How to use AI with PDFs, spreadsheets and images safely

Before uploading a file to AI, confirm that the data and tool are approved. During analysis, require traceable evidence such as page numbers, cells, and quoted passages, then verify important calculations and conclusions in the original file.

AI can work across PDFs, spreadsheets, images, and audio.

That makes it useful for real business work—and creates more ways for a plausible answer to hide a source error.

The safest approach keeps the original file, evidence, and accountable reviewer visible.

Check permission before capability

Ask:

• Does the file contain personal, sensitive, confidential, privileged, credential, or payment information?
• Is this product and account approved for that information?
• Who can access uploaded files and generated output?
• How long is the data retained?
• Can the provider use it to improve services or models?
• Does processing occur outside Australia?
• Is the use consistent with customer or employee notices and permissions?

The OAIC’s commercial AI guidance recommends due diligence on the product, data flows, security, human oversight, and privacy obligations.

Working with PDFs

PDFs can contain:

• Real text.
• Scanned images requiring optical character recognition.
• Tables with complex layout.
• Footnotes and sidebars.
• Headers and repeated boilerplate.
• Pages that are missing or out of order.

Ask the AI to:

1. State the document title, date, and page count it can access.
2. Identify unreadable pages or sections.
3. Cite page numbers for every important conclusion.
4. Quote a short supporting passage.
5. Separate direct statements from inference.

Check the cited page yourself.

Working with spreadsheets

AI is useful for exploring a workbook, explaining formulas, finding unusual values, and suggesting analyses.

For exact work:

• Confirm the correct sheet and range.
• Check hidden rows, filters, units, dates, and blank values.
• Ask which columns and formulas were used.
• Recalculate totals with spreadsheet formulas or code.
• Reconcile the result to a known control total.
• Preserve the original file and version.

Do not let a nicely worded summary conceal an incorrect denominator.

Working with images

AI can describe photographs, diagrams, screenshots, forms, equipment, and visual defects.

Limitations may come from:

• Low resolution or blur.
• Cropped context.
• Similar-looking components.
• Small labels.
• Lighting, perspective, or occlusion.
• Inference about people, emotion, health, identity, or intent.

For inspections, safety, medical, or identity-related use, involve a qualified person and use purpose-built processes.

Working with audio and transcripts

Transcription can confuse names, accents, technical terms, numbers, and speakers.

Keep the recording where lawful and appropriate, then:

• Correct names and key terms.
• Check decisions, commitments, amounts, and dates against the audio.
• Mark uncertain passages.
• Confirm who said what.
• Obtain required consent and follow retention rules.

Use a two-pass prompt

Pass 1: source inspection

Inspect the supplied file.
List its title, date, structure, readable sections, missing or unreadable content,
and any limitations that could affect analysis. Do not draw conclusions yet.

Pass 2: bounded analysis

Answer the question using only the supplied file.
For each conclusion, provide the page, cell/range, image region, or transcript timestamp.
Separate direct evidence, calculation, and inference.
If the evidence is insufficient, say so.

This makes gaps visible before the final answer.

Preserve provenance

Record:

• Original filename and version.
• Upload date.
• Tool and account used.
• Prompt or workflow version.
• Output and citations.
• Reviewer and approval date.

For recurring work, this helps investigate errors and update the process.

OpenAI’s working with files guide demonstrates the range of document, spreadsheet, image, and connected-file tasks now possible. Business value comes from combining that capability with source discipline.

Review what business data should not go into a public AI tool and use the fact-checking process for consequential findings. Rising Tide’s AI training teaches these habits using real team workflows.