New AI Format Keeps Your Work's Context, Boosts Accuracy to 96%
Your work documents lose the 'why' – this AI fixes that and makes answers far more reliable.
When you finish a report or analysis, the final document often leaves out the research, calculations, and assumptions behind it. That context is lost, making it hard for others (or AI) to understand or build on your work. A new study introduces the Portable Agent Format (PAF), which bundles the work product with its compiled context, evidence, and rules for interaction. Think of it as a zip file that includes not just the final report but also the notes, sources, and guidelines for how to use it.
To test if this context is useful beyond the original process, researchers created four new PAF packages: an investment proposal, fund research, financial-statement analysis, and an engineering incident review. Two were made by specialized AI agents and two by a general-purpose assistant. They then asked 400 questions and evaluated 12,000 responses. A hosted AI reader using PAF answered 96% of questions correctly, compared to 86% with standard instructions plus raw files. On answerable questions, it scored 94% versus 79%. The gap varied by package.
A smaller offline AI reader (0.8 billion parameters) achieved 69% overall but only 54% on answerable questions. Its error rate was 2.9% (materially wrong or unauthorized answers), much better than 56% with source and instructions alone. However, offline retrieval failed completely on 40 historical-record questions in one financial package. The study suggests that compiled knowledge should be part of the delivered work product, but independent standards and end-to-end governance are still unproven.
For everyday professionals, this means less time answering follow-up questions and more confidence in AI-assisted work. But don't expect perfect offline performance yet – the technology is still evolving.
- PAF packages your work with its context, so AI can answer questions accurately without extra explanation.
- In tests, a hosted AI with PAF got 96% of questions right, versus 86% with standard methods.
- Offline use is less reliable, with only 69% accuracy and some failures on historical data.
Why It Matters
This could save you hours answering questions about your work and reduce misunderstandings.