Eleven AI terms that decide whether you get anything for your money

Last year a client sent me an invoice. The line item read: „AI agent for marketing – 4,800 euros, implementation“.

When he showed it to me, I asked what the agent actually does. He said that every Monday it generates three Facebook posts and sends them to his scheduler. Nice. But that is not an agent. That is a macro with one API call.

He paid 4,800 euros for something an intern would build for him in Zapier for a fraction of that price.

I am not writing this to belittle him. I am writing it because it keeps happening, and always for the same reason: nobody explained to him the difference between what he ordered and what he was sold.

These are 11 AI terms that explain that difference. It is not a glossary for developers. It is a glossary for the people who sign the invoices.

When is AI just an expensive chatbot, and when does it actually work?

The difference is whether the model works with your data, or only with what it remembers from training.

When someone says „we have deployed AI“, they usually mean one of three scenarios. And that difference is the whole difference between those 4,800 euros and something that genuinely saves work.

A simple chatbot with no access to current company data answers mainly from the model’s learned knowledge and the context of the conversation. On prices, deadlines or products it can therefore be wrong – asked „how much is shipping to you“, it will answer confidently and incorrectly.

An assistant that looks things up in your documents before it answers. That is called RAG, and for a company it is usually the right choice.

A model fine-tuned on your data. That is called fine-tuning, and in most cases where a company wants it, it is not what the company needs. I will explain why further down.

LLM: what it actually is, and why it is not ChatGPT

An LLM is a technology. ChatGPT is a product. It is the difference between a combustion engine and a specific car.

A language model is a neural network that learned to predict the next piece of text from an enormous amount of text. Out of that simple mechanism comes the ability to read, write, summarise and sort language.

The architecture behind most of today’s dominant LLMs comes from Attention Is All You Need (Vaswani et al., Google, June 2017), which introduced the transformer. Other architectures exist, but the transformer prevails today. The first model OpenAI called GPT came out on 11 June 2018. It only got the label „GPT-1“ retroactively.

The authorship of the phrase „large language model“ itself cannot be attributed to one person or one paper. The period from 2018 to 2020 helped popularise the term, and the word „large“ moved to the foreground with GPT-3, which had 175 billion parameters and was trained on 300 billion tokens.

Why this matters to you: when someone tells you „we use GPT-4“ or „we have our own LLM“, ask about the model, not the brand. There are dozens of models, and the price differences are a hundredfold.

Prompt: why is „write me an email“ a weak brief?

The prompt is the entire input the model receives: role, context, examples, output format.

That 4,800-euro order had one problem in it that this term describes precisely. The owner wrote in the brief: „write posts about our products“. That is a weak brief. The model then generates output according to a probability distribution, and you get one of many possible answers, not the one you needed.

„Write me an email“ is a simple brief that is also a prompt, just a very weak one. The difference is in the quality of the brief, not in the definition. A prompt is: „You are the sales director of a Slovak manufacturing company. Write an email to a client whose order is two weeks late. Tone: matter-of-fact, no more than one apology. Length: 120 words. End with a concrete proposal for a new delivery date.“

The same model, two completely different outputs.

The word „prompt“ was used in computing long before AI, at the command line. As a named discipline in the LLM context, „prompt engineering“ appears on 15 February 2021 in a paper by Reynolds and McDonell. It is a significant early contribution, not a proven origin of the field. A systematic review of the field came out in June 2024 (The Prompt Report, arXiv 2406.06608).

The most common mistake: the belief that a longer prompt is better. In practice it gets killed by too much context without structure. It relates to a term coming a few lines below.

RAG: and why does it not eliminate hallucinations?

RAG means the model finds relevant passages in your documents before it answers and combines them with what it already knows. It is not answering purely from documents.

The full name is retrieval-augmented generation, and it comes from a paper by Lewis et al. at Facebook AI Research, arXiv 2005.11401, submitted on 22 May 2020. The authors described it as giving the model a second, untrained memory – an index of documents – and teaching it to draw from that.

The difference in practice: a company assistant that can show, for every sentence, the document and the page it came from. Versus a model that „remembers“ it from training and never says where from.

And now the part you will not read in marketing materials: RAG does not eliminate hallucinations. It reduces how often they occur, but the model can hallucinate around the retrieved source. The original paper does not put it that way, and later research (Ji et al., 2022) confirms it: RAG does not „consistently reduce“ the hallucination problem.

Why it matters: whoever believes RAG has removed hallucinations stops checking the output. And that is exactly the moment a non-existent price ends up in an email to a client.

Embedding: how does a machine measure whether two things mean the same?

An embedding is a translation of text into a list of numbers, such that things similar in meaning end up numerically close to each other.

It is how a machine „measures“ similarity of meaning without understanding the words. „Advertising“, „advert“ and „campaign“ sit close together in that space. „Advertising“ and „bread roll“ sit far apart.

The idea is old. John Rupert Firth wrote in 1957 that a word is characterised by the company it keeps. The mathematical basis dates from the late 1980s. A major milestone that brought embeddings into the form today’s AI knows them in was word2vec (Mikolov et al. at Google, arXiv 1301.3781, 16 January 2013).

A company use case: have 10,000 complaints clustered by meaning and you find that 3,400 of them say the same thing in different words. Manual sorting will not reveal that.

The mistake I see: the assumption that close embeddings mean correctness. They do not. Two articles carrying the same disinformation will have nearly identical embeddings.

Context window: why „we have a million tokens“ is not enough

The context window is the model’s working memory. Whatever does not fit into it at that moment, the model does not use then. It can, however, be stored outside the context and retrieved by a tool or through RAG.

Measured in tokens, it includes the brief, attached documents, the entire conversation history and the answer itself.

The numbers that hold today (verified 25 October 2026): Gemini 4 Argon has a shared context window of 1,048,576 tokens for input, processing and output. Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 have a million tokens of context and 128,000 maximum output. For comparison: GPT-3 had two thousand tokens in 2020, GPT-4 eight thousand in 2023.

And here is the thing competing articles leave out: context rot. Accuracy and the ability to recall a piece of information decline with input length, before the limit is even reached. Chroma Research confirmed it on eighteen frontier models in 2025. Anthropic names it directly in its own documentation.

So „we have a million-token context, we will put the whole company in there“ only partly works: a big window is an advantage, but accuracy declines with input length, and you pay for every token in the window on every call. Prompt caching changes how much you pay for those tokens, not whether they count.

Hallucination: and why the term is not from the AI era

A hallucination is an answer that contains untrue information presented as fact, fluently, confidently and in good grammar.

The model is not broken. It is doing exactly what it was trained to do, just without verification against reality.

An oddity that does not show up in AI articles: the term is older than the LLM. Wikipedia dates it to computer vision. Eric Mjolsness introduced it, as a way of adding detail to an image, in his doctoral thesis in 1986. The „face hallucination“ algorithm was published by Baker and Kanade in 1999. For LLMs it was formalised by the survey by Ji et al. (arXiv 2202.03629, ACM Computing Surveys, 2022).

In marketing it looks like this: the model generates the statistic „67% of marketers already use AI“, with persuasive phrasing and no source. You copy it into a deck for a client.

The mistake that needs naming: a hallucination is not deception. The model has no intent. And a hallucination is not creativity either. When it invents a new slogan, that is work. When it invents a new law, that is a hallucination. The difference is whether the output claims something about the world.

Fine-tuning: the most expensive mix-up in the field?

Fine-tuning adjusts the model’s behaviour and abilities by training it on further data. RAG supplies it with external information at the moment of answering. If you need to work with current prices, products or documents, RAG is usually the more practical choice. If you need the model adapted to a specific task or way of working, fine-tuning can make sense. Whoever confuses the two pays twice.

Fine-tuning is the process by which an already trained model is taught further on a smaller dataset so that it adapts to one task, tone or format. Its weights change and the model genuinely remembers the new behaviour.

For language models, ULMFiT established it as a standalone method (Howard and Ruder, arXiv 1801.06146, 18 January 2018). GPT-1 used the same recipe four months later. The efficient LoRA variant cuts the number of trainable parameters from billions to millions.

An example that makes it clear: you fine-tune the model on 500 support replies the company approved, so that it writes in exactly its tone. Brief, without unnecessary apologies, always with a link to the complaints form. In this example the goal is to teach the model a way of answering, not to keep a current product catalogue inside it.

And this is exactly where the most expensive mistake in the field happens. „We will fine-tune the model on your documents so it knows about your products“ sounds good and is a bad brief. Facts change every month. The model will not remember them reliably or accurately. Three months later it will be selling a product that no longer exists, or a price that no longer applies.

The split is simple, and IBM and Databricks phrase it the same way: fine-tuning updates the model itself, RAG enriches a finished model with external data at the moment of the question.

Agent: and why do companies buy an „agent“ that is just a macro?

An agent can dynamically decide on further steps and choose tools according to the goal. A classic workflow executes a pre-designed procedure. Neither is automatically better. For repeated, predictable tasks, simple automation can be cheaper, faster and more reliable.

And that is exactly the term that cost that client 4,800 euros.

The term „agent“ is old in computer science. But „AI agent“ in today’s sense is a matter of 2022 to 2024. The technical basis of the „think and act“ pattern was laid down by ReAct (Yao et al., arXiv 2210.03629, 6 October 2022).

In its architecture guide of 19 December 2024, Anthropic introduces a distinction that anyone buying an „AI agent“ today should know: it separates workflow from agent. A workflow is a predefined sequence of steps. An agent is the one that searches for the procedure.

Why it matters: a chatbot gets a question and answers it. An agent gets „prepare content for next week“. It looks at the analytics, picks three topics, writes the copy, generates the visuals, sets up the posts in the scheduler and returns a report. Nobody told it in what order to do that.

The difference from classic automation (Make, Zapier): there the procedure is hard-coded and can fail for several reasons. The input changes, an API drops out, authorisation fails, or the data changes. An agent searches for the procedure, which is why it can also fail by finding the wrong one. That is why Anthropic explicitly recommends an iteration limit as a safety net.

If someone sells you an „agent“ and it is in fact a hard-coded procedure with one API call, it will work until the first situation nobody programmed.

Multimodal model: why it is not „it can generate an image“

Multimodality means the ability of AI to work with several types of data, for example text, image or sound. It can include understanding as well as generation. The fact that a model can create an image from text does not automatically mean that it can also analyse images reliably.

A model that turns text into an image may not be able to read a chart in a screenshot.

One of the major milestones was CLIP, introduced in 2021 (Radford et al., arXiv 2103.00020). It learned the relationships between images and text on roughly 400 million image-text pairs. Multimodality reached products with GPT-4V (2023) and Gemini (2023 to 2024).

A company example: photograph an invoice at the counter, the model reads it and pulls the amount, the supplier and the due date into a spreadsheet.

Two mistakes. The first: the assumption that a model is equally good in all modalities. The CLIP paper itself admits it struggles with abstract tasks such as counting objects in an image. The second, the most expensive one in practice: that it will reliably read complex tables and forms in PDFs. That is still a separate technical discipline, and most companies underestimate it.

Token: and why is Slovak more expensive?

A token is the smallest unit of text the model works with, and you pay by the number of tokens.

The model does not see letters or words. It sees tokens, usually parts of words or punctuation. The word „tokenization“ splits into „token“ and „ization“. A common word like „the“ is one token.

This mechanism was established by the paper on BPE (Sennrich, Haddow and Birch, arXiv 1508.07909, 31 August 2015), which solved the problem of rare words in neural translation by breaking words down into subword units.

And here is the thing no Slovak AI glossary mentions: tokenisation varies significantly between languages. In some language combinations the difference was as much as fifteenfold. Petrov et al. (Language Model Tokenizers Introduce Unfairness Between Languages, 2023) measured it across seventeen tokenizers, including those deliberately trained as multilingual.

Even with Slovak it is therefore worth tracking real token consumption, especially in automations that process large amounts of text. At large volumes it is a direct cost line, not a curiosity.

Inference: where does the money actually go?

Training is paid once, inference on every use.

Inference is the phase in which an already trained model processes new input and generates output. The opposite of training: in training the model learns, in inference it is used.

An example that holds it together: training is building the production line. Inference is every piece that goes through it.

Prices per million tokens (25 October 2026): Claude Fable 5.1 at $10 for input and $50 for output. Opus 5.5 at $4 and $20. Sonnet 5.5 at $2 and $10. Haiku 5.5 from 10 cents for input and 50 cents for output. These are API prices, not a ChatGPT, Gemini or Claude subscription. The differences between models are large, and it is a decision made on every call. That does not mean that every call with the most expensive model costs a hundred times more. It depends on the number of tokens, caching, tools and the nature of the task.

A mistake in Slovak: confusion with statistical inference, that is, estimating the properties of a population from a sample. They are two completely different terms, and they get mixed up in Slovak because both are translated the same way.

The second mistake: „the bigger the model, the better“, regardless of price. Most of the ongoing cost of an AI product is inference, not training. That is why in practice a smaller model with a good prompt and RAG is often chosen over the biggest model for everything.

So what should you take from this?

Three decisions that actually follow from those terms:

First: work with current facts, not with the model’s outdated memory. If you need current prices, products or company documents, in most cases start with RAG or a direct connection to your company systems. Use fine-tuning where you need the model adapted to specific tasks, behaviour or output format. The two approaches can also be combined.

Second: a smaller model with good context beats the biggest model with chaos. Inference cost is what you pay every month. And the context window is not a warehouse, it is a desk, and you can get lost at a desk too.

Third: ask about the mechanism, not the name. „We have an AI agent“ means nothing. The question is whether it searches for a procedure or executes a fixed script. „We fine-tuned the model“ also means nothing. The question is whether those facts have somewhere to come from at the moment of the question.

When we went through that invoice with the client, at the end he said one thing that annoyed me: „Well, it sounded good.“

It did. And that is the whole problem. Those terms are not used to explain things. They are used to sound good, and to make it impossible to ask what exactly you get for the money.

Over the years in business I have learned one thing about technology: the ones that actually work can be explained without you having to know how they are built. When someone cannot explain what their system does differently from a cheaper version, it is usually because it does nothing differently.

Frequently asked questions

Do I need my own LLM?
Almost never. Owning your own model makes sense at volumes where it pays off financially, or when you have specific requirements about where the data sits. Sensitive data can also be processed through suitably secured external models. For most companies it is cheaper and better to use an existing model with RAG over their own data.

Is RAG the same as a vector database?
No. Vector search is only one part of it. In practice it is combined with full-text search and reranking of results.

Why does the model answer differently than last time to the same question?
Because generation is not deterministic. The model picks from a probability distribution, and temperature settings influence that. That is why a one-off test means nothing. My practical advice: repeat the measurement at least three times so that you see the spread, not chance. Incognito mode alone does not guarantee a completely neutral result either.

When is AI entirely pointless?
When the problem is in the process, not in the language. If a company does not know its prices, AI will not find them out. If it has a mess in its data, the model inherits that mess.

How much will it cost me?
It depends on volume, not on technology. At small volumes it is a few euros; at tens of thousands of questions a month it is a separate line in the budget, and that is when it pays off to look at the price per token rather than the vendor’s marketing material. The specific sum depends on the model, the input length, the output and the tools used.

Have you come across something similar? Did you order „AI“ and get a macro? Write to me, I would like to collect a few of those stories.

Sources

Links to the original papers and documentation the article draws on:

• Lewis et al. – Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (RAG), 2020
• Anthropic – Building Effective Agents, 19 December 2024
• Chroma Research – Context Rot, 2025
• Anthropic – models and pricing (as of 25 October 2026)
• Google AI for Developers – Gemini models and context window
• Vaswani et al. – Attention Is All You Need (transformer), 2017
• Radford et al. – Learning Transferable Visual Models From Natural Language Supervision (CLIP), 2021
• Ji et al. – Survey of Hallucination in Natural Language Generation, 2022
• Yao et al. – ReAct: Synergizing Reasoning and Acting in Language Models, 2022
• Howard and Ruder – Universal Language Model Fine-tuning (ULMFiT), 2018
• Mikolov et al. – word2vec, 2013
• Sennrich, Haddow and Birch – Neural Machine Translation of Rare Words with Subword Units (BPE), 2015
• Petrov et al. – Language Model Tokenizers Introduce Unfairness Between Languages, 2023
• Reynolds and McDonell – Prompt Programming for Large Language Models, 2021
• Schulhoff et al. – The Prompt Report, 2024

Model prices and context windows are verified as of 25 October 2026. This section changes with every generation of models, so check it again the next time you use this article.