A habit has quickly developed in the office since the introduction of AI assistants: people spend time polishing command sentences. Adding the word "professional," assigning the role "you are a senior analyst," and extending it to a paragraph. However, what usually determines the quality of the answer is not the sentence itself, but the materials attached alongside it.
The difference is easy to test yourself. Request a summary of quarterly performance without attaching anything, and the result will sound reasonable but entirely fabricated. Attach the sales file, and the result becomes verifiable.
Why models fabricate when the materials are empty
Language models do not have an internal mechanism to verify facts. They construct answers from patterns, not from checking sources. When materials are not provided, those patterns are the only ones available, and the results can sound convincing yet be incorrect. This is termed hallucination.
The approach that has now become standard to mitigate this is retrieval augmented generation, commonly abbreviated as RAG. The principle is simple: before answering, the system first retrieves relevant snippets from the documents we provide, then constructs the answer based on those snippets. The School of Computer Science Binus notes that its main advantage is not only accuracy but also traceability. The generated answers carry the source documents, allowing users to verify them.
The scale of improvement is significant. A study at Universitas Multimedia Nusantara implementing RAG based on a vector database reported an accuracy rate of 86.84 percent.
For daily work, what needs to be taken from this is not the technical terminology, but the logic: answers based on materials can be traced, while answers based on the model's memory cannot.
What constitutes good materials
It is not just about attaching files. Several factors usually determine the results and are often overlooked.
Clarity of version. If a draft is attached, the answer will reflect that draft. Specify the version, or do not attach it at all.
Definition of columns. A column named "status" could mean payment status, shipping status, or account activity status. The model will guess, and while its guesses are not always wrong, they cannot be held accountable.
Explicit time range. Data from three years summarized as "latest trends" will yield trends based on the model's version, not yours.
And what is often forgotten: also include materials that contradict. If only a good monthly report is attached, the conclusion will be positive.
Honest limits
Attaching materials does not automatically make the answers correct. Incorrect materials still lead to incorrect conclusions. The difference is that the errors can now be traced back to the specific line, rather than evaporating as a well-written sentence.
This is also the reason why the skills required are shifting. The Ministry of Communication and Digital, after an evaluation by Bappenas, has shifted the focus of its digital literacy program from device introduction to upskilling and AI skills. The PwC Global Workforce Hopes and Fears 2025 survey shows why this is urgent: in Indonesia, 69 percent of workers have used AI in the past year, but only 16 percent use it daily, and daily users report significantly higher productivity increases, 96 percent compared to 75 percent.
The gap between merely trying and truly utilizing largely lies in the habit of preparing materials.
Sources
- School of Computer Science Binus – Retrieval-augmented Generation Sebagai Solusi untuk Masalah Halusinasi dalam AI Generatif, October 9, 2025 → socs.binus.ac.id
- Universitas Multimedia Nusantara – Implementasi Teknik Retrieval Augmented Generation untuk Menghilangkan Halusinasi pada Large Language Model Berbasis Vector Database → kc.umn.ac.id
- PwC Indonesia – AI adoption is boosting productivity, particularly among Indonesia's Gen Z, press release February 23, 2026 → pwc.com/id
- Warta Ekonomi – Komdigi Geser Fokus Literasi Digital ke Upskilling dan AI → wartaekonomi.co.id