Your staff already paste emails into ChatGPT. Your accounting software has a new “AI” button. A vendor promises an assistant that “knows your business”. Somewhere between the hype and the shrugging, you need a working answer: what is this technology, what can it do reliably, and where would it actually help a company like yours?
This guide explains generative AI without the jargon: how it works, what it is good and bad at, the three ways companies use it, and how to take a sensible first step.
Generative AI in one paragraph
Generative AI is software that produces new content (text, images, speech, code) from an instruction written in plain language. The best-known kind is the large language model, or LLM: the engine behind ChatGPT, Claude, Gemini and Microsoft Copilot. You give it text, it gives you text back. Ask it to summarise a 30-page tender, draft a reply to an annoyed customer, or pull the supplier name and VAT number out of an invoice, and it does it in seconds.
The important word is generates. It doesn’t look up a stored answer. It writes a new one every time.
How a language model writes
A language model is trained on a very large amount of text: books, websites, code, documents. During that training it learns one skill: given some text, predict what comes next. It works in tokens, small pieces of text that are often a whole word and sometimes part of one. (We explain them in tokens and context windows.)
When you ask a question, the model writes its answer one token at a time. For each new piece it looks at everything so far (your instructions, any document you gave it, the words it has already written) and picks a likely continuation. Then it does it again.
One piece at a time
Reply to Mr Novak: 10 of his 12 pallets ship today, the other 2 follow on Thursday at no extra cost.
Dear Mr Novak, thank you for your order. Ten pallets leave our warehouse today. The remaining two will follow on Thursday, and there is no extra delivery charge. Kind regards, Maria
After that first stage, providers train the model further on examples of good answers and on feedback from people, so it follows instructions and behaves like an assistant rather than a very clever autocomplete.
Three consequences matter for business use:
- It sounds right before it is right. The model is built to produce fluent, plausible text. Plausible and correct overlap most of the time, but not always. When it lacks the facts, it can still write a confident answer. People call this a hallucination, and you can design around it.
- It knows only what it was trained on and what you give it. It has never seen your price list, your contracts or yesterday’s orders, unless you put them in front of it.
- The same question can get a slightly different answer. That’s useful for drafting and awkward for anything that must come out identical every time.
How it differs from the software you already have
Your accounting system, ERP and CRM follow rules someone wrote. Same input, same output, every time. Generative AI works the other way round: it copes well with messy input, but you can’t fully predict what it will produce.
Traditional software vs generative AI
| Traditional software | Generative AI | |
|---|---|---|
| How it works | Follows rules someone wrote | Predicts likely text from patterns |
| Handles messy input (emails, scans, calls) | No | Yes |
| Writes new text | No | Yes |
| Same input, same output | Yes | Partly |
| Exact calculations | Yes | Partly |
| Easy to explain why it did something | Yes | Partly |
That’s why the most useful setups combine both. The AI reads and writes; ordinary software checks, calculates and stores the result. And if a simple rule does the job, use the rule. Our older post on workflows, AI workflows and agents shows how the pieces fit together.
What generative AI is good at
Six kinds of work where it already helps, day to day:
- Drafting. First versions of emails, job ads, product descriptions, quotes and reports. A recruitment agency can turn a hiring manager’s bullet points into a job ad in a minute; a consultant then edits it.
- Summarising. Long email threads, meeting transcripts, tenders, contracts. A support lead can read “what happened in this ticket so far” in five lines before calling the customer back.
- Extracting. Pulling specific fields out of unstructured text: supplier, invoice number, amounts and due date from a PDF; order lines from a customer’s email. An accounting firm can turn a box of scanned receipts into rows that a person reviews.
- Classifying. Sorting incoming messages by topic, urgency or language, so a support team’s tickets land in the right queue.
- Translating and rewriting. A distributor can publish product descriptions in Slovak, Czech, Hungarian and German, and adjust the tone for each market.
- Answering questions from material you provide. Give it your policies, manuals or database records, and it can answer staff or customer questions from them, with a link to the source. This technique is called RAG, and it’s worth understanding how it works.
The further a task drifts from “work with the text in front of you” towards “know a fact on your own”, the more you need a source, a tool or a person to check.
Where it’s dependable
- Rewriting and translating 1
- Summarising 2
- Sorting and extracting with validation rules 3
- Answering from your documents with a link to the source 4
- Facts and figures from memory 5
← Works well with text you give it Needs a source, a tool or a check →
What it’s bad at
- Facts it wasn’t given. Ask about a client’s contract terms without giving it the contract, and it will guess. Ask about a niche regulation and it may invent a paragraph number.
- Exact arithmetic and counting. The model sees numbers as text pieces. It is usually right on simple sums and unreliable on long calculations, VAT reconciliations or totals across hundreds of rows. Do the maths in ordinary code, or give the model a calculator or a database query as a tool.
- Recent events. Its training stops at a cutoff date. Unless it can search the web or read your systems, it doesn’t know what has happened since.
- Guaranteed consistency. Two runs can phrase things differently or, at the edges, decide differently. For decisions with legal or financial weight, you want rules, logs and a person.
- Knowing when it doesn’t know. By default it prefers to give an answer. You have to design for “I don’t know”.
Not just text: images, speech and code
Text
This is where most business value is today, and what most of this guide is about.
Images
Image models create or edit pictures from a description: product mock-ups, social media visuals, illustrations. They’re handy for marketing drafts. Check the tool’s licence terms, and be careful with real people’s faces and other companies’ logos. Separately, many language models can also read images, such as a photo of a delivery note or a scanned form. For an operations team that’s often more useful than generating pictures.
Speech
Speech-to-text turns calls and meetings into transcripts. Text-to-speech reads text aloud in a natural voice. Combine both with a language model and you get voice agents that can take a booking or answer “where is my order?” over the phone.
Code
The same models write and review code. Developers, ours included, use them every day, with a person responsible for what ships.
Three ways to use it in a company
1. Chat assistants
ChatGPT, Claude, Gemini, Microsoft Copilot. Staff open a window and type. It’s cheap to start and good for individual productivity: drafting, summarising, thinking out loud. The limits: people copy and paste between the assistant and your systems, results depend on how well each person writes prompts, and you need clear rules about which data may go in. Use the business plans rather than personal accounts, so you get admin controls and contract terms that cover how your data is used.
2. AI features in tools you already pay for
Your helpdesk suggests replies, your CRM summarises accounts, Microsoft 365 or Google Workspace drafts documents. No project needed, just a setting and often an extra licence. This works well when your process fits the tool as it is.
3. Models connected to your systems
Your own software sends text to a model through an API (a way for one program to talk to another) and uses the result inside a defined process. This is where AI stops being a side tool and becomes part of how work gets done. Take a distributor that receives 200 order emails a day:
Order emails at a distributor
- Order email arrives Free text, a PDF or a photo
- AI reads it Customer, items, quantities, delivery date
- Rules check Product codes, stock, credit limit
- Person handles exceptions Only orders that failed a check
- Order in the ERP
This option takes some building, but it’s the one that changes how much work a team can handle. If your data can’t go to a public AI service, the same setup can run on open-weight models such as Llama, Mistral or Qwen, hosted in your own environment.
Where it already fits in a normal company
You’re not late, and you’re not alone. According to Eurostat, 20% of EU enterprises with 10 or more employees used some form of AI in 2025, up from 13.5% a year earlier. Among large companies it was 55%. The most common use was analysing written language.
Typical first uses look like this:
| Team | A sensible first use | Who stays in charge |
|---|---|---|
| Finance and accounting | Reading invoices and receipts into the system | An accountant reviews flagged items |
| Sales | Call notes into the CRM, follow-up drafts | The rep approves before sending |
| Customer support | Sorting tickets, suggested replies | The agent edits and sends |
| Operations | Order emails into the ERP | Rules, plus a person for exceptions |
| HR and recruitment | Job ad drafts, CVs parsed into fields | The recruiter decides; no automatic rejections |
| A private clinic | Answers to booking and preparation questions | Anything medical goes to a nurse |
Notice the pattern. The AI does the reading and the first draft. A person, or a rule, makes the call.
What to do next
- Give your team a business-grade assistant and a one-page policy on what data may go in. Let people find small wins on their own.
- Pick one process with lots of repetitive text work: emails, documents, forms. Measure how long it takes today and how often it goes wrong.
- Run a small pilot with real examples and a person checking every output.
- Connect it to your systems only once the pilot proves it saves time.
Our guide on bringing generative AI into your company walks through that plan step by step.
If you already have a process in mind and want to know whether it needs AI or just a good rule, a 30-minute call is a quick way to find out. Get in touch.
