The $2,400 Card Charge With No PO Number: Capping City Hall AI Spend Before It Becomes a Budget Amendment

9:14 a.m., and there is a charge on the P-card statement nobody can code

The Finance Director for a city of about 42,000 people is doing the monthly purchasing-card reconciliation on the Tuesday after the statement closes. Everything ties out except one line: $2,412.68 to an AI vendor, on the Deputy Clerk's card, no purchase order number and no object code that fits. It is not fraud. The Deputy Clerk had a 4,100-page records request and a council packet due Thursday, and she solved her problem the way any competent person would. She just solved it outside the budget.

That single line starts a chain. The charge has to be coded to a fund. If it lands in the General Fund's IT line, the IT line is over, and it shows up at year end in the budget-to-actual schedule the auditors put in front of the council. Over by enough and the Manager brings a budget amendment resolution to a public meeting, where the newspaper that covers your council reads consent agendas for a living.

None of that is about whether AI was worth using. It was. It is about the fact that until 2 July 2026, AI was one of the very few things a city bought that had no ceiling on it.

How AI got into the budget without ever being in the budget

Walk any city hall and you will find the same pattern. Nobody adopted an AI budget. What happened instead is that six or seven people each found a use, each expensed it, and each charge was small enough to slide under the purchasing threshold that requires quotes. Your policy probably says anything under $5,000 needs no formal quotes, $5,000 to $25,000 needs three written quotes, and above that goes out for sealed bids. A $60 monthly subscription never trips any of that. Twelve of them, across four departments, across a fiscal year, is real money that never once appeared in a line item the council voted on.

The second problem is fund accounting, and it is the one that gets cities in trouble. Your water and sewer operations are an enterprise fund. They are supposed to pay their own way, and the rate study you paid a consultant to do assumes a certain cost of administration. If the Utility Billing Clerk's AI use is quietly riding on the General Fund's IT line, you have a cost allocation problem, and cost allocation problems are exactly what the audit management letter is for.

AI spend governance, in plain terms, is the ability to set a dollar ceiling per department per month, be warned before you hit it, and print a report showing who spent what — the same three things you already require of your fuel card program. Until this summer, the major AI vendors did not give a public finance office any of the three.

What the 2 July governance update actually gave a finance office

On 2 July 2026 the Claude Enterprise governance update added the pieces a municipal finance office has been asking for: a usage and cost dashboard, spend limits set at the organization level and at the individual user level, automatic alerts when a budget reaches 75% and again at 90%, default model settings, and entitlements that decide which staff may use which model. It also added reporting access so the numbers can be pulled into whatever you already use for reporting rather than screenshotted into a memo.

Translate that into your language. The spend limit is an appropriation ceiling. The 75% alert is the same signal you want on the overtime line in the third week of a bad snow month. The model default is the difference between the Clerk's office drafting minutes on the inexpensive model and the two people doing records-request review getting the expensive one. Entitlements are the AI equivalent of who has a key to the fuel island.

What changed compared to 2024 and 2025 is not the intelligence. It is that the vendor finally accepts that some customers cannot legally spend money they did not appropriate, and built the controls to match. That is the whole story, and it is enough to move AI from a card charge you discover to a line you adopt.

flowchart TD
  A["Department head asks for AI access"] --> B{"Is there a line in the adopted budget?"}
  B -->|No| C["Waits for the next budget amendment"]
  B -->|Yes| D["Finance sets monthly cap and default model"]
  D --> E["75% alert emails the department head"]
  E --> F{"Still climbing at 90%?"}
  F -->|Yes| G["Manager approves a one-time increase, or it stops"]
  F -->|No| H["Monthly report attaches to the P-card reconciliation"]

Setting the caps the way you set a fuel budget for the plow fleet

Do not start by asking what AI costs. Ask what each department is allowed to spend, exactly as you do with fuel, uniforms and training. Set the ceiling to that number and let the alerts tell you whether you guessed right.

A workable first pass for a mid-sized city gives Community Development the largest cap, because permit completeness review is where the volume is, and gives the Clerk's office the second largest, because records requests arrive in waves. Police get an administrative cap for policy manual updates and grant narratives, explicitly not for anything touching a case file.

Then encumber it. One purchase order, one object code, one line in the adopted budget with a name a council member can read aloud without needing it explained. When somebody asks at the June hearing what the city spends on AI, the answer should be a single number.

The arithmetic on a 42,000-resident city

These figures are an illustration, not a quote. Assume 23 licensed seats across six departments, and caps set as follows.

The 75% alert on the citywide ceiling fires at $1,035 and the 90% alert at $1,242, both of which arrive with days left in the month. That is the point. An alert on the 22nd is a decision. An overage discovered on the 6th of the following month is a budget amendment.

Against that $16,560, the honest comparison is what the same work cost last year: suppose the Clerk's office ran 380 hours of temporary clerical help at $24 an hour, which is $9,120, and the city attorney billed 40 hours of records review at $285, which is $11,400. If the AI line removes even half of each, the caps pay for themselves and you still have a hard ceiling on the downside. That last part is the actual product. You are not buying cheaper labor; you are buying a number that cannot run away from you between council meetings.

What a cap will not save you from

A spend limit tells you nothing about whether the work was any good. A minutes draft that misstates a motion is a problem no dashboard catches. Keep the rule that the Clerk certifies the minutes and reads every line before they go in the packet.

A cap set too tight is worse than no cap. If the Deputy Clerk hits her ceiling on the 14th during a heavy request month, she will go back to a personal account on her own card, and now city business is being conducted on a personal login that your retention schedule does not cover and your open-records response cannot search. Set the ceiling where the work actually lives, and make the one-time increase easy enough that nobody routes around it.

Governance also does not answer the legal questions. Whether your prompts and outputs are public records, how long you retain them, and what your state's 2026 AI statute requires if you sit in Texas, California, Colorado, New York, Utah, Nevada, Maine or Illinois — those are conversations with your city attorney, not settings in a dashboard.

Start with one department and one number

On Monday, pull three months of P-card statements and grep them for every AI vendor name you can find. Add it up. That number, whatever it is, is your current unbudgeted AI spend, and it is the number you take to the Manager. Then pick the single department with the most usage, set one cap, turn on the 75% and 90% alerts, and run it for one month before you touch anything else. One department's real curve teaches more than any vendor's estimate.

If your city's inbound phone volume is part of the problem — the utility billing line in the first week of the month, the permit counter at 8 a.m., the after-hours calls that go to voicemail and come back as complaints at the next council meeting — CallSphere builds AI voice and chat agents that answer those lines, book appointments, and capture the caller's details around the clock. Worth noting for the same reason as everything else here: ask for the usage report and the monthly ceiling before you sign, so the phone line shows up in your budget as a line, not as a surprise.

Frequently asked questions

Do we have to put AI software out for competitive bid?

That depends on your purchasing policy thresholds and your state's procurement statute, and above some dollar amount the answer is usually yes. Many cities sidestep it by buying through a cooperative purchasing agreement that was already competitively awarded, which most policies accept. Have your Purchasing Agent confirm in writing which vehicle you are using before the first invoice.

Which fund should this be charged to?

Whichever funds actually use it, in proportion. If Utility Billing holds a meaningful share of the seats, the enterprise fund should carry that share and your cost allocation plan should say so. Burying it all in the General Fund IT line is what draws a comment in the management letter.

What do I tell the council when they ask?

One sentence and one number: the city has a hard monthly ceiling of X dollars across Y departments, alerts fire at 75% and 90%, and here is last month's actual by department. Councils are rarely opposed to the tool. They are opposed to not knowing the number.

Are the prompts and outputs public records?

Assume yes until your attorney says otherwise. In practice that means keeping the work inside the city's account rather than on personal logins, and fitting AI drafts into the Clerk's existing retention schedule instead of inventing a category for them.