Most AI demos go well. The presenter types a question, an answer appears in seconds, and a draft shows up that looks close to something your office would write. Someone asks whether it would work on the city's own documents, and the answer is yes. Then the meeting ends, and the part the demo skipped is still waiting for you: how to buy this tool in a way that satisfies your procurement rules, your records obligations and your auditors, and what you will say when a council member asks why the city chose it.
Buying AI software is still buying software. Most of what your purchasing staff already know still applies: scope the need, compete it where required, evaluate fairly, negotiate a contract, manage the vendor. A few things about AI tools change the work at each step, though. The product can behave differently next quarter than it did in the demo. Its output can read smoothly and still be wrong. The vendor may want your data for purposes beyond serving your city. And quality is hard to judge until you test the tool on your own material.
This checklist covers the steps where those differences matter, in the order you will reach them.
Procurement thresholds, bidding rules, cooperative purchasing authority, public records law and data protection requirements differ by state, and often by charter or local ordinance. Nothing below is legal advice. Treat each item as a question to bring to your city attorney, your purchasing lead and your state clerk or procurement association.
Why AI purchases need their own checklist
A permitting platform does roughly the same thing on day 400 that it did on day one, unless you approve an upgrade. AI tools differ in a few specific ways, and each one shows up later in this checklist.
- The product changes underneath you. Vendors update the models their tools run on. An update can improve results, make them worse, or change the tone of drafts without any change to the screens your staff see.
- Quality depends on your data. A tool that does well on a vendor's sample documents may struggle with your scanned minutes from fifteen years ago or your department's shorthand.
- Errors are fluent. A wrong date or a misattributed council action reads just as confidently as a correct one.
- Data moves in less obvious ways. Staff type prompts, upload documents and receive outputs, and the system may keep logs of all three.
- Records questions follow the data. Prompts, drafts and outputs may be public records under your state's law.
- Accountability has to be assigned. When a staff report contains an error, the city is responsible, not the software.
None of this makes AI purchases riskier than every other purchase. It means the usual steps need a few extra questions, and those questions are easier to ask before you sign.
Define the problem before you meet a vendor
The most common mistake with AI procurement is starting from the product. "We should get an AI tool for agendas" is a purchase looking for a reason. A problem statement describes the work.
Compare these two starting points:
- "We need AI for staff reports."
- "Department staff spend a large part of each agenda cycle searching old packets and minutes for prior council actions on a topic, and those searches often miss relevant history, which leads to follow-up questions at the meeting."
The second version tells you what to measure, what data the tool will need, who the users are and what success looks like. It also leaves room for an answer that is not AI. Sometimes the fix is a better records index, a staff report template, or a submission deadline departments can actually meet.
Questions for the problem statement
Write the answers down. One or two pages is enough.
- What task are we trying to improve, and who does it today?
- How long does it take now? Measure a sample of real cases rather than estimating.
- What goes wrong today, and how do we find out when it does?
- What would "better" look like in terms we can check: time, completeness, fewer corrections, fewer follow-up questions?
- What records would a tool need to read, and what data must never go into it?
- Who will review every output before it is used, and do they have time?
- What happens if the tool is unavailable on a deadline day?
Set the risk level early
Not every AI use carries the same stakes. A tool that helps staff search past minutes for internal research sits at a different level than one that drafts language residents will rely on, and both differ from anything that affects eligibility, enforcement, permits or benefits. Decide early where your use case sits, because it shapes how much testing, review and contract protection you need. Your AI use policy, if you have one, may already define these tiers. The NIST AI Risk Management Framework is a useful general reference, even though it is voluntary and not written for city procurement specifically.
Choose a procurement path that fits your rules
With the problem defined, work out with your purchasing officer which path applies. The options below exist in many places, but whether each is available to you depends on your state and local rules.
- Small purchase. Below a set threshold, many jurisdictions allow simpler quoting. Watch for subscription costs that add up across years or departments and cross a threshold you did not expect.
- Request for information (RFI). Useful when you are not sure what to ask for. You learn what vendors offer and refine requirements before a formal solicitation.
- Request for proposals (RFP). Lets you evaluate on criteria beyond cost, which suits tools where accuracy and data handling matter as much as price.
- Cooperative or piggyback contracts. Buying through a contract another agency has already competed can save time. Confirm your authority to use it, and read its data and AI terms closely. Those terms were negotiated for someone else's needs, and you may want an addendum.
- Sole source. Requires a justification under most rules. An impressive demo is unlikely to meet the standard.
Look for AI that is already in the building
Find out what AI tools staff already use. Free accounts, browser extensions and AI features added to existing software can all bring terms the city never negotiated. An informal inventory clarifies the problem and may show you are already paying for capabilities nobody knew about.
Write requirements as outcomes and tests
AI products change fast enough that a feature list can be outdated before proposals arrive. Where your rules allow, write requirements as outcomes and say how you will test them. "The tool must show the source document, page and meeting for factual statements in a draft, and evaluators will verify a sample of those citations" is easier to evaluate fairly than technical terms copied from brochures. If a pilot is part of the evaluation, say so in the solicitation, along with how pilot results will be scored.
Design a pilot that produces evidence, not impressions
A pilot is where you learn whether the tool works on your material, with your staff, under your deadlines. It is also where many AI purchases drift: a free trial starts, a few people like it, and months later the city is paying for something nobody formally evaluated. A good pilot is bounded, documented and ends with a decision.
Put terms in place first
A free pilot is still a data relationship. Before real city records go into a vendor's system, you need written terms covering confidentiality, data use, retention and deletion when the pilot ends. Ask your city attorney whether a pilot agreement is needed and whether a free pilot raises procurement or gift questions under your local rules. Make sure the pilot does not convert automatically into a paid subscription.
Set success criteria and a baseline
Write down what the tool must do to pass, using the measures from your problem statement. For example:
- Factual statements in drafts trace to a source document that actually says what the draft claims.
- Reviewers spend less time correcting drafts than they would spend writing from scratch.
- Searches on test topics find prior actions that staff confirm are correct and complete.
- No confidential test material appears where it should not.
Then measure how the work goes today. Time a sample of current tasks and note how many rounds of corrections a typical staff report goes through. Keep the method simple enough to repeat during the pilot. Criteria set in advance protect you from judging the tool on how impressive it felt.
Build a test set that includes hard cases
Use real work from your own records, including the cases most likely to cause trouble:
- Old scanned documents and poorly scanned or handwritten minutes
- Items that were amended, tabled, continued or reconsidered
- Topics where council changed direction over several meetings
- Similar names, addresses or project titles that are easy to confuse
- Dollar amounts, ordinance numbers and dates
- Questions where the correct answer is "the records do not say"
If you compare vendors, use the same test set for each. If a vendor asks to choose its own examples, politely decline.
Log errors and end with a written decision
Pick reviewers who know the subject matter, including at least one skeptic. Give them a simple error log: date, task, what was wrong, how serious it was and how they caught it. That log is the most useful document the pilot produces, because it shows where the tool fails and how much review time it really requires.
At the end, compare results to your criteria and write a short memo: proceed, proceed with conditions, extend with specific open questions, or stop. The memo becomes part of your procurement record.
Data terms to settle before signing
Most of the AI-specific risk in a contract sits in how data is handled. Before you negotiate, map the data in plain terms with IT and the clerk's office.
Map what goes in, what comes out and what stays
- Inputs. Documents the tool reads (packets, minutes, ordinances, resolutions, staff reports), plus anything staff type or upload.
- Outputs. Drafts, summaries, search results and flags.
- What the system keeps. Logs of prompts and outputs, copies of your documents and searchable indexes built from them. Ask the vendor to describe each in writing, with how long it is kept.
- Who can see it. Vendor staff, subcontractors and third-party services the product relies on (often called subprocessors).
- Where it is stored. Where data is stored and processed, and whether that can change.
- How it is separated. How your data is kept apart from other customers' data, so one customer's records never appear in another's results.
Questions to resolve in writing
- Will the vendor, or any provider it uses, train or improve models with city data? If the answer is "no, unless you opt in," get that in the contract, not only on a policy page.
- Can the city export all of its data, including outputs and logs, in a format other systems can read?
- How will the vendor help with a public records request covering prompts, drafts or outputs?
- Can the city place records on legal hold inside the system?
- Can retention settings match your retention schedule?
- How is data deleted at contract end, and will the vendor confirm deletion in writing?
Data that needs extra care
Personnel files, records exempt from disclosure, health information and law enforcement data may carry state or federal requirements beyond your general public records law. Decide with your city attorney which categories may go into the tool at all, and tell staff before the first upload.
Evaluation criteria and a scoring table
Write your criteria and weights before you open proposals, and give them to evaluators in writing. Weights are a local decision. A city using AI for internal research may weight accuracy and data handling most heavily. A city drafting public-facing documents may put more weight on source traceability and the review workflow.
| Criterion | What good looks like | How to verify |
|---|---|---|
| Accuracy on your records | Correct results on your test set, including hard cases | Pilot error log and reviewer scores |
| Source traceability | Statements link to the specific document, page and meeting | Check a sample of citations against the originals |
| Handling of gaps | The tool says when the records do not answer a question | Include unanswerable questions in the test set |
| Human review workflow | Outputs are clearly drafts and easy to correct | Observe reviewers during the pilot |
| Data use and isolation | No training on city data without consent; data separated from other customers | Written answers and contract language |
| Security | Clear access control, encryption and incident response practices | Security questionnaire reviewed by IT |
| Records support | Retention, export, legal holds and records request support | Test an export and a sample search |
| Audit trail | Logs show who did what and when | Request a sample log export |
| Accessibility | Usable by staff who rely on assistive technology | Documentation plus hands-on testing |
| Fit with current systems | Works with your agenda, records and document tools | IT review, confirmed in writing |
| Exit and portability | Data returned in usable form, with deletion confirmed | Contract terms and a test export |
| Full cost of ownership | All costs over the term, including training and staff review time | Cost worksheet built by purchasing |
A few practices make scoring hold up:
- Score evidence, not demos. A demo shows ideal conditions. The pilot and written answers show what the tool does for you.
- Get claims in writing. If a vendor describes a capability on a call, ask them to confirm it in writing. Features change, and verbal descriptions are easy to misremember.
- Count review time as a cost. If staff must check every citation line by line, that time belongs in the comparison.
- Call references that look like you. Ask cities of similar size and structure what surprised them after go-live.
Contract clauses to request (and review with counsel)
Your city attorney decides what goes into the contract and how it is worded. The topics below are things to raise, not model language. Some vendors include these terms already, and others will negotiate. Raise them before the contract is final, when the city has the most influence.
Data ownership and use
- The city owns its inputs and outputs.
- The vendor and its providers may not use city data to train or improve models without written consent.
- Subprocessors are disclosed, with notice before new ones are added.
- Storage and processing locations are stated, with notice before they change.
Security and incidents
- The vendor maintains stated security practices and provides documentation on request.
- The vendor notifies the city of incidents affecting city data within a timeframe the city sets, and cooperates with the response. CISA publishes general guidance IT staff may find useful here.
- Insurance requirements, including cyber coverage, are set by your risk manager or attorney.
Changes to the product
- Advance notice of material changes to the underlying model or to how output is generated.
- A reasonable chance to test significant changes, or at least review results soon after.
- New AI features are not switched on for city data without the city's agreement.
- Remedies, possibly including termination, if a change materially reduces performance against acceptance criteria.
Records, acceptance and exit
- Cooperation with public records requests, subpoenas and legal holds.
- Configurable retention and city access to audit logs.
- Acceptance criteria tied to your pilot measures, with stated consequences if they are not met.
- Return of all city data in a usable format within a set period, deletion confirmed in writing (subject to legal holds), and described transition assistance.
- Termination for convenience, if your rules and attorney support it.
Order of precedence
Many products include online terms the vendor can update at any time, plus click-through agreements at login. Ask your attorney to confirm that the signed contract controls over those terms. Liability caps, indemnification and warranties are common sticking points and squarely for counsel. Your job is to explain the use case and risks clearly enough for counsel to weigh them.
Monitoring after go-live
Signing is the midpoint, not the finish. Because AI tools change and their errors are quiet, the city needs a light, steady way to keep watching.
Name one business owner, with a backup, usually in the department that relies on the tool most. IT, the clerk and the city attorney all have roles, but one person should hold the error log and the calendar. Keep the pilot error log running in production, and pull a small sample of finished work periodically to check against source records, even though staff review every output. Sampling catches errors that slip past busy reviewers and shows whether review habits are holding.
When the vendor announces a model update or new feature, the owner should decide whether it matters for your use case and rerun part of the pilot test set if it does. Keep the test set for exactly this purpose. Before renewal, compare results to your original problem statement and baseline. GFOA and ICMA both offer general resources on managing technology investments that can help shape that review.
Post-purchase checklist
- Business owner and backup named in writing
- Error log in place and reviewed on a set schedule, such as quarterly
- Periodic output sampling against source records
- Vendor change notices reviewed, with retesting when needed
- Access removed promptly when staff leave or change roles
- Audit logs reviewed periodically
- Use stays within approved scope; new higher-risk uses get their own review
- Annual reassessment against the problem statement and baseline
- Full test export completed and verified while the relationship is healthy
- Renewal decision documented
Next steps
You do not need to finish every item before you start. A few moves this month will put the rest in order:
- Write a one or two page problem statement for the task you most want to improve, with a measured baseline.
- Take an informal inventory of AI tools and features staff already use.
- Meet with your purchasing officer and city attorney to confirm the procurement path and the pilot terms you need.
- Build a test set from your own records, including the hard cases.
- Draft scoring criteria and weights before you invite any vendor in.
When you reach the pilot stage, run every product through the same test set and error log. If you are evaluating tools for agenda items and staff reports, Govera is one option to include: ask to test its source-cited drafts against your own records, and check the citations yourself.



