How state and local government applies the NIST AI RMF
The NIST AI RMF is voluntary, has no certification, and no auditor issues a NIST AI RMF report, so a council, board, state agency or contract asking whether the NIST AI RMF for state and local government is in place is really asking for an internal program measured against the 72 subcategories. The organization owns its controls and must be able to show how it governs, maps, measures and manages each AI system. An implementation partner can prepare systems, inventories and evidence, but it cannot sign off on the result. Only an independent assessor or auditor can issue a third-party report, and only if the buyer or regulator specifically requires one.
The right answer depends on three factors: how many AI systems the entity runs, whether it builds them or procures them, and whether a state law applies. San Jose's published self-assessment shows the pattern. Its scores were highest for Map and lowest for Measure, and the city attributed its low Measure score partly to subcategories written for model development, since most of its AI systems are procured. A public body that buys AI should therefore check each subcategory against what it can actually control in a vendor contract, such as testing access, monitoring rights and documentation the vendor must supply. Confirm early which subcategories assume in-house development and record how the organization will meet the intent of each one.
State law adds a separate layer. Texas HB 149 applies to governmental entities as defined in the statute, excluding hospital districts and institutions of higher education. Public bodies in Texas should confirm with counsel whether they are covered, then check two specific duties: the disclosure in Section 552.051 when an AI system interacts with consumers, and the prohibitions on social scoring in Section 552.053 and on certain biometric identification in Section 552.054. Build an AI inventory that flags public-facing and biometric uses first, so the legal review starts from a documented list rather than memory.
Where the gaps appeared and what to fix first
Govern covers the policies, roles and culture that every other function depends on. San Jose already had a centralized AI governance owner in its IT department, an AI governance working group and a digital privacy advisory taskforce, but its self-assessment still scored Govern at an average of 2.3 on a scale from 1 (new, no action yet) to 4 (established and constantly improving). The gap was a formal citywide AI policy that sets out roles and responsibilities, executive sponsorship and AI principles, so that the working group does not depend on informal authority.
- Written citywide AI policy: A council-approved or executive-approved policy that names the AI principles the city commits to and the scope of systems it covers.
- Named roles and responsibilities: A RACI or equivalent matrix showing who approves new AI use, who reviews risk and who can stop a system.
- Executive sponsorship: A record that a senior official, such as the city manager or a deputy, owns AI risk decisions and reports on them.
- AI education across departments: Training records for department staff who select, buy or operate AI tools, not only for IT.
- AI inventory linked to policy: A list of every AI system in use, including procured tools, tied to the policy's approval steps.
Check whether your city can show the signed policy, the approval chain for a recent AI purchase and the training completion records for at least one non-IT department.
Map establishes what each AI system is for, who it affects and what could go wrong. San Jose scored Map higher than most functions, at an average of 2.8, which suggests context-setting was already reasonably developed. The practical lesson is that Map work must be kept current as departments add tools, because a context statement written at procurement can go stale once a system is used for a new purpose.
- Intended purpose and context: A documented statement of what each system does, which residents it affects and which city mission it serves.
- Impact and benefit analysis: Records of likely benefits and harms, including harms to residents who cannot opt out of city services.
- Procured versus built classification: Notes showing whether a system is vendor-supplied, configured or internally developed, since most of San Jose's systems are procured.
- Component and third-party mapping: A list of vendor models, data sources and integrations that feed each system.
- Operator and oversight roles: Documentation of which staff use each system and what human review applies to its output.
Ask whether each mapped system has a current purpose statement that a department head has reviewed in the past year.
Measure asks whether the city can test AI systems for trustworthiness and track them over time. San Jose's lowest average score was Measure at 1.7. The city attributed part of this to subcategories written for the technical development of AI models, while most of its systems are procured. Cities should read Measure subcategories as applying to the configuration, data and outcomes they control, and should document how they test vendor systems they cannot inspect.
- Robustness testing: Test results showing how a system performs on inputs it was not tuned for, including edge cases typical of city services.
- Drift monitoring: Periodic comparisons of output quality against a baseline set when the system went live.
- Stress testing: Records of how the system behaves under peak load or unusual demand, such as an emergency-response surge.
- Error documentation: A log of known errors, their frequency and the corrective steps taken.
- End-user feedback and appeals: A reporting channel for residents and staff, with records of how complaints were reviewed.
Confirm with each vendor which test results they will share, since procurement contracts often limit access to model internals.
Manage turns the results of Map and Measure into decisions: whether to deploy, how to respond to risk and when to retire a system. San Jose scored Manage at 2.3. Its review pointed to the need for evaluation procedures that end in action, including decommissioning. A city that cannot retire a system once its value falls, or cannot show how it responded to a logged error, will struggle to show residents that AI use stays under control.
- Go or no-go decisions: Recorded approvals before a system goes live, with the risks accepted and by whom.
- Response to high-priority risks: Documented mitigations, such as human review thresholds or rollback steps, for risks rated most serious.
- Decommissioning procedure: A step-by-step plan for turning off a system, handling its data and notifying affected departments.
- Incident communication: Records of how errors were reported to leadership and, where relevant, to affected residents.
- Vendor contingency plans: Exit or fallback arrangements for systems supplied by third parties.
Check whether the city has retired at least one AI system under a written procedure, or whether decommissioning exists only as a plan.
Which Texas rules apply to governmental entities using AI
Texas applies a specific set of AI rules to governmental entities through HB 149, the Texas Responsible Artificial Intelligence Governance Act, enacted by the 89th Legislature. The act defines a *governmental entity* as a state or local administrative unit exercising governmental functions, and it excludes hospital districts and institutions of higher education. A city, county, school-adjacent municipal body or state agency that performs governmental functions should assume the act reaches its AI use unless counsel confirms a specific exclusion. The first task is an inventory question: which systems the entity runs, buys or develops, and which of them touch residents or public decisions.
Section 552.051 is the disclosure rule most likely to reach a public-facing team. When a governmental agency makes available an AI system intended to interact with consumers, the agency must disclose, before or at the time of interaction, that the person is dealing with AI. The disclosure must be clear, conspicuous and in plain language, and it must avoid dark patterns. A chatbot on a permit portal or a 311 service is a typical place to check. Review the wording, placement and timing with the team that owns the interface, and keep a dated screenshot or configuration record as evidence of the disclosure.
Section 552.053 bars governmental entities from using AI for social scoring, and Section 552.054 limits developing or deploying AI that uniquely identifies individuals from biometric data in certain cases. Both sections turn on what a system does, not on how it was procured, so a vendor-supplied tool can still fall within them. Ask each vendor and business owner to describe the system's outputs and any biometric inputs in writing. Confirm with counsel how these sections apply to your specific systems before go-live, and record that review in your AI inventory alongside the NIST AI RMF Govern and Map outcomes.
What makes a government AI program larger or smaller
A state or local AI program grows with the number of systems it must govern, who builds them, and how many offices use them. The six drivers below set how much inventory, policy, testing and evidence the NIST AI RMF work will need, and each one can be controlled by deciding early what is in scope.
Number of AI systems in inventory
Every mapping, measurement and management activity starts from a list of systems. Before scoping, build an inventory that records each AI system, its owner, its purpose and whether it makes or supports decisions about residents. San Jose's review covered all 72 AI RMF subcategories for its systems, which shows how quickly the workload grows once the inventory is complete. Ask each department to report AI use, including tools bought as features in existing software, so the count reflects reality rather than the systems the IT office already knows about.
Share procured versus built
Most government AI is bought, not built, which changes where evidence comes from. San Jose noted that many of its AI systems are procured rather than developed internally, and attributed part of its low Measure score to subcategories written for model development. Sort each system into built in-house, configured from a vendor product or purchased as a service. Built systems need development and testing evidence; procured systems need vendor documentation, contract terms and your own monitoring records.
Departments and locations
Each department, field office or facility that runs its own AI tool adds separate owners, data sources and user groups. A county with a central IT office may still have a permitting team or a public health unit operating tools the central office never reviewed. Identify every department and location that uses AI, name an accountable owner in each, and confirm whether the central AI governance owner has authority to require their participation in reviews and inventory updates.
Consumer-facing systems triggering Texas disclosure
Texas Section 552.051 requires a governmental agency that makes an AI system available to interact with consumers to tell the consumer, before or at the time of interaction, that they are dealing with AI. Chatbots on service portals, automated phone assistants and AI-driven application intake forms are the usual examples. Flag each system that interacts directly with the public, record how and when the disclosure appears, and review wording in clear plain language. Consumer-facing systems also deserve priority in risk review and monitoring.
Third-party vendor contracts
Each vendor contract that supplies AI, data or hosting adds dependencies that the agency must monitor and can only partly control. Check whether contracts give access to model documentation, incident notices, performance information and data-handling terms, and whether they allow the agency to suspend or replace a system. Count contracts by the systems they support, and check renewal dates so reviews are scheduled before terms roll over. The AI RMF's third-party subcategories give the checklist for what these contracts should cover.
Evidence maturity
A program with clear policy, named owners and repeatable testing is far easier to assess than one relying on memory and email threads. San Jose's self-assessment scored its practices on a scale from new, with no action yet, to established and constantly improving, and its average scores ranged from 1.7 for Measure to 2.8 for Map. Score each subcategory honestly against that scale before choosing a target. Lower maturity means more time spent creating records, so the gap between current and target scores is the best predictor of effort.
How a city or agency works through the framework
The sequence moves from naming who owns AI governance to publishing results that leadership and the public can review. Each phase produces a document or record that the next phase depends on, so the agency can show how its AI oversight was built and where it still has gaps.
Name the governance owner and working group
Leadership designates one accountable owner for AI governance, usually within the IT or digital services function, and charges a cross-department working group with reviewing AI use. San Jose's review started from this kind of structure: a centralized owner, a working group and a privacy advisory taskforce already in place. The step is complete when the owner, the group's members and their decision rights are written down and approved by the executive sponsor. Set the timing with leadership, since the group's first meetings depend on budget and staffing.
Build the AI inventory
Staff from each department list every AI system in use, including systems procured from vendors and tools embedded in software the agency already runs. For each entry, record the owner, the purpose, the data it touches, the people affected and whether it interacts with residents. Procured systems matter here because San Jose noted that most of its AI is bought rather than built. The output is a maintained inventory that the working group reviews, and it is complete when every department has confirmed its list in writing.
Score current maturity for each subcategory
The working group checks the agency's practices against the 72 AI RMF subcategories and assigns each one a current maturity score on a defined scale. San Jose used a scale from 1, meaning new with no action yet, to 4, meaning established and constantly improving. Agencies should agree on scoring rules before scoring, record the evidence behind each score, and have a second reviewer check a sample. The output is a scored baseline with category averages, which shows where the agency is weakest. Confirm the scale and evidence rules with your oversight body before the scores are used for decisions.
Prioritize gaps by public impact
The working group ranks low-scoring subcategories by the risk each AI system poses to residents, staff and public services, using the inventory and the Map results. Measure-related gaps often rank high for agencies that procure AI, because they depend on evaluation methods that the vendor or the agency must supply. San Jose's low Measure score partly reflected subcategories written for model developers, so agencies should tailor expectations to what they can actually control. The output is a ranked gap list with an owner for each item, and the step closes when leadership accepts the ranking.
Adopt the citywide AI policy and procedures
Using the ranked gap list, the agency drafts a formal AI policy that sets roles and responsibilities, executive sponsorship and AI principles, then writes the procedures that put it into practice. San Jose identified these as its first needs after its review, along with AI education across departments and formal end-user feedback channels. Legal counsel should check the policy against Texas rules where they apply, including the disclosure duty for consumer-facing AI and the prohibitions on social scoring and certain biometric uses. The output is an approved policy and a set of procedures, and the step is complete when departments have received training on both.
Run evaluations and set up monitoring
The agency runs the evaluation procedures its policy requires for each system in the inventory. San Jose's review named robustness testing, drift monitoring, stress testing, error documentation and decommissioning procedures as areas to formalize. Evaluations should be repeated after changes and recorded so that the next maturity score rests on evidence. The output is an evaluation file for each system and a monitoring schedule owned by a named person, and the step closes when the first full cycle of records is filed with the working group.
Publish or report results
The working group compiles the rescored baseline, the policy status, the inventory and the open gap list into a report for the executive sponsor and, where the agency chooses, for the public. Reporting should state the scoring method, the evidence behind the scores and the gaps still open, so that readers can judge the result. The report is not an audit or certification, and the agency should say so, because NIST does not validate or endorse any organization's use of the framework. The output is an approved report with a date and an owner for the next review, which feeds the next cycle of the sequence.
What a government team needs before scoring its AI governance
Before a city, county or state agency scores itself against the 72 AI RMF subcategories, it needs a few working foundations in place. The six items below are the ones San Jose's review pointed toward, and each one determines whether the scoring produces evidence or only opinion.
A complete AI system inventory
List every system that makes or supports decisions, including tools bought from vendors and AI features switched on inside existing software. San Jose's review assessed governance across all AI in use, and most of its systems were procured rather than built. To check the inventory, ask each department to name the vendor tools it uses, then compare that list against procurement records and software subscriptions. Anything missing from the inventory cannot be scored.
A named executive sponsor
Assign one executive who owns the outcome of the assessment and can make decisions across departments. The AI RMF's governance categories expect accountability to sit at a senior level, and San Jose's post-review findings called for executive sponsorship in a formal citywide policy. Check that the sponsor has written authority over AI purchases and that the name appears in the policy, not only in a meeting invitation.
Documented roles and reporting lines
Write down who owns each AI system, who reviews its risks and who can stop it. San Jose already had a centralized AI governance owner in its IT department and a working group, yet its post-review findings still called for formal roles and responsibilities. To check this, pick three systems and ask a staff member outside IT to name the accountable owner. If the answer varies, the roles are not documented.
A privacy or legal contact
Name a person who can answer privacy and legal questions about each AI system, such as a privacy officer, records officer or city attorney. Texas rules for governmental entities touch on disclosure and prohibited uses, and public records obligations shape what documentation can be shared. Confirm that this contact sees the inventory before scoring begins, so that privacy and legal risks are assessed alongside technical ones.
Procurement language requiring vendor documentation
Add clauses to purchasing templates that require vendors to supply model documentation, testing results, known limitations and incident notice. Most public agencies depend on procured AI, so the Measure evidence for those systems usually sits with the vendor. Ask procurement staff whether current contracts give the agency a right to receive this material, and whether renewal dates allow new terms to be added.
A list of applicable state laws
Compile every state law that governs your agency's AI use, starting with the Texas Responsible Artificial Intelligence Governance Act (HB 149) if the agency is in Texas. HB 149 defines a governmental entity as a state or local administrative unit exercising governmental functions and excludes hospital districts and institutions of higher education, so confirm whether your agency falls inside that definition. Record which obligations apply to each system so the scoring reflects legal requirements, not only the framework.
Self-assessment, in-house program or outside help
This table compares three ways a city, county or state agency can approach the NIST AI RMF: an internal self-assessment modeled on San Jose's published review, an in-house program supported by an outside readiness partner, and an independent assessor who reviews evidence. Read each row to see who issues output, when the option fits and what it costs in independence and credibility.
| Option | Who issues the output | When it fits | Main trade-off |
|---|---|---|---|
| Internal self-assessment modeled on San Jose | Only the agency, through its own governance documents and scoring | Early stage agencies that need a baseline across the 72 subcategories and have an internal owner for AI governance | Low cost in outside fees, but scores reflect the agency's own judgment and carry no independent weight |
| In-house program with an outside readiness partner | The agency still owns and signs all policies, scores and decisions; the partner prepares systems and evidence | Agencies that must build a citywide AI policy, evaluation procedures and feedback channels and lack staff time or method experience | Faster progress on documentation and testing, but the agency must keep the program running after the partner leaves |
| Independent third-party assessor for evidence review | Only the independent assessor, in its own report or findings | Procurements, oversight bodies or public commitments that ask for outside validation of evidence | Adds independent credibility, but the NIST AI RMF has no certification, so the assessor reviews evidence rather than issuing a certificate |
| Who maintains the program after the assessment | The agency, in every option | Any option, since AI systems change and the framework is revised | Ask any outside party to confirm in writing which artifacts the agency keeps and who updates them |
| Scoring method | Set by the agency unless the assessor specifies its own method | Self-assessment and readiness work usually use the 1 to 4 scale from San Jose's review | Confirm the scale and scoring rules before starting so results can be compared across years |
| Metric burden | Set by the agency, with room to prioritize | Agencies that find many metrics overlap, as San Jose reported for MEASURE 1.1 and MEASURE 2.2 | Trimming metrics saves effort, but any reduction should be documented and justified to reviewers |
| Procured AI systems | Depends on contract terms and evidence the vendor provides | Most city AI is procured, so the agency relies on vendor evidence for model-development subcategories | Ask vendors in writing for test results, change logs and documentation before relying on them |
Why the assessment stalls and how to avoid it
San Jose's feedback to NIST shows where a self-assessment against the AI RMF loses momentum: the sheer number of subcategories, overlapping metrics and AI systems the city buys rather than builds. The cards below cover what caused those problems and the practical steps that keep an assessment moving.
Checking 72 subcategories by hand
San Jose scored its governance against all 72 AI RMF subcategories and reported the exercise as tedious and time-consuming. A city, county or state agency with a smaller team faces the same workload in full. Before starting, decide which subcategories apply to the systems in scope and record the reason for each exclusion, so reviewers can see that the omissions were deliberate rather than overlooked.
Scope to the system inventory
The framework does not require every subcategory to be scored against every system. Start from a current inventory of AI systems, classify each by impact and by who uses its output, and assess the high-impact systems in depth. Agencies that cannot name every AI use in their own departments should fix that gap first, because an assessment built on an incomplete inventory will not hold up under review.
Redundant measurement metrics
San Jose said many metrics felt redundant and cited MEASURE 1.1 and MEASURE 2.2 as examples. Overlapping requirements push assessors to answer the same question twice. Read each metric against the evidence it needs, and merge metrics that draw on the same test results rather than writing separate answers for each.
Procured AI and the Measure score
San Jose attributed its low Measure score partly to subcategories written for teams that develop AI models, while most of its systems are procured. For a procured system, the agency must obtain evidence from the vendor, such as test documentation, performance data and incident history, and write its own monitoring and feedback procedures. Confirm in each contract which vendor test results the agency can receive and how often.
One shared evidence register
Many subcategories ask for the same artifacts, such as a risk register, a system description or a monitoring report. Map each artifact once in a shared evidence register that lists the document, its owner, its review date and every subcategory it supports. Reviewers then trace one document to several answers instead of receiving duplicate narratives for each subcategory.
Playbook is dated and revision is pending
The Playbook's audit log shows its last content update in August 2023, and NIST's AIRC says it will be updated after the AI RMF itself is revised. Treat the Playbook suggestions as dated input, not current guidance. Record which version of the Playbook your assessment used, and check NIST's pages for the revised framework before you finalize scores or policy language.
Questions government teams ask about the NIST AI RMF
Is the NIST AI RMF required for cities, counties or state agencies?
No. The NIST AI RMF is a voluntary framework, and NIST's FAQ says neither private nor public sector organizations are required to use it. A requirement can still come from elsewhere: a state law, a grant condition, a procurement clause or a city council policy that names the framework. Before you start, check whether any of those instruments applies to your entity and what exact version or profile it cites. Your legal counsel should confirm which obligations bind your agency, since a voluntary framework can become a contractual expectation.
Is there an NIST AI RMF certificate my agency can earn?
No. NIST does not validate or endorse any organization's use of the framework, and it runs no certification, accreditation or conformity assessment program for it. Any vendor or consultant who offers an AI RMF certificate is not offering a NIST credential. What you can produce is a self-assessment or an internal assessment that documents how your governance maps to the 72 subcategories, with evidence behind each. If a buyer or regulator asks for independent assurance, ask them to define what they will accept, because that is a different engagement from the framework itself.
How much effort does an AI RMF self-assessment take, and what drives it?
Effort depends mostly on how many AI systems you run, how much of your AI is procured rather than built, and how much governance already exists. San Jose's published review scored its governance against all 72 subcategories and reported that checking practices against the framework was tedious and time-consuming, partly because many metrics felt redundant. Drivers to plan for include the number of systems in the inventory, the number of departments involved, whether you have a documented AI policy, and how much evidence must be collected or reconstructed. Ask any outside party for a written scope that lists the systems, departments and subcategories covered, along with deliverables and dates. No effort figure is given here because none has been supplied.
Can our own staff run an AI RMF assessment without a consultant?
Yes, a government team can run its own self-assessment, and San Jose's example was built with its own IT governance owner, working group and privacy taskforce. The main requirement is clear ownership: someone must hold the inventory, a working group must review the Map and Measure evidence, and executive sponsorship must back decisions. Outside help is most useful for gap analysis against the Govern subcategories, drafting a citywide AI policy, or testing evaluation procedures that staff have not run before. Be specific about which tasks you keep in-house and which you buy, and confirm the split in writing before work begins.
Which Texas rules apply to a governmental entity that uses AI?
Texas HB 149, the Texas Responsible Artificial Intelligence Governance Act, applies to governmental entities, defined as state or local administrative units exercising governmental functions, excluding hospital districts and institutions of higher education. Three provisions matter most. Section 552.051 requires an agency that makes an AI system available to consumers to disclose, before or at the time of interaction, that the consumer is interacting with AI, in clear and conspicuous plain language without dark patterns. Section 552.053 bars governmental entities from using AI for social scoring. Section 552.054 restricts developing or deploying AI to uniquely identify individuals using biometric data in certain cases. Have counsel confirm how each applies to your specific systems.
Does the pending NIST revision change what our agency should do now?
Not the core work. NIST says AI RMF 1.0 is being revised as part of the White House AI Action Plan and has not yet published a revised version. The July 2025 Action Plan recommends removing references to misinformation, diversity, equity and inclusion, and climate change. NIST expects a formal review no later than 2028. Keep your inventory, risk documentation and evaluation records current regardless, since they serve any version. Before you rewrite policy, confirm which version your state or local rules cite, and watch the NIST AI RMF page for the revision. Organizations that rely on Playbook suggestions should note that NIST says the Playbook will be updated after the framework is revised.
Who owns the controls, who prepares evidence and who assesses
The government entity owns its AI policies, risk decisions and AI systems, and that ownership does not move to anyone else when outside help is brought in. The AI RMF is voluntary, and NIST does not validate or endorse any individual organization's use of it, so a city, county or state agency decides for itself which subcategories to adopt, what risk tolerance to set, and who signs off on a go or no-go decision for each system. Those decisions belong in the entity's own governance records, such as the formal citywide AI policy and executive sponsorship that San Jose identified as gaps after its review. Before engaging anyone, the agency should confirm which roles sit with its own staff, including the AI governance owner, the working group and the privacy function.
An implementation partner such as DeployOpen can help prepare the systems, documentation and evidence that support those decisions, working inside the entity's own environment under a written scope that names the subcategories, systems and deliverables covered. Typical work includes building an AI inventory, drafting the Map documentation for each system, setting up the logs and monitoring that Measure relies on, and organizing evaluation records. Because most AI systems in local government are procured rather than built, the agency should ask whether the partner will also review vendor documentation, contract terms and third-party monitoring, since those are where much of the evidence lives. The partner prepares material; the agency approves it and keeps it.
Only an independent assessor or auditor issues a report or attestation about the program. An internal self-assessment, such as the scoring San Jose published against all 72 subcategories, is useful for planning and transparency, but it is not an external opinion, and no AI RMF certificate exists. When a procurement officer, oversight body or state office asks for independent assurance, the agency should first confirm what that party will accept, then define the assessment scope in writing so the assessor tests the same systems and evidence the entity relies on.
How to choose a partner for the work
This sequence lets a city, county or state agency define what needs assessing, hire a partner against a written scope, and keep the evidence and decisions in its own hands. Each step ends with an output the agency can check before the next one begins.
Build the AI inventory first
The agency lists every AI system in use or in procurement, including tools bought from vendors and features switched on inside existing software. Department owners, the IT governance owner and legal or privacy staff each confirm the entries for their area, noting the purpose, the data involved and whether a resident interacts with the system. The output is a dated inventory that defines the scope of any assessment. Without it, neither the agency nor a partner can say which systems the AI RMF review covers.
Decide which subcategories apply
The agency's governance owner works through the 72 AI RMF subcategories against the inventory and marks which ones apply to which systems, and which do not. San Jose's published self-assessment is a useful reference for how a city scored each subcategory, but the agency should make its own choices. The output is a written list of in-scope subcategories with a short reason for each exclusion, so an outside party can see what was left out and why.
Request a written scope and deliverables
The agency sends the inventory and the subcategory list to each candidate and asks for a written scope before any work begins. The scope should name the systems covered, the subcategories assessed, the documents and evidence to be produced, the people who will be interviewed, and who approves each deliverable. DeployOpen works on fixed-scope engagements with scope, deliverables and timelines agreed in writing before work starts. The output is a signed scope document that the agency can compare across bidders.
Confirm public-sector and state law experience
The agency asks each candidate to describe prior work with state or local government AI governance, and to explain how it would apply the Texas rules on AI disclosure, social scoring and biometric identification where they apply to the agency. Ask for named deliverables from comparable work, described in general terms, and check how the partner handles records retention and public-records obligations. The output is a short written response that the agency's legal counsel can review against its own obligations.
Confirm the partner issues no certificates
The agency asks each candidate to confirm in writing that it will not issue any certificate, attestation or conformity finding for the AI RMF work. NIST does not certify or validate AI RMF use, so any claim of certification should be treated as a red flag. The partner should describe its role as preparing systems, documentation and evidence, with the agency owning its governance decisions and any independent review coming from a separate party. The output is a signed statement of role that sits in the contract file.
Agree the handover of documents and evidence
The contract states which documents, policies, metrics, test records and decision logs the partner produces, the format in which they are delivered, and who holds the master copies. The agency should keep copies in its own repositories and confirm it can read and edit every file without the partner. Ask how changes made after handover will be recorded. The output is a handover checklist signed by the agency's governance owner and the partner's lead before the engagement closes.
Review the completed evidence against scope
Before final payment or sign-off, the agency checks each deliverable against the written scope and the subcategory list. Staff confirm that the evidence matches the systems in the inventory, that open items are recorded with owners, and that the self-assessment reflects the agency's own judgment rather than the partner's. Where a review by an outside party is planned, the agency gives that reviewer the same scope and evidence set. The output is a closing memo that records what was delivered, what remains open and who owns each item.
How DeployOpen prepares government AI governance evidence
DeployOpen is a Chainscore Labs company that plans, implements, documents and supports self-hosted open-source platforms inside the client's own infrastructure. For state and local government, we prepare the systems, records and evidence that support an agency's own NIST AI RMF self-assessment, while the agency keeps ownership of its governance decisions and the independent review it chooses to commission.
How we scope
- We agree a written scope covering which AI systems, departments and AI RMF categories the agency wants to assess.
- We confirm which subcategories the agency owns, which are procured from vendors, and which need vendor input.
- We record the scope, deliverables and timelines in writing before any implementation work begins, since engagements are fixed-scope.
How we build
- We set up a self-hosted AI system inventory and governance record that lists each system's purpose, owner and procurement source.
- We build evidence collection for evaluation procedures, such as robustness testing, drift monitoring and error documentation, where the agency wants them.
- We integrate feedback and incident logging so end-user reports and decommissioning decisions leave a traceable record.
How we hand over
- We deliver documentation mapped to the agreed AI RMF subcategories, with the evidence location and owner for each item.
- We train the staff who will maintain the records, so the agency can keep its self-assessment current after handover.
- We state in the handover what remains the agency's responsibility, including policy approval, executive sponsorship and any external assessment.
How we support
- Ongoing support, upgrades and continuous testing are available as a retainer, scoped in writing.
- We can update evidence when the revised AI RMF is published, since NIST has said the Playbook will be updated after that revision.
- Questions about Texas disclosure, social scoring or biometric rules can be scoped with the agency's legal counsel, who sets the compliance position.
