Why we started building acrylic

You ask AI for a draft, copy it into a document, fix the formatting, and check for errors. If the work draws on several files, you switch between windows, compare the figures, and update the document by hand. AI’s answers have improved, but turning them into a finished document still takes work.

Something different is happening in software development. A developer describes a change, and AI finds the relevant files, edits the code, and runs the tests. The developer reviews the changes. People are starting to delegate tasks to AI, not just ask it questions.

Could document work follow a similar path? What would it take to let AI work directly on a report, spreadsheet, or presentation?

We saw a need for an environment where AI could understand a document’s structure, make changes, and let people review the result. That’s where acrylic began.

What it takes for AI to edit documents

Editing a document is more complicated than generating text. Reports contain paragraphs, tables, styles, headers, and footnotes. In a spreadsheet, formulas link values across cells. In a presentation, the position and order of text, images, and shapes affect how the message comes across. Those structures need to stay intact when the content changes.

Consider what happens when quarterly results change. Updating a number in a spreadsheet is only part of the job. The report’s tables, narrative, and conclusion may also need revision. If the same figures appear in a presentation, that file needs updating too. Correcting a number can break a formula. The content can be right while the formatting or layout is disrupted.

Reading the file as plain text isn’t enough. AI needs to understand where each element belongs and how it connects to the rest of the document. It needs to change the relevant parts, preserve everything else, and save a file that people can open and keep editing.

That’s why we’re building both the engines that read and save documents and the editors people use to work on them.

Sensitive documents are harder to use with AI

Technical complexity is only part of the challenge. The documents where AI could save the most time often contain information people can’t readily share with an external service.

Contracts contain customer details and commercial terms. Financial reports include figures that haven’t been made public. HR documents hold personal information, and internal meeting notes may include plans that haven’t been finalized. These documents take time to read and review. They often follow similar formats and require careful comparison, tasks where AI could help save time.

But before using AI, people have other questions to answer. Can I upload this file to an external service? Where will the information be processed? Would that comply with my organization’s security policy?

For sensitive documents, the concern about sending information outside the organization can outweigh the convenience. People end up using AI for material they can share, while important, time-consuming work remains difficult to hand over.

We don’t see security as an enterprise feature to add later. If AI is going to be useful for everyday document work, security needs to be part of the design from the start.

Why we’re addressing document editing and data control together

Accurate editing isn’t enough if people can’t trust the product with important files. Keeping data on a computer isn’t enough either if AI can’t edit the document: someone still has to copy its response into the file and fix the formatting. Both capabilities are necessary for AI to change how document work gets done.

We’re building the document editor and the environment where AI runs together. In acrylic, you can open HWPX, DOCX, XLSX, and PPTX files and continue the work you’ve already started. The AI agent finds the files it needs, compares their contents, and applies changes to the documents.

For example, you might ask, “Find the figures that changed in this spreadsheet and update the quarterly report.” acrylic compares the two files, identifies the tables and sentences that need updating, and applies the changes to the report.

The file is saved in its original format. You can review the changes and keep editing in the same workspace. Even when AI carries out the task, you make the final decisions.

You should be able to choose where AI runs

Different tasks call for different ways of using AI. A cloud model may be a good fit for work that requires more capable models or complex reasoning. For documents containing personal or internal information, keeping the file and the request on your computer may matter more.

We’re building acrylic so people can get started with cloud AI and choose local models when their computer can support them. In local mode, documents and requests are processed on your computer rather than sent to an external AI service.

Local AI won’t be the best choice for every task. The models you can run and their performance depend on your computer. A cloud model may be better suited to complex work. People should be able to choose based on the task and the information involved.

Work that uses information you can share with an external service can be handled in the cloud. Sensitive documents can be processed locally. Organizations with specific security requirements should be able to run AI in an environment they manage themselves. We believe that control over where data is processed will make it possible to use AI for more important work.

People need to keep using their existing files

Much of our work ends up in reports, spreadsheets, and presentations. Research and analysis go into reports. Figures are organized in spreadsheets. Plans become presentations, and decisions are recorded in contracts, policies, and internal documents.

Every organization has its own formats and ways of working, but the practical requirements are similar. People need to keep using existing files and share them with others. After AI makes changes, those files still need to open and remain editable in the original application.

If a new AI tool requires people to convert years of documents or learn a new way of working, it can create more work. We start with the files people already use.

We’re building one workspace for DOCX reports, XLSX spreadsheets, PPTX presentations, and HWPX documents. AI needs to understand the structure of each format and save files that people can continue to use.

We want people to pick up their existing work and finish it with less manual effort, with AI helping them along the way.

Can we delegate document work to AI?

Writing code and working on documents are different tasks. Code follows defined syntax, and tests can check the results. Documents vary widely in purpose and format. Two organizations may write the same kind of report differently. Even a factually correct report may need revision if its tone or table layout doesn’t meet the organization’s needs. Bringing the coding-agent approach to document work means accounting for those differences.

Still, software development offers a useful example. A person describes the outcome, AI finds the relevant files and carries out the task, and the person reviews the changes and makes the important decisions. We believe document work can follow a similar pattern.

AI should be able to work on the files themselves: create a document, revise an existing one, compare several sources, and save a result that people can keep editing.

We call this an Agentic Office Suite: an environment where AI carries out document tasks and people review the results.

The change we want to make

We’re still building acrylic. We need to read and save different document formats more accurately and work reliably across files whose contents depend on one another. We also need to improve how people review AI’s changes to help prevent errors in the content and damage to the document’s structure.

For local models, we need to improve performance and processing speed, and establish which computers can run them well. There’s plenty left to do, but we know what we want to make possible.

AI should be able to edit real documents, with people reviewing the results and continuing the work. People should also be able to use AI with sensitive documents in an environment suited to the information they contain.

AI agents are changing how people write software. We believe they can change how people work with documents too. That’s why we started building acrylic.