Government records AI — FOI requests, faster

You FOI'd the agency. Six months later you get a 2,000-page PDF — partially redacted, mostly garbage, with the relevant 30 pages buried somewhere inside. This is the FOI experience for reporters, watchdogs, and academic researchers everywhere.

A SeekFiles AI assistant scoped to the FOI dump turns a 40-hour read into a 4-hour query session. Doesn't replace the read; targets it.

The workflow

Step 1 — Upload the dump

PDFs from FOI responses, RTI requests, Sunshine Act requests. Even scanned, even redacted, even disorganised. Upload them all to a folder; SeekFiles will OCR and index everything.

Step 2 — Build an investigation-scoped Assistant

Name it after the investigation ("Pork Barrel Audit 2024" / "Mayor X Travel Reimbursements" / "DPWH Contract Series 2023").

Step 3 — Start with broad orientation

Ask:

• "What are the date ranges covered by these documents?"
• "What government offices are represented?"
• "Are there any documents marked confidential or partially redacted?"
• "What appears to be the most common document type — memos, contracts, receipts, emails?"

Get the lay of the land before you dive.

Step 4 — Targeted questions

Now drill in:

• "Find any reference to Vendor X. Quote the surrounding context."
• "Show me all expenditures over 1M with vendor + date."
• "Are there any documents that mention Person Y? Cite the page."
• "What was the total spent on travel in Q2 2023, by destination?"

Each answer cites the specific page — your reporting can defensibly say "page 472 of the FOI release."

Step 5 — The "what's missing" pass

This is what AI is genuinely bad at and you have to do manually: look at what's NOT in the documents. Gaps in date ranges, suspiciously absent vendors, documents whose responses are unusually short. AI can't tell you what isn't there. You can.

For journalism specifically

• Source protection. If your FOI dump contains witness names or whistleblower references, treat the AI carefully — use a privacy-tier subscription that doesn't train on your data.
• Verification standard. Citations from the AI are starting points, not endpoints. Verify the page yourself before printing.
• Cross-document patterns. "Show me every vendor that appears in both the 2022 and 2024 budgets." This is where AI multi-document Q&A shines.

For academic researchers

Same pattern, longer timeframe. Researchers studying public administration, urban planning, public health policy often FOI multi-agency document sets. Multi-year, multi-source. SeekFiles lets you query across the union of those documents in one chat.

For watchdogs and civic groups

Public datasets, audit reports, agency annual reports. Upload all years; ask trend questions ("How has spending on X changed since 2018?"). Cite directly to the audit page in your report.

What never to do

• Don't trust AI summaries of high-stakes documents without reading the source. The summary may be 95% right and 5% wrong; the 5% is where lawsuits live.
• Don't publish citations you haven't verified. AI sometimes pulls the right paragraph from the wrong page if OCR was messy.
• Don't FOI lazily and outsource the work to AI. Strong investigations still require strong manual reading. AI helps with the long tail.

A practical pilot

Take one FOI dump you've been putting off. Upload it. Spend an hour with the Assistant. You'll find at least three leads you would have missed by skim-reading.

That's the value. Not "AI did the reporting." Just "AI found me the haystack needles."