-limited pilot and defined comparison when evidence is weak or the use is high-impact and a pilot is feasible.
6. Maintain a public inventory of covered systems, purposes, vendors, contract dates, policies, and evaluation schedules, with lawful security redactions.
7. Make renewal an affirmative written decision explaining outcomes, total cost, incidents, changes, residual risk, alternatives, and exit readiness.

The pending ALPR bills provide a useful test case. Agencies can map the proposed statutory safeguards against their own policies and contracts now, identify which controls require statewide law, and implement the remaining controls through procurement before legislation is resolved.

Bottom line

The key question is not whether a product is innovative. It is whether the public agency can document lawful authority, a defined need, credible evidence, operational capacity, data and security control, meaningful human oversight where decisions are consequential, a resilient supplier, an evaluable contract, material changes after deployment, and a workable exit path. Procurement, change management and exit belong in one lifecycle record. That method can be reused across surveillance and sensor systems, automated administrative decisions and, when a concrete project warrants it, digital infrastructure.