# More documents don't mean better answers. Unnecessary ones make them worse.

Jaro Rais · 6 min read · 2026-10-21 · Working with AI
Source: https://itashu.co/en/know-how/more-documents-do-not-mean-better-answers

You uploaded every policy, price list and set of meeting notes into AI. Yet the answers keep getting less accurate. That is no accident. We explain four reasons and how to fix them.

![A desk buried in papers where one important sheet gets lost, next to a desk with three neat sheets](https://itashu.co/img/kh-02-1280.webp)

## What belongs on the desk and what doesn't

Everything AI sees at a given moment is called context: your instructions, uploaded documents, the whole conversation so far and the results of tools. Today's models handle huge context, up to a million tokens, roughly 750,000 words. Companies draw a simple conclusion: the more we give it, the better.

A July 2025 study by Chroma tested 18 leading models. **All of them became less reliable as context grew.** Researchers call this “context rot”.

- **With a person:** A colleague has three hundred papers on their desk. They physically fit, but the one important sentence gets missed.
- **With AI:** Context is the desk. The model's attention is limited and is split across everything on the desk.

Working with context is exactly this decision. The technical term is [context engineering](https://itashu.co/en/know-how/ai-does-not-need-a-better-model#kapitola-02). For you as a user it means four mistakes you can avoid.

## 01 Poisoning: a wrong conclusion gets recorded as fact.

AI misreads something, notes down a wrong conclusion and from then on treats it as truth. To AI, its own notes look just as trustworthy as your documents.

- **With a person:** A wrong decision makes it into the meeting notes. For a month everyone works by it, because “it's in the notes”.
- **With AI:** A wrong note in the conversation affects every answer that follows.

**Fix: verify before it gets recorded.** Have important conclusions confirmed from the original document, not from AI's answer.

![Four fixes: verify, summarize, select and delete](https://itashu.co/img/art-02-fixes-1280.webp)

## 02 Distraction: a long history starts to drive the work.

A language model essentially continues the pattern it sees. When a conversation is very long, the strongest pattern becomes what is already in it. AI then repeats old approaches instead of solving what is needed now.

**Fix: summarize and start over.** When a conversation gets long, have AI summarize where things stand and continue in a new conversation. Keep in mind that every summary loses something, like a copy of a copy. Important things belong in permanent documents, not in the conversation history.

## 03 Confusion: too many tools and documents at once.

The more tools and documents AI has available, the harder it is to pick the right one. In one test, a model failed a task with 46 tools but completed the same task with 19.

- **With a person:** You don't give a new colleague access to all 30 of the company's systems. You give them the three they need for their work.
- **With AI:** The same applies. Invoicing doesn't need access to the calendar, marketing and the warehouse.

**Fix: choose only what the task needs.** If something more is needed, you can add it.

## 04 Clash: the old and the new version sit side by side.

Your documents include last year's price list and the current one. Or AI realizes it was wrong, but the original wrong note stays in the conversation. The model has two conflicting truths in front of it and sometimes picks the wrong one.

**Fix: delete the old.** Remove [outdated documents](https://itashu.co/en/know-how/language-model-frozen-in-time) from the project. Adding the new version is not enough.

> **In practice:** “An accounting firm uploads four years of policies into one Claude project so that AI has ‘everything’. Asked about the VAT rate, AI answers once by the old policy and once by the new one. The solution isn't a better model. The solution is keeping only valid documents in the project.”

## Four mistakes, four fixes: less on the desk, better answers.

| MISTAKE | WHAT HAPPENS | FIX |
| --- | --- | --- |
| Poisoning | a wrong conclusion is taken as fact | verify from the original |
| Distraction | a long history drives the work | summarize and start a new conversation |
| Confusion | too many tools and documents | choose only what's needed |
| Clash | old and new version at once | delete the old |

**Each of these fixes takes something away.** AI works better when it has less in front of it.

## What this means for setting up AI in your company

- One project per area of work, not one project for the whole company.
- Only valid documents in a project. Someone has to be responsible for keeping them current.
- Short, clear instructions. Long instructions aren't more thorough, they just overload.
- End a long conversation with a summary and continue in a new one.

**Ask yourself:** Who in your company decides what AI sees, and who deletes what is no longer valid?

---

Tell us where AI gives you inaccurate answers. We'll show you what to clear off its desk. https://itashu.co/en/consultation