How AI file access works: uploads, local folders and connected sources
Understand what an AI can read, how to share useful files, and how to turn an approved folder into reports, searchable collections and repeatable workflows.
Your files are context. Give that context a structure.
“Can you read my files?” has several different answers. You might attach a single document, connect a service or open a local project. The useful question is which files this task can actually read and which outputs it can create.
Three ways files enter the task
An upload gives the service a copy of the selected material. A connected source lets a supported integration retrieve records under its authorization. A local agent works with paths available in its execution environment. These arrangements differ in freshness, persistence and permissions. Ask which arrangement the task is using.
A document attached last month may not reflect today’s changes. A connected search can return only some matching records. A local folder can contain files the parser cannot open. Have the assistant identify the exact files or records used in the output and their dates when relevant. This makes “I used your files” a claim you can inspect.
Start with a three-folder workspace
Create inputs, working and outputs folders. Put a few approved copies in inputs. Keep intermediate scripts or notes in working. Ask for finished artifacts in outputs. A README at the top can explain the goal and which materials are authoritative. The names are a convention for this exercise, not a required product format.
Begin with an inventory: filename, type, apparent purpose, whether it was successfully read and any limitation. For a large collection, have the assistant inspect the structure and a representative sample before opening everything. Exclude irrelevant files so the task stays focused.
A request you can adapt
Inventory the approved files in inputs. Report filename, format, what you actually read and anything you could not parse. Identify duplicates by content where practical, but do not move or delete files. Propose a structure for the result and create new artifacts in outputs. Preserve the inputs.
Choose the representation that preserves meaning
A screenshot can communicate layout but is a poor substitute for a table when you need exact arithmetic. A scanned PDF may need text recognition, and that recognition can confuse columns or characters. A spreadsheet formula is different from its displayed value. Ask the agent to explain extraction limits that could affect the conclusion.
Provide a small reference example when the formatting carries meaning. If indentation indicates a hierarchy, say so. If red text marks a proposed change, identify that rule. For a spreadsheet, specify which sheet and which rows are data, as opposed to notes or totals. Better input orientation can prevent an entire analysis from being built on the wrong range.
Ask for artifacts that preserve a trail
A useful transformation leaves enough information to trace a result back to its input. Keep source filenames in extracted tables, original row numbers in exceptions, and source dates in a report. When the assistant combines files, ask for an input manifest with counts before and after.
Check the output in the application where it will be used. Open the spreadsheet and inspect formulas. Open the document and inspect page breaks. Run the generated script on a fresh sample. File creation is a step toward a finished result, not proof that the result is usable.
Create documents people can actually use
Use the smallest useful access
Keep the first exercise to public, synthetic or otherwise approved material. Local access does not establish offline processing. Connected services may expose additional records, and a project can include more folders than the one you meant to use. Check the actual scope and the provider’s data settings for sensitive work.
If the task is blocked from reading a path, name the needed file and the reason it matters. A focused upload or narrowly approved folder may solve the problem. Giving an agent access to an entire drive is rarely necessary to summarize a handful of documents.
Take this into your next task
Identify the access method, inventory what was actually read, preserve the inputs and make every important output traceable.
Questions people ask
Can a website silently read my computer’s folders?
An ordinary page cannot browse arbitrary local folders. This site’s CSV demo reads only a file you explicitly select or text you paste. A local agent or connected application has its own access controls.
Does local file access mean my documents stay entirely on my computer?
Not necessarily. A locally running client can still send selected content to a hosted model. Check the service, tools and applicable data rules before supplying sensitive files.
Sources and editorial notes
Reviewed September 6, 2026. Product capabilities depend on the app, plan, region and workspace settings. Workflows and prompts are authored teaching material, not recorded model results or measured time-saving claims.
Keep going
- AI for Spreadsheets: From Export to Checked Report
- Organize a Messy Folder with AI
- Getting Started with Codex: Your First Project
Related Power of AI pages
Keep reading with Start here, everyday uses, the tool guide, the writing workshop, and sources and standards.