# What an AI agent can and cannot do with your files | CoLateral Source: https://colateralai.com/resources/ai-agents-on-engineering-files [Home](/) [Resources](/resources) # What an AI agent can and cannot do with your files Give an agent a clear task and a source you can check. August 18, 2026 Updated October 2, 2026 4 min read Workspace preview with invented sample content and staged agent conversations. Layouts may vary by release. In this guide [Start with access and scope](/resources/ai-agents-on-engineering-files#section-1) [Choose work with a checkable result](/resources/ai-agents-on-engineering-files#section-5) [Watch for four common failure modes](/resources/ai-agents-on-engineering-files#section-8) [Keep the review beside the source](/resources/ai-agents-on-engineering-files#section-17) [Know where data can go](/resources/ai-agents-on-engineering-files#section-20) [Where CoLateral fits](/resources/ai-agents-on-engineering-files#section-24) Access to files can make an agent useful for drafts, code and repetitive work. It also makes the current revision, permissions and actual output worth checking. ## Start with access and scope A chat assistant can work from what you type or attach. An agent may also read a working folder, run commands or write files, depending on the connection and permissions you give it. Some connections operate locally; others use a browser or cloud environment. Check the selected path rather than assuming they all reach the same files. Before the first request, identify the current source, the output you want and where that output should go. Ask the agent to name missing information instead of inventing it. Keep a recoverable copy when the task can change valuable work. A useful request: Read only the current files in this example folder. Draft a short summary with the filename and section behind each point. Flag missing or conflicting information. Do not edit the source files or publish the result. ## Choose work with a checkable result - Summarize documents with references you can open and compare. - Extract a table or index, then compare a sample of rows with the source. - Reformat or reconcile lists using a rule and a before-and-after example. - Draft a meeting record, content outline or covering note from supplied material. - Write code or build a small tool, then inspect and run it against an expected result. These tasks have an output you can inspect. An agent can still miss a page, misread a value or produce incorrect code. A fluent explanation helps you understand its work, but the source and the actual result are what you check. ## Watch for four common failure modes ### The wrong revision A summary can look convincing while describing an outdated file. Keep the current material clearly named, identify it in the request and check the files the agent actually used. If it cannot find the revision, resolve that gap before using the answer. ### An assumption presented as a fact A missing value can become a plausible sentence or table entry. Ask for unknowns and assumptions to be listed separately. Compare extracted figures with the original page, especially after copying them into another document. ### A number without a reproducible calculation A number in prose does not establish how it was calculated. For a result that depends on arithmetic, ask for an executed formula, sheet or script with its inputs and units. Then check the method and a known example. Correct execution alone does not establish that the chosen method is appropriate. ### Changes outside the task An agent with write access may change more than you expected. Set the intended scope and inspect the changed files or version-control diff. Permissions and confirmation behavior depend on the agent and action; a request for a narrow edit is not proof that every edit stayed within it. ## Keep the review beside the source Put the task, current source and draft where you can compare them. Record what you checked and what remains unresolved. For code, run the behavior you need. For an extracted table, check representative entries and any critical values. For a published message, review the exact text and destination. Review and professional responsibility remain with the person using the result. For regulated work, the qualified professional responsible decides what is appropriate and what may be issued. The workspace and an agent response do not certify that decision. ## Know where data can go Local execution and local model inference are different. An agent running on your computer can send prompts and material it reads to its provider. Cloud and browser paths have their own data handling. Review the connected account, provider terms and permitted files before using sensitive material. CoLateral canvas content saves locally in its application or browser profile. Some signed-in records, including saved snapshots and command-bar conversations, can also be mirrored to cloud storage. Those selective copies do not back up every project file, canvas or setting. See the [Privacy Policy](/privacy) and keep independent backups of valuable work. CoLateral supplies no model credits. Supported agents use your own provider accounts, with their own setup, permissions, usage limits and any charges. ## Where CoLateral fits CoLateral puts notes, sheets, documents, browser cards, terminals, tools and supported agents on a visual project canvas. Keep the source beside the task, work manually when that is simpler and connect an agent when its capabilities fit. The [first-project guide](/resources/first-project-workspace) starts with a small note you can save and reopen without an AI account. For a document workflow, the [drawing intake](/tools/project-intake-drawings) and [PDF Sheet Indexer](/tools/pdf-sheet-indexer) pages explain two optional tools and their scope. Their output still needs to be compared with the source. Check [setup answers](/faq) before choosing an agent connection. ## Common questions ### How is an AI agent different from a chat assistant? The available actions and access differ. An agent may read a working folder, run commands or write files, depending on its connection and permissions. Check the selected path and the source it actually used. ### Can I trust an agent to calculate a result? Ask for reproducible arithmetic in an executed formula, sheet or script. Check the inputs, units, method and a known example. A correct run does not establish that the method is appropriate for the task. ### Do my files stay on my machine? Canvas content saves locally, but connected agents can send prompts and material they read to their providers. Some signed-in records can also be mirrored to cloud storage. Check the selected connection and Privacy Policy before using sensitive material. ### Does every action require my approval? Permissions and confirmation behavior depend on the agent and action. Review the selected controls, set the task scope, inspect actual changes and keep recoverable copies of valuable work. ## Tools mentioned here These file-workflow tools are available inside the CoLateral desktop app. [Tool Drawing Intake (AI) Upload drawings, review AI-suggested walls, openings, lintels, and project info with evidence, then generate a draft schematic layout. See the tool](/tools/project-intake-drawings) [Tool PDF Sheet Indexer Ingests PDF drawing sets and builds a structured sheet index with revision history and delta-sheet flags. See the tool](/tools/pdf-sheet-indexer) ## Keep reading [Getting started Keep your first project in CoLateral Create a small project brief, check a local save and reopen it. Learn where supported agents, files and tools fit when you are ready. Read](/resources/first-project-workspace) ## Bring your next project together. Bring your files, notes, tools and supported agents together in CoLateral. Start with an invented example, choose your connection and inspect the result. [Get the desktop app](/download) [What a license costs](/pricing)