# Five AI myths that lead companies to bad decisions

Jaro Rais · 7 min read · 2026-10-23 · Working with AI
Source: https://itashu.co/en/know-how/five-ai-myths

AI makes things up all the time. You can see how it thinks. An agent can handle everything on its own. Some of these sentences used to be true, others may be true one day. Today they lead to bad decisions.

![Five cards: a crossed-out myth on top and what is true today below](https://itashu.co/img/art-04-hero-1280.webp)

## A check question: “Are you sure?” won't improve the answer

When AI answers and you ask “are you sure?”, many expect it to double-check its answer. The opposite can happen. **The model takes your question as a signal that you want a different answer, and adapts.** Even if it was right in the first place.

- **With a person:** A new colleague who changes their mind every time the boss asks “are you sure?” isn't thorough. They are anxious.
- **With AI:** If you want a check, ask for it directly: “Verify this answer in the attached document and say where it is.”

## 01 Myth: AI makes things up all the time

[Making things up](https://itashu.co/en/know-how/why-ai-makes-things-up), technically hallucination, is not solved. But today's best models are far from the fabrication machines of a few years ago. Three things helped:

- **Tools:** the model looks the answer up on the web or in a database instead of fishing it out of memory.
- **Admitting uncertainty:** models have learned to say “I can't verify this”.
- **Time to think:** models that reason before answering can check their own work.

How often models are wrong depends on the task. When summarizing a supplied text, the best models have error rates in the low single digits of percent. With factual questions answered from memory, there are far more errors. It isn't zero, but it isn't “all the time” either.

**What it means for you:** AI without access to your material will guess. AI that answers from your documents and can cite them is wrong far less often.

![A chain whose links fade: reliability drops with every step, with checkpoints between sections](https://itashu.co/img/art-04-myths-1280.webp)

## 02 Myth: you can see how AI thinks

Some models show their process step by step. It looks like a view into their head. **It isn't.** Research shows that this record may not match how the model actually reached its answer. It can be an after-the-fact justification of an answer the model is leaning toward.

**What it means for you:** The process is useful for checking the logic, not as proof of correctness. For important decisions, verify the result, not just the reasoning.

## 03 Myth: training is expensive, using AI is cheap

Headlines write about hundreds of millions spent on training models. Yet a growing share of computing power is consumed by use itself. According to estimates cited by IBM, it was about a third in 2023, roughly half in 2025, and the estimate for this year is two thirds.

The reason: models that reason generate 10 to 100 times more text per question. Most of all with agents.

**What it means for you:** You pay for use, not for training. Putting the most expensive model with long reasoning on a simple task is throwing money away. [Choose the model by the task.](https://itashu.co/en/know-how/most-expensive-model-for-everything)

## 04 Myth: a large context replaces a database

Today's models read hundreds of pages at once. It's tempting to [upload a whole CRM export, all your contracts or a year of accounting](https://itashu.co/en/know-how/more-documents-do-not-mean-better-answers) and query them like a database. It works in a demo. In everyday work it fails, for five reasons.

**Finding is not the same as combining.** A model almost always finds one specific sentence in a long text. But when it has to combine several pieces of data from different places, accuracy drops fast. In the NoLiMa test by Adobe Research (2025), at a length of roughly 50 pages, 10 of the 12 models tested fell below half of their performance on short text.

**The middle gets lost.** Models make the best use of information at the beginning and the end. What sits in the middle gets overlooked more often. The Lost in the Middle study from Stanford (2023) showed this. Research by Chroma in 2025 confirmed across 18 models that reliability drops as text gets longer.

**A database calculates, a model estimates.** A database computes total revenue, the number of clients or a date filter the same way and exactly, every time. A language model does it by reading thousands of rows and can easily skip a row or get a sum wrong.

**Every question costs the whole text.** The model processes all the inserted content again for every question. With a large export, you pay for hundreds of pages on every question and wait longer for the answer.

**An export goes stale fast.** It is a snapshot of one day. A database always has the current state and also guards access rights, meaning who may see which data.

**What it means for you:** [A large context is good for working with a few documents at once](https://itashu.co/en/know-how/language-model-frozen-in-time), such as comparing two contracts or summarizing a report. Questions about company data belong to AI connected to the system, which queries the database and explains the result to you.

> **In practice:** “A sales manager uploads a year's CRM export into AI and asks which clients from Bratislava bought in spring and haven't been in touch since summer. That requires combining three things from different places. It is more reliable to let AI query the CRM directly through a connection, not read an export.”

## 05 Myth: an agent can handle a whole process entirely on its own

An agent is a [model in a loop](https://itashu.co/en/know-how/ai-does-not-need-a-better-model#kapitola-04): it decides, takes a step, looks at the result and continues. It handles individual steps well. The problem is chaining them, because **errors multiply.**

Simple math is enough. If each step is 95% reliable, a chain of 20 steps turns out well in about 36% of cases. With 50 steps, roughly 8%. And an agent that handles an inquiry from the incoming email to the sent quote can easily take dozens of steps.

**Measurements confirm it.** In March 2025, the research organization METR found that the best models at the time completed almost all tasks that take a person under four minutes. On tasks taking more than four hours, they succeeded in under 10% of cases. According to METR, the length of tasks agents can complete doubles roughly every seven months. So it is growing fast, but a long process without checks is still a risk.

- **With a person:** You let even a capable newcomer work on their own, but you look at the result at important milestones.
- **With AI:** The agent works on its own in short stretches. Between them, a person or a second model that verifies individual steps checks its work.

**What it means for you:** A fully autonomous agent for a long process is a risk today. An agent with checkpoints is a reliable helper.

## Myths either used to be true, or will be one day

| MYTH | THE TRUTH TODAY | WHAT IT MEANS FOR A COMPANY |
| --- | --- | --- |
| AI makes things up all the time | far less with tools and documents | give it your documents |
| You can see how it thinks | you see a narrative, not the thinking | verify the result |
| Training is the expensive part | use is the expensive part | model by task |
| Large context is a database | finds one thing, can't combine | connect to systems |
| An agent handles everything alone | errors chain together | checkpoints |

Sources: IBM Technology; Adobe Research, NoLiMa (2025); Liu et al., Lost in the Middle, Stanford (2023); Chroma, Context Rot (2025); METR, Measuring AI Ability to Complete Long Tasks (2025).

**Ask yourself:** Which of these myths decides today how your company uses AI, or doesn't?

---

Tell us what you hear about AI in your company. We'll show you what of it applies to you. https://itashu.co/en/consultation