How AI Handles Budgets, Audits, Transcripts and Bills

AI is good at one research job above all others: finding the right page in a long document. It is mediocre at reading tables and poor at knowing which of two similar columns you meant. Once you know that split, a 248-page budget stops being a weekend and becomes an evening.

Finding primary sources

AI search engines are good at telling you what exists and who publishes it. They are bad at handing you the right copy. A prompt like this one asks for the map and nothing else:

What official documents would show how the invented Halden County
spends its road budget? For each, name the office that publishes
it and where on a county website it usually lives. Do not
summarize the contents.

A good answer is a list: the adopted budget (clerk or auditor), the annual financial audit (independent auditor, filed with the county), road department capital plans, commission minutes. The documents themselves live on the county's own website. A copy the AI links on a third-party file site may be a draft, an old year or somebody's edited version.

Worked example: the Halden County budget

Halden County is invented. Its proposed fiscal 2027 budget is a 248-page PDF posted on the clerk's page. Theo Branck, a retired bookkeeper who sits on a citizen budget committee, has one night before the budget hearing.

Step 1: Use AI as an index

The PDF is public, so uploading it to a cloud tool is fine. (Under 100 pages, Claude reads text and visuals; this one is longer, so the tool may read text only. Charts and scanned pages can be missed. Each tool's help page lists its limits.) Theo's first prompt asks for pages, not conclusions.

[budget PDF attached]

Give me the page numbers in this budget for: (1) the general fund
summary table by department, (2) any list of new positions,
(3) debt service, (4) transfers between funds. One line each with
page numbers. If you can't find one, say NOT FOUND.

It came back with pages 12, 37, 88 and 91. Page 88 was right. Page 91 was a glossary that mentioned transfers. The real transfer schedule was on page 203, which Theo found by searching the PDF for "transfer out." The index saved him time. It was not perfect.

Step 2: Pull the table, then foot it

Theo asked for the general fund summary from page 12 as CSV, with all three columns: 2025 actual, 2026 adopted, 2027 proposed. Then he pasted it into a spreadsheet and added the columns himself.

Footing the 2027 column: 7,412,000 + 3,905,000 + 4,035,000 + 702,000 + 1,214,000 = 17,268,000. It matches. The 2026 column: 6,840,000 + 3,250,000 + 4,120,000 + 610,000 + 1,180,000 = 16,000,000. Matches. The 2025 column: 6,512,000 + 2,980,000 + 4,290,000 + 575,000 + 1,141,000 = 15,498,000. Matches.

On his first try the AI had given the parks 2027 figure as $720,000. The column added to $17,286,000, which didn't match the printed total, and that one mismatch sent Theo back to page 12, where it said $702,000. Two digits swapped. Footing caught it in under a minute.

Step 3: The wrong-column trap

Theo then asked the AI which department grew fastest. It said the jail, up 31 percent. That number felt high, so he did the math.

Here is the full year-over-year picture Theo built in his own spreadsheet.

Parks grew faster in percentage terms than the sheriff, but the sheriff's raise is six times bigger in dollars. Dollars tell the size of big items and percent gives context. The jail line is the big one either way: $655,000 is more than half of the county's total increase of $1,268,000 (655 divided by 1,268 is 0.517).

One more context check: a general fund grows from more than property tax, and Idaho Code 63-802 limits how fast most taxing districts can grow their property tax budgets (generally 3 percent a year, plus amounts for new construction and annexation, with other exceptions). A 7.9 percent spending increase does not mean a 7.9 percent tax increase. The revenue page and the clerk can say where the money comes from.

Step 4: Questions, not conclusions

Theo's last AI prompt: "Based only on pages 12, 37 and 88, what questions should a resident ask the sheriff about the jail increase?" He got eight. Three were good (Is the increase from staffing, medical contracts or housing inmates for other counties? How many new positions? Is any of it one-time?). He took those to the hearing.

Meeting transcripts

Auto-transcripts are full of errors in names and numbers, and AI summaries of them drift toward whatever got discussed most. A long debate about a rate hike can turn into "the council approved the rate hike" even when the motion failed. A log of motions holds up better than a summary because every line points to a timestamp.

[transcript pasted]

Make a table of every motion: timestamp, mover, seconder, exact
wording if stated, result and vote by name. Then list every dollar
figure spoken, with speaker and timestamp. Mark anything unclear
as UNCLEAR. Do not summarize the discussion.

Each line can then be confirmed at its timestamp in the recording. Ten motions take about ten minutes. Official minutes, once approved, are the record of the action; the recording is the record of who said what.

Comparing versions of a bill

Bills change between printings. A one-word edit ("shall" to "may") can gut a requirement, and AI summaries are exactly the kind of reader that skims past one word. A two-pass check catches it.

1. Mechanical diff. Both versions come from the legislature's own site. Word's Review, then Compare, then Compare Documents marks every insertion and deletion. Free text-diff tools do the same.
2. AI explanation. The marked-up changes (not the two whole bills) go to the AI with a request like: "For each change, explain in one plain sentence what it does. Quote the old and new wording. Flag any change to who must act, deadlines, penalties or money."

The diff is the ground truth. If the AI explains six changes and the diff shows seven, you know which one it missed.