Insights / AI
How AI actually works (and forgets)
By Bakhrom Rikhsiboev, founder of First Door Studio. Last reviewed November 2026.
When a chatbot answers you, it is not looking up a fact in a single official book, and it is not "remembering" you the way a person would. It is a large language model: a system trained to predict the next word, many times in a row, based on patterns in the text it has seen and the text you just gave it.
That is enough to draft an email, summarize a page or answer a common question. It is also why these tools forget, guess and sometimes invent. If you plan to use one in a business, those limits matter as much as the impressive parts.
The context window: a short desk, not a long memory
A model can only "see" a limited amount of text at once. That limit is called the context window. It includes the hidden instructions (the system prompt), the current conversation, and any documents you pasted or the system retrieved. Once the window is full, older turns get dropped or compressed. The tool is not quietly storing your last month of chats unless you have built a separate memory or a database for that.
This is why an assistant can sound sharp on question one and lost on question twelve, or why it ignores a policy you mentioned at the start of a long thread. It didn't betray you. It ran out of desk space.
The fix is not "buy the biggest model and hope." It's to keep prompts short and current, and to fetch the right document at the right time instead of pasting your entire drive into the chat. That pattern is called retrieval-augmented generation, or RAG, and we walk through it in how companies actually build AI products.
Hallucinations: confident answers with no source
A hallucination is when the model states something that isn't in its input and isn't true, in a tone that sounds sure. It might invent a refund limit, a legal citation, a "study," or a feature your product doesn't have. It does this because it is guessing a plausible next sentence, not checking a ledger.
You reduce that risk by grounding answers in your own documents, by telling the model to say "I don't know" when the source isn't there, and by having a person review anything that could cost money, make a promise or give advice. You cannot remove the risk entirely by wording the prompt more strongly.
Guardrails: the fences around the model
Guardrails are the rules and filters around the model, not the model itself. Common ones include:
- A system prompt that sets the role, the tone and the topics it must refuse
- Input checks for prompt injection ("ignore your instructions and...") and for data you never want it to process
- Output checks that catch personal data, forbidden claims, or answers that go past your policy
- Tool permissions, so it cannot refund an order or read payroll just because someone asked
None of these are optional once an assistant talks to customers or can take an action. A demo without them is a demo, not a product. Our safe business AI assistant guide and the AI checkpoints cover what we require before something goes live.
API costs: you pay per use
When you call a model through an API, you typically pay for the amount of text that goes in and the text that comes out (tokens), sometimes plus a fee for tools or for a hosted account. A long system prompt, a large retrieved document and a rambling answer all cost money. A busy Monday can cost more than a quiet one.
Costs stay predictable when you cap usage, cache repeated questions, send easy questions to a smaller model, and log what you spent. A budget alert that only emails you is not a cap. You need a hard stop, or you can get a surprise bill.
If you're choosing between a simple website assistant with approved answers and a custom AI that reads your documents, start with how we scope custom assistants, or book a free audit and we'll tell you which is enough.