Generative AI — What It Is in Plain Words
Why this matters
- Telling apart tasks where a model composes the result from tasks where a database lookup or an ordinary bot is enough
- Understanding where the variation in answers comes from and how it is kept within bounds
- Asking a vendor on what volume and by what method quality was tested
What generative AI is in plain words
Picture an assistant who has read ten years of company correspondence. Asked to draft a reply to a customer letter, the assistant finds no ready answer in the folder and writes a new text modelled on the thousands already read.
A generative model works the same way. A search box returns something that somebody has already written. A generative model assembles new text for a specific request, fragment by fragment.
That defines where it fits. It is strong wherever the result differs slightly each time: letters, descriptions, document analysis, drafts. Where the answer is identical for everyone, an ordinary template costs less.
How it works
A model knows language and general patterns, but it does not know a particular company's price list or internal policy. The relevant fragments are inserted into the request — a technique known as retrieval from an in-house knowledge base. The answer is then built from the supplied documents and can be verified against a link to the source.
The second consequence of a probabilistic nature is invented detail. Where data runs short, the model fills the gap with plausible text. The phenomenon has its own entry: hallucination. The cure is dull: supplied sources, a constrained answer format, spot checks.
Under the hood of nearly every working system today sits a large language model. Images, speech, and code come from its relatives, built on the same principle.
An example from practice
The figures below are a planning benchmark for a mid-sized sales department, not a report on a specific rollout. The method is stated so that it can be repeated on any company's own data.
The department prepares roughly 300 sales proposals a month. The measurement: two weeks, a stopwatch on 40 requests, from the customer's letter to the file sent out. Average manual assembly takes 35 minutes, of which 25 go to the text and to picking line items from the price list.
The model receives the customer's letter, an export of the price list, and three samples of past proposals. It returns a completed draft. The account manager edits it and sends it.
The benchmark after launch is 12–15 minutes per proposal at the same quality. Verification works like this: the same 40 requests, the same stopwatch, plus the share of proposals sent out with no corrections to the figures. The order of magnitude of the sum the manual version consumes can be worked out with the process cost calculator.
What this gives a business
- Tasks with no written rules become solvable. Free-form letters and documents used to mean manual handling. Now they mean spot checks.
- The first working version arrives in weeks. A single stretch of a process goes live faster than a specification for full automation gets approved.
- Costs can be calculated in advance. The unit cost of one operation comes from the volume of the request and the answer, then multiplied by the monthly flow.
- Freed-up hours show in a metric. Savings appear only where it has been decided in advance where those hours will go: more requests per manager, shorter response times, less overtime at peak.
Case-study reviews offer a benchmark: routine work takes up to 40% of a manager's working time (shtab.app, a collection of Russian company cases). The figure is worth testing in-house — a week of calendar measurements across two employees yields a share of its own, and it is usually different.
When this can be skipped
A process with strict rules and an identical result is cheaper to close with ordinary integration. Moving data between two systems, checking an amount against a limit, sending a scheduled mailing — here a generative model adds cost and variation.
Low volume does not pay off either. A dozen documents a month are processed by hand faster than a quality check can be set up. A checklist helps assess whether a specific process is ready.
What is worth checking
First: where the rule an employee decides by actually lives. If it exists only in an experienced manager's head, it gets written down first. Standardisation comes before automation.
Second: what verification of an answer looks like. A sound system shows which document a fact came from and can answer "no data available." Without that, all one hundred percent has to be checked, and the savings disappear.
Third: ask the vendor for a measurement on the company's own sample — 30–50 real cases, the share of correct answers, the share passed to a human. A demonstration on hand-picked examples proves only that the examples were hand-picked.
Frequently asked questions
What is generative AI in plain words?
It is a model that composes new text, an image, a table, or code in response to a request. The finished answer is not stored anywhere: it is assembled for that specific question out of patterns learned from large volumes of data. That is why a single system handles both routine and unusual requests without a separate script for each case.
How does generative AI differ from ordinary automation?
Ordinary automation executes rules that have been written down: if the field is empty, reject; if the amount exceeds the limit, send for approval. A generative model works where the rules were never written out — customers phrase things differently, and documents arrive in free form. The price of that flexibility is an answer that varies slightly each time and has to be checked.
Why are the model's answers different every time?
The model picks the next fragment of text by probability rather than by a fixed rule. The same request produces answers that are close in meaning but differ in wording. For reports and figures, the fix is formatting: the model fills in predefined fields, and the calculations stay with the accounting system.
Does this require the company's own data?
For general tasks such as drafting a letter, no. As soon as knowledge of a price list, a policy, or a customer's history is needed, the relevant fragments are inserted into the request. This approach is called retrieval from an in-house knowledge base. Training a dedicated model for a single company is required far less often than commonly assumed.
Which task comes first?
Pick one area where text is written by hand every day and where mistakes surface immediately: replies to routine requests, sorting incoming email, drafting a sales proposal. Measure the time per operation for two weeks before launch. Take the same metric afterwards — the comparison then rests on one volume and one sample.
Let’s discuss your project?
Tell us about your process — we’ll suggest where AI pays off fastest.
Related articles
What Is a Large Language Model, in Plain Language
A large language model is a program that continues any text with the most likely next word. It learned from books, articles and correspondence, and it answers in human language.
What Is a Large Foundation Model
A large foundation model is an AI trained once on an enormous body of data and then usable across dozens of company tasks without separate training for each.