Skip to main content
DoneThat

AI Adoption GuideGovernmentReport

Public-facing data summarizer

LLM converts technical program performance data into accessible plain-language summaries for public and media release.

Government processPlanFundAuthorizeDeliverInspectEnforceReportClose

By Don, DoneThat’s AI coach · updated

The public sentence has to match a locked cell

A public-facing data summarizer earns its keep only when every claim in the plain-language brief traces to a cell in a table the agency has already released. The model turns technical program language into copy residents and reporters can use. It does not add a finding, a percentage, or a success story the freeze does not contain.

Quality on this page is a cited sentence, not a smoother paragraph. If the released extract lists units completed in a named cell, the brief may state that figure and name that cell. If the table has no completion rate, the brief has no completion rate. Empty stays empty. The public information officer still publishes the rest.

This rewrite sits at the end of a packet that is already public. It does not assemble statutory exhibits. Use the legislative compliance report assembler when the audience needs attachments and citations to law. It also does not explain why a line moved. Use the variance explainer when the question is the swing, not the wording.

Freeze the extract before anyone drafts

Lock one extract: file name, as-of date, tab, and cell range. That freeze is the only source the summarizer may read. If program staff later send a "better" spreadsheet, it does not enter the draft until it is released under the same rules as the first table and the freeze is updated.

Agencies already keep program tables in portals and records systems from vendors such as Microsoft, CivicPlus, OpenGov, and Granicus. Treat those products as a class of places a table can live. None of them is the publisher of the media note. The PIO is. Do not prompt the model against a live view that can change while the release is in review.

If operations flagged unusual rows before the table went public, finish that work with an anomaly-surfacing dashboard first. A public summary that "explains" a spike the agency has not locked is a new claim, not a rewrite.

One walkthrough is enough. A housing rehabilitation program has already posted its quarterly performance table. The locked extract lists applications received in cell B4, units completed in cell B7, and a blank for average days from approval to start. That blank is part of the freeze. Staff may know a typical wait from an internal tracker that was never released. That tracker is not in the packet, so it is not in the brief.

Draft only what you can cite

Write the brief as sentences, each with a cell cite on the same line. Do not hide the cite in a footnote the PIO will miss on deadline.

Sentences that pass look like this: the program received the applications shown in cell B4 of the locked quarterly table; the program completed the units shown in cell B7 of the same table. Those lines restate released cells. They do not interpret them.

Sentences that fail invent a relationship the table never stated. "Most applicants moved into completed units" is not allowed unless a cell states that relationship. A lead that opens with a success rate is not allowed unless the freeze contains a success-rate cell. Inventing a rate so the note sounds like news is a quality miss, not a style choice.

Keep the register plain: short sentences, named programs, named periods, and the same nouns the table uses. If the table says "units completed," do not upgrade that to "families housed" unless a released cell uses that phrase or defines the mapping in the freeze.

When the same facts must go out in more than one language, generate those versions from this cited brief, not from a second pass at the raw table. Hand the locked sentences to the multilingual communication generator so translations cannot drift from the freeze.

Leave silent cells silent

A blank cell is an instruction. The model must not fill it with an estimate, a prior-period figure, or a hedge that still asserts a number. Phrases such as "on the order of" or "roughly in line with last year" are still claims if no released cell supports them.

If a reporter will ask for average days from approval to start, the brief can say the released table does not report that figure. That sentence cites the blank. It does not invent a wait time.

The first failure mode is a claim with no cell: a vivid clause that sounds true because someone said it in the hallway. Hallway knowledge belongs in a later release after it is in a locked table, or it stays out.

The second failure mode is treating the model output as published. A draft in a chat window, a shared document, or an email thread is not a release. Mark it as draft. Keep it off the newsroom list, the homepage, and social channels until the PIO signs.

The third failure mode is an invented success rate, often labeled "impact" or "results so far." If the table is silent on rate, the brief is silent on rate. A partial, cited note is the correct public product. A complete-sounding note with unsourced rates is not.

The PIO still publishes when some lines are blank. Completeness is not the test. Traceability is.

What the PIO checks before the wire

Read the draft against the freeze, line by line. For each sentence, point at the cell. If you cannot, cut the sentence or replace it with an explicit "not reported in this table."

Check quotes and headlines the same way. A subject line that adds a percentage the body never cites is still a public claim. A pull quote that turns "units completed" into "crisis solved" is still a public claim.

If the table later gets an erratum, freeze the erratum as a new extract and rewrite only the sentences whose cells changed. Do not silently patch the old brief and leave the old cites in place.

Vendors in the Microsoft, CivicPlus, OpenGov, and Granicus class may host the file, version it, or display it. They do not clear quotes. Clearance stays with the PIO, including who may send the note and when.

A reporter who wants more than the freeze contains gets a referral to the next public table, not a private number from the model. If the next table is not yet released, the answer is that it is not yet released.

What this page does not cover

This page does not assemble statutory attachments or decide which exhibits belong in an oversight packet. That work belongs with the legislative compliance report assembler.

It does not explain variance. If the public question is why units completed moved, lock the facts here first, then attach or point to the variance explainer.

It does not replace language-access workflow. Translations follow the cited brief. They do not get a second chance to interpret the spreadsheet.

It does not decide which anomalies are fit to print. Surface and resolve those before the freeze. A summarizer that "helpfully" narrates an unresolved spike is writing news the table did not release.

The working rule is small on purpose. Lock the released table. Summarize with cites. Leave blanks. The PIO publishes.

Is this worth automating for you?

Whether this pays back depends on how much time it takes your team today. Most teams estimate that from memory, and the estimate is usually wrong in one direction or the other.

DoneThat reconstructs where the time actually went, with no timers to forget, so you can measure the baseline before committing to a project and check the gain afterward.

Measure the baseline first