AI and PDF: what comes next

Every document tool is adding AI right now, and most of it is decoration. This post sets out what we think AI genuinely improves in document work, what it does not, and the rules we are holding ourselves to as we build the features on the PDFCraft roadmap.

Why this matters

Documents are where organisations store their obligations. A lease, a policy schedule, a set of tender conditions — these are long, dense and consequential. Reading them properly takes time most people do not have, so they skim, and skimming is where mistakes happen.

That is a real problem worth solving. Rotating a scanned page is not. Knowing the difference is the whole point.

Where AI genuinely helps

Question answering

The obvious case: open a 90-page contract and ask what the termination notice period is. A good answer includes the page reference so you can verify it yourself in ten seconds rather than reading for an hour.

Summarisation with structure

Less obvious, and arguably more useful: generating an outline of a long document with page links. Long PDFs are hard to navigate, and a reliable table of contents solves a boring problem that affects every single reader.

Translation of a page, not a paragraph

Translating an official letter is common and stressful. The value is in keeping the layout intact so the translated version still reads like the document it came from.

Extraction into a table

Pulling dates, amounts and party names out of a stack of invoices into rows is exactly the kind of tedious, error-prone work a model handles well — provided every extracted value can be traced back to the page it came from.

Where AI does not help

• Merging, splitting and rotating pages. These are deterministic operations that already take one click in the toolkit.
• Compression. This is arithmetic, not judgement.
• Placing a stamp or signature. You know where it goes; a model guessing is worse than you deciding.
• Anything where a wrong answer looks exactly like a right one and nobody will check.

The privacy problem, stated honestly

Our current tools are local. Your file is opened by your browser, changed in memory and written back to your device. That is a strong guarantee precisely because there is nothing to promise — no server ever sees the document.

Large language models are different. Running a capable model usually means sending text somewhere. We are not going to pretend otherwise, and we are not going to quietly change the privacy characteristics of the site because a feature is fashionable.

The rules we are committing to

1. Local tools stay local. Nothing that works in your browser today will start uploading tomorrow.
2. Anything that needs a model will say so on the page, before it runs, in plain language.
3. AI features are opt-in. You will never trigger one by clicking a familiar button.
4. Answers cite pages. If a feature makes a claim about your document, it will point at where it found it.
5. No silent scope creep. If what we send changes, the page text changes with it.

Common use cases we hear about

Common mistakes with AI document tools

• Pasting confidential text into a general chatbot without checking the retention terms.
• Accepting a summary as a substitute for reading a clause you are about to be bound by.
• Using extraction on low-quality scans without spot-checking the output.
• Assuming a translated document is legally equivalent to the original.

What you can do today

While the AI features are still in development, the mechanical work is already handled. Prepare documents properly and they will be far easier to work with later: merge related files, remove blank pages, fix sideways scans, and keep filenames consistent. Our document filing guide covers a system that takes minutes to adopt.

Summary

AI belongs in document work where reading is the bottleneck — questions, outlines, translation and extraction — and nowhere else. We will ship those features when they are accurate enough to be trusted, label them clearly, keep them optional, and leave the local tools exactly as local as they are today.