AI Training for City Staff: What to Teach and in What Order

A lot of AI training in city government starts in the wrong place. A vendor runs a webinar on features, someone circulates a list of clever prompts, and staff learn how to get a tool to produce text but not when to stop and check it. That order is backward. Before anyone writes a better prompt, they need to understand what the tool is actually doing and where it breaks. Before they type anything at all, they need to know which information cannot go in.

This guide lays out a curriculum in six modules, taught in this order:

1. What the tools do and where they fail
2. Data rules
3. Prompt hygiene
4. Verifying output
5. Records implications
6. Role-specific exercises

Each module depends on the one before it. Staff who do not understand why a tool invents details will not take verification seriously. Staff who have not learned the data rules will paste the wrong document into a prompt while practicing. Staff who cannot verify output should not be running exercises with real work products.

After the modules, you will find guidance on running short sessions and on keeping a training record that holds up when a council member, auditor or records requester asks who was trained on what. This assumes your city has an AI use policy or is drafting one. The data and records modules depend on state law and local policy, so have your city attorney and records manager review those materials before you teach them.

Module 1: What the tools actually do and where they fail

Start with a plain explanation, not a technical one. Most text tools staff will encounter are built on large language models. In practical terms, they produce the most likely next words based on patterns learned from a large body of text. Unless a tool is connected to a specific set of documents, it is not looking anything up. It is generating text that sounds right.

That one fact explains most of what goes wrong, and staff should hear it in the first ten minutes.

The kinds of tools staff will run into

• General chatbots. Open-ended tools that draft text from their training. Their knowledge of your city is thin or nonexistent.
• Tools grounded in specific documents. These draw on a defined set of records, such as your agendas, minutes and staff reports, and may point to where a statement came from.
• Transcription and meeting summary features. Often built into video meeting or recording software.
• AI features inside software you already license. Summaries or drafting help a vendor adds to an existing product, sometimes turned on by default.

Tell staff which of these the city has approved, and just as clearly, which it has not.

Failure modes to name out loud

Hands-on testing turns up the same problems again and again. Teach them by name, because a problem with a name is easier to spot:

• Invented specifics. Statute section numbers, dates, dollar amounts, vote counts and quotes can be generated as easily as real ones, and they look identical.
• Confident tone regardless of accuracy. Tone tells you nothing.
• Stale or general knowledge. A general tool may describe how something works in most places, not in your state, and not after your last ordinance amendment.
• Mixing up similar items. Two meetings with similar agendas, two versions of an ordinance, two applicants with similar names.
• Dropped details. Summaries of long packets or recordings can skip conditions, amendments or dissenting comments.
• Agreeing with the premise. Ask a leading question and the tool will often build an answer around your assumption, even a wrong one.

A demonstration that works

Show these problems live with your approved tool. A few exercises that tend to be instructive:

• Ask a general tool for your state's open meetings notice requirement, including the statute citation. Then have staff look up the actual statute and compare.
• Ask it to summarize a city ordinance by a number that does not exist and see whether it declines or produces a summary.
• Give it a set of published minutes and ask who made the motion on a specific item. Check the answer against the minutes.

You cannot predict what the tool will do on a given day, and that is the point. Right or wrong, staff had to check to know.

End with a working rule: treat every output as a first draft from a fast, articulate new hire who has never worked for your city and never attended a meeting.

Module 2: Data rules, before anyone types

This module comes second because it has to happen before staff start practicing. The first prompt someone writes is often built from whatever document is open on their desk, and that document may be a personnel file or a closed-session memo.

Three categories, not a hundred rules

Group information into three categories that match your policy, with examples from your own city's work:

Your categories may differ. Privacy law, rules for criminal justice data and public records exemptions vary by state. Confirm the lists with your city attorney and IT before training.

The account matters as much as the data

The most common data mistake is not malicious. It is a staff member using a personal account on a consumer tool for city work. Explain why that matters:

• A licensed tool may be governed by contract terms the city negotiated on storage and model training. A personal account is governed by consumer terms the city never saw.
• Staff usually cannot tell from the screen how long data is kept or who can see it.
• Material in a personal account is harder for the city to find when a records request or litigation hold arrives.

A simple rule: if the tool is not on the approved list, treat anything you enter as if you posted it publicly. And remember that entering data includes uploading files, pasting email text, uploading screenshots, running transcription on a recording and connecting a tool to a shared drive.

Exercise: the sorting stack

Print 12 to 15 short scenarios on cards. For example: "A code enforcement officer wants to tighten the wording of a violation letter that includes the property owner's name." "A clerk wants a summary of last month's published minutes." "A supervisor wants help drafting a performance improvement plan." Have small groups sort them into the three categories, then discuss the ones where groups disagreed. The disagreements are where the teaching happens.

Module 3: Prompt hygiene

Notice the word hygiene. This is not a module on tricks for more impressive output. It is about writing prompts that reduce errors, protect data and make output easier to verify.

Habits worth teaching

• State the task, the audience and the format. "Draft a two-paragraph background section for a staff report to city council, in neutral language" leaves far less room for drift than "write about this."
• Supply the source instead of asking from memory. Attach the adopted ordinance or the published minutes, and tell the tool to use only that material.
• Tell it what to do when information is missing. Ask for a marker such as [NEEDED: date of planning commission hearing] instead of a guess.
• One task per prompt. A summary, a recommendation and a press release in one request makes each part harder to check.
• Strip identifiers you do not need. Improving the wording of a letter does not require the resident's name and address.
• Do not lead the witness. "Why is this project a good fit?" invites advocacy. "Summarize the factors the planning commission considered" invites a summary.

Weak and better prompts side by side

Exercise: rewrite three prompts

Give each participant three weak prompts drawn from their department's work. They rewrite each one, run both versions in an approved tool using public documents, and compare which output was easier to check. The comparison makes the case better than any slide.

Module 4: Verifying output

This is the core skill, and it deserves the most time.

A repeatable check

Give staff a procedure they can run every time, not a general instruction to "review carefully."

1. Mark every specific. Highlight every number, date, name, dollar amount, vote, citation, legal statement and quote.
2. Trace each one to a source. Open the document it should have come from and find it. If a tool points to a source document and page, open that page. A citation is a lead, not proof.
3. Look for what is missing. A condition of approval, an amendment, a dissenting vote, a public comment that changed the outcome.
4. Check framing and tone. Watch for adjectives that advocate and summaries that tilt toward one side.
5. Check currency. Is the ordinance cited still the adopted version? Has the budget been amended since?
6. Decide and own it. Edit, discard or approve. Once it leaves your desk, it is your work product, not the tool's.

Expect verification to take time

Be honest with staff and supervisors. Checking a draft properly takes real time, and on short documents it can take nearly as long as writing from scratch. The value is often in organizing material and producing a structure, not in skipping review. Supervisors who expect AI-assisted work to arrive instantly will get unverified work.

Exercise: seeded errors

This is the most useful exercise in the curriculum. Before the session, the facilitator takes an AI-drafted paragraph based on public city documents and plants several errors: a wrong meeting date, a motion attributed to the wrong council member, a condition of approval that was never adopted, a dollar figure off by a digit. Staff get the draft and the source documents and hunt for problems.

Debrief on two questions. Which errors did people catch, and how? Which did they miss, and why? The ones that sounded plausible are usually the ones that slip through, and that experience does more to change habits than any warning.

Who signs off

Accountability usually stays with whoever would have been responsible without AI: the report author, the clerk for minutes, the department head for a memo. Your policy should say this, and training should make sure everyone has heard it.

Module 5: Records implications

Teach this after verification, because staff need to understand the full workflow before they can see where records are created inside it. The answers vary by state and by city. Teach your city's local answers, confirmed by your city attorney and records manager, not a general rule borrowed from elsewhere.

Questions staff should be able to answer

• Are prompts and outputs public records here? In many states, whether something is a record depends on its content and its connection to public business, not its format.
• Where should AI-related records be kept? In the city's approved tool, a project folder, or attached to the final document. "In my chat history" on a personal account is the wrong answer.
• Which retention rules apply? Staff do not need to memorize the schedule, but they should know who to ask about drafts and working files.
• What about personal devices? AI exchanges on personal phones about city work may still be subject to records requests in some states.

Meeting transcripts and AI summaries

• The official minutes are the minutes the body approves. An AI summary or transcript is a working aid for the clerk unless your city has decided otherwise.
• Transcripts and recordings may be records in their own right, with their own retention requirements.
• Transcription accuracy varies with audio quality, crosstalk and proper nouns. Names of residents, streets and projects are frequent trouble spots.
• A recording that includes a closed session should not go into a transcription tool unless counsel has approved that use.

Also cover disclosure (if your policy requires noting AI assistance, show exactly what the note looks like) and litigation holds, which may reach AI prompts and outputs related to the matter.

Exercise: follow a records request

Walk through a realistic request: "All records relating to the Elm Street rezoning from January through June." Ask each participant where AI-related material for that project would live in their workflow, and whether they could find it in a reasonable time. Gaps found here are better found in training than during an actual request.

Module 6: Role-specific exercises

Now put it together with work staff actually do. Group participants by role, and use public city documents for every exercise.

Clerks and records staff

• Compare an AI transcription of a past meeting against the approved minutes. List every discrepancy in motions, votes and names.
• Use an approved search tool to find past council actions on a recurring topic, then confirm each result in the original minutes.

Planning and community development

• Draft the background section of a staff report from a closed application file, then trace every fact to the file.
• Ask a tool to list applicable code sections for a sample project, then check each against the current adopted code. Invented or outdated citations tend to surface here.

Finance and budget staff

• Draft a narrative explaining a budget variance from an adopted budget document and verify every figure against the source table.
• Have the tool total a set of line items, then check the arithmetic independently. Never rely on a text tool's math unchecked.

Communications and public information

• Rewrite a legal notice in plain language, then confirm every legally required element survived.
• Discuss translation. AI translation of public materials should go through qualified human review, and your state may have specific language access requirements.

Department heads and council support staff

• Review a facilitator-prepared AI draft with seeded problems as if approving it for council. What questions would you ask the author?
• Prepare a short briefing on an agenda item from the packet, then check it for neutrality and for omissions a council member would care about.

Police, fire and other specialized departments often operate under additional rules for criminal justice, health or emergency information. Coordinate their exercises with their own compliance staff.

Running short sessions that fit a working week

Long training days compete with council meetings, permit counters and budget season. Short, hands-on sessions work better.

Format

• Length: 30 to 45 minutes, one module per session. Module 4 may need two.
• Split: roughly a third explanation, two thirds hands-on work and discussion.
• Group size: 6 to 12 people, so everyone does the exercises.
• Mix: cross-department groups for modules 1 through 5, role-based groups for module 6.
• Materials: real public documents from your own city. Staff take local examples seriously.
• Takeaway: a one-page card with the data categories, the verification steps and who to call.

A sample schedule

Consider granting access to approved tools only after sessions 1 and 2, so people know the limits and data rules before they have a login.

Who should facilitate

Pick someone who has used the approved tools and will show their failures honestly, often the clerk, records manager or an IT lead, with the city attorney reviewing modules 2 and 5. Vendor training is useful for learning features but should not replace the city's own sessions on data, verification and records. Small cities can share sessions with neighbors or draw on their state clerk association or municipal league.

New hires and refreshers

• New hires: modules 1, 2, 4 and 5 during onboarding, before tool access, with role exercises in their first month.
• Policy or tool changes: a short refresher on what changed.
• Annual refresher: rerun the seeded errors exercise with fresh material to check whether habits have slipped.

Keeping a training record that holds up

If a problem reaches council, an auditor asks about controls or an employee disputes discipline involving AI use, the city needs to show who was trained, on what, and when.

What to record

For each employee and each module:

• Name, title and department
• Module completed and date
• Version of the training materials and of the AI use policy in effect
• Delivery format and facilitator
• Whether hands-on exercises were completed
• Signed or electronic policy acknowledgment
• Tool access granted after training, and by whom
• Refresher due date

The record means little if you cannot show what the training said, so keep a dated, version-numbered copy of the slides, handouts, scenario cards and exercises. Store training records where your city keeps other employee training records, usually with human resources, and retain them under the series your records manager identifies.

Training record checklist

• Record template created with all fields above
• Policy acknowledgment form reviewed by city attorney
• Training materials dated and version-numbered
• Storage location and retention confirmed with records manager
• Tool access linked to completed training, agreed with IT
• Refresher dates calendared
• New hire process documented
• Program owner named

Next steps

You do not need to finish everything before starting. A practical sequence for the next month:

1. Confirm the local answers. Meet with your city attorney and records manager to settle the data categories and records rules for modules 2 and 5.
2. List the approved tools. Work with IT so the curriculum covers exactly what staff can use, and what they cannot.
3. Build the seeded errors exercise first. It takes the most preparation and teaches the most.
4. Run a pilot group. Take 6 to 8 staff from different departments through the sequence, then revise based on what confused them.
5. Set up the training record before the first full cohort, not after.

If your city is evaluating tools that draft from its own records, use them as practice material for module 4. Govera, for example, cites the source document, page and meeting behind each draft, which makes tracing faster, but the training point stays the same: staff open the source and review every output before it goes anywhere.