Cities and counties are asked to do more each year with staffing that does not grow at the same rate, which makes automation genuinely attractive rather than merely fashionable. They are also among the organizations least able to absorb a failed technology purchase, because the budget cycle rarely allows a second attempt in the same year and the failure is a matter of public record.
That combination argues for assessing readiness first. Not as caution for its own sake, but because the assessment is the cheapest way to produce a defensible basis for whatever you do buy.
Where the value usually is
Local government has an unusually good fit with the things current AI is actually reliable at, because so much of the work is text.
Constituent contact is the obvious case. A city receives the same questions repeatedly, in varying phrasing, through several channels, and answering them consumes staff who have other duties. Improving how residents find an answer, and helping staff draft consistent responses, is well within reach and does not require anything experimental.
Internal knowledge search is the one departments ask for once they have seen it. Ordinances, council decisions, procedures and prior correspondence are scattered across systems, and staff answer from memory because locating the authoritative version takes longer than the answer is worth. That is a real cost, and it touches no personal information, which makes it a sensible first candidate.
Document processing, permit and licensing workflow, agenda and minutes preparation, and IT service management round out the list. A fuller treatment with the practical caveats is in AI use cases for local government.
Public records is the question a vendor will not raise
This is the part that differs most from the private sector, and in our experience it is the part least likely to have been discussed before a demonstration.
If a member of staff uses an AI tool to draft a response, the draft may be a public record. So, potentially, may the instruction they typed to produce it, and the conversation history the tool retained. Whether that is so depends on your state's law and on your own retention schedule, and the range of answers across the country is wide enough that no page can tell you which applies to you.
What an assessment establishes are the facts your records officer and counsel need in order to decide: which tools are in use, what they retain and for how long, whether that retention is under your control or the supplier's, whether the content can be produced on request, and whether it can be disposed of when your schedule says it should be. A tool that cannot produce records on demand, or cannot delete them on schedule, is a records management problem regardless of how useful it is.
Settle this before deployment. Retrofitting a retention position onto a tool already in daily use across four departments is considerably harder than choosing a tool that can meet it.
The systems are older than the ambition
Cities and counties tend to run a long tail of specialized applications, several of them a decade or more old, bought from vendors serving only this market, and holding exactly the information a useful AI application would need.
The assessment question per system is specific: is there a documented interface, a supported export, or neither. Ask it in writing, per system, and be suspicious of a general assurance. The most common avoidable cost in this work is discovering after selecting a platform that the permit system will only produce a formatted report.
Two other technical findings recur. Identity is often managed in more than one place, which means an AI service granted access in one of them is invisible in the others. And network and endpoint capacity in smaller jurisdictions is frequently sized for the current workload with no headroom, which is a scheduling constraint rather than a blocker but needs to be known in advance.
Security, sized for the actual threat
Local government is attacked persistently and often has a security team of one, or of nobody. The AI-specific questions sit on top of that rather than replacing it.
The additions worth examining are whether staff can send resident information to an external service without anything preventing it, whether the organization can see what has been sent, whether an approved tool's access has been scoped or simply granted, and what a supplier's terms permit them to do with your content. That last one deserves reading rather than asking about, because the answer usually lives in a linked policy rather than in the agreement itself.
The wider set is in AI cybersecurity risks to assess before deployment, and CISA's artificial intelligence resources are the reference point most jurisdictions are expected to be aware of.
Governance that a council will accept
Governance in a city or county has an audience that other organizations do not have: an elected body, and through it the public. That changes what the document has to do.
A workable position names who may approve a tool and who may decline one, states which categories of information may never go to an external service, establishes that AI output is a draft rather than a decision, and identifies the decisions that require a human judgment somebody is accountable for. It should be short enough that a council member will read it and specific enough that a department head can apply it.
Two additions matter locally. Say something about disclosure: whether the public will be told when AI has been used in preparing a response or a document. And be explicit about decisions affecting individuals, because eligibility, enforcement, permitting and benefits determinations are exactly where automated decision making causes harm and exactly where a resident will ask who decided. Our proportionate starting position is in an AI governance framework for public-sector organizations, and the NIST AI Risk Management Framework is the reference most often requested by a governing body.
Staff, and the conversation nobody wants to have first
Public sector staff hear "automation" and, reasonably, hear something about their jobs. Leaving that unaddressed does not avoid the conversation; it just means the conversation happens without you.
The organizations that adopt this well say plainly what is being automated and what is not, involve the affected teams in choosing candidates, and are honest that some roles change. They also invest in the thing that actually determines success: whether the people receiving AI output can tell a good answer from a plausible wrong one. In local government that competence usually exists, because staff know their own subject matter well, but it needs to be turned into an explicit checking step rather than assumed.
Departments do not move at the same speed
One finding is specific to the sector and worth planning around: readiness is rarely uniform across a jurisdiction. A finance department running a modern platform with clean records and a clerk's office running a system bought in 2011 will score very differently on the same dimensions, and an organization-wide average conceals both.
The practical consequence is that a city should usually assess centrally and sequence departmentally. Governance, data classification, access review and the acceptable-use position are organization-wide and should be settled once. Which department goes first is a separate decision, and the right answer is usually the one with accessible data and a willing team rather than the one with the loudest problem.
Elected leadership sometimes wants the opposite: visible work in the department residents complain about most. That is a legitimate priority and it is worth saying plainly what it costs, which is usually a longer first phase spent on integration.
Procurement, and why the assessment helps
Public purchasing needs a defensible basis, and "the demonstration was impressive" is not one. An assessment produces the document that makes the rest of the process straightforward: what was needed, what was considered, what was rejected and why, what risks were identified, and what was done about them.
It also tends to reduce what you buy. Once candidate use cases have been tested against data availability and integration, the shortlist is usually shorter and cheaper than the one you started with.
If a cooperative contract is your likely route, AI readiness consulting for TIPS members sets out how that works, and what TIPS cooperative purchasing is covers the basics.
Where to start
The AI Readiness Self-Assessment gives you a score across the seven dimensions in about eight minutes, with recommendations for your weakest areas. It is self-reported and says so.
To work through it with your own departments, the public-sector readiness checklist is the working document, and ten questions to ask is the version to take into a first management meeting.
Where the answers need checking rather than believing, an assessment of your environment examines the systems, records and policies themselves and produces a roadmap. Or simply tell us what your jurisdiction is weighing up and we will tell you what we would look at first.