From PDF to Insight: How AI Reads Your School Documents

The most common question I get from school leaders after their first quality review: "How does it actually work?"

Fair question. After conducting 300+ school quality reviews across 13 states, I helped build SchoolQualityReview to codify what I look for into a system that does the document analysis automatically. When you upload a document, you see a progress bar, then indicator scores start appearing. But between "upload" and "scores," a lot happens — and understanding the process helps you trust the results.

Here's exactly what happens to your documents. No black boxes.

Step 1: Document Upload and Text Extraction

When you upload a file — PDF, DOCX, XLSX, PPTX — the first thing that happens is text extraction. The system needs to convert your document into something it can read.

For digital PDFs (the kind you create from Word or Google Docs), this is straightforward. The text is already there — we extract it directly using standard PDF parsing.

For scanned PDFs (from a physical scanner or fax), there's an extra step. The system uses AI vision to read the scanned pages, converting images of text into actual text. This works well for clean scans, but very low-quality scans or handwritten notes produce less reliable results.

For spreadsheets (XLSX), the system extracts cell data and structures it so the AI can interpret financial tables, attendance data, and assessment scores.

The key thing to understand: the quality of your original document matters — for any AI tool, not just ours. A well-formatted digital PDF produces cleaner text than a third-generation photocopy scanned at low resolution. When possible, upload digital-native files.

Practical tip: If your school is still printing documents and re-scanning them as PDFs, stop. Export directly from Google Docs or Word. You'll get better results from any AI tool and save yourself the scanning step. Anything that started digital should stay digital.

Step 2: Document Chunking for AI Analysis

A 50-page strategic plan can't be fed to an AI model all at once. Documents are split into chunks — small enough to process but large enough to preserve context.

Chunks are split at natural boundaries (section headings and content breaks) so that a finding about "governance practices" doesn't get cut in half mid-sentence.

Why does this matter to you? It means every part of every document gets read. A piece of evidence on page 47 of your board minutes is just as likely to be found as something on page 1. There's no "the AI only reads the first few pages" problem.

This is worth asking any AI vendor: does your system process the entire document, or just the first few pages? Some tools truncate long documents to save costs.

Step 3: AI Assessment Against 82 Quality Indicators

This is where the real work happens. The 82 indicators span academic performance, organizational health, and financial sustainability — the same areas I evaluated during in-person reviews, now codified into measurable criteria. For each document, the system creates assessment tasks: one for every indicator that might have relevant evidence.

For each task, the AI receives:

The text chunk from your document
The indicator definition (what "clean audit reports" means, what evidence looks like)
Scoring criteria (qualitative 0-100, or quantitative threshold comparison)

The AI then does what a human reviewer would do, but across all 82 indicators simultaneously:

Search for relevant evidence — Is there anything in this chunk that relates to this indicator?
Extract evidence excerpts — Pull the specific sentences or paragraphs that serve as evidence
Assess quality — How strong, specific, and direct is this evidence?
Score — Assign a score based on the evidence found

For qualitative indicators, the score reflects evidence strength: 90 means clear, direct evidence; 50 means indirect or incomplete; 0 means nothing relevant was found.

For quantitative indicators, the AI extracts a specific value (attendance rate: 94.2%) and compares it against a defined threshold (≥ 92%). Meets, doesn't meet, or unclear — no subjective judgment involved.

For calculated indicators, the system computes financial ratios using deterministic formulas. No AI interpretation — just math.

Step 4: Score Aggregation and Rollup

Once all tasks complete, the system aggregates results at four levels:

Indicator rollups — Evidence from all documents combined into a single score per indicator
Standard rollups — Indicator scores averaged within each quality standard (e.g., Academic Standard 1, Organizational Standard 3)
Domain scores — Standards rolled up into Academic, Organizational, and Financial domain averages
School summary — An AI-generated narrative interpreting the overall pattern

Your final score for any indicator reflects everything the AI found across all your documents, not just one.

Step 5: Evidence Citations and Traceability

This is the part that matters most for trust. Every score links back to the specific evidence that produced it. Click on any indicator finding and you see the AI's assessment, direct quotes from your documents, which document each quote came from, and a relevance score for each piece of evidence.

If the AI says your governance practices scored 75, you can see exactly which board minutes it cited, which sentences it pulled, and judge for yourself whether the score is fair.

This is what I always wanted from the manual review process. In 20 years of conducting school quality reviews, the hardest part was showing a school leader exactly where a finding came from. I'd have notes spread across a legal pad and three binders. Now every conclusion traces back to a specific sentence in a specific document — and you can verify it yourself.

How Long Does It Take?

The whole process — upload to final report — typically takes 1-3 hours depending on how many documents you upload. The assessment stage is the longest because it's doing the most work: potentially thousands of individual evaluations (82 indicators multiplied by the number of document chunks). Each one is a thorough analysis, not a keyword scan.

For context, the equivalent manual review takes 4-6 weeks and $10,000-$25,000 in consultant fees.

Common Questions

"Does the AI hallucinate findings?" Every finding is grounded in evidence from your actual documents. The AI doesn't invent evidence — it reports what it found. If it found nothing, the indicator scores zero. You can verify every citation against your original document.

"What if two documents contradict each other?" The rollup aggregates evidence from all documents. The combined score reflects both strong and weak evidence, and the citations show you both sources so you can judge for yourself.

"Is my data secure?" Documents are stored in Google Cloud with organization-level isolation. No other user can access your documents. Data is encrypted at rest and in transit.

Want to see this in action? The free tier includes 6 indicators so you can test the full pipeline — upload, extraction, assessment, evidence citations — before committing. Try it with your documents →