What is generative AI? Explained for business, in plain language

AI is the umbrella term for software that does tasks that normally take human thinking. Within it, there are roughly two flavours: AI that predicts and AI that creates.

Traditional AI: predicting and recognizing

Traditional AI, often called machine learning, learns patterns from examples and uses them to predict or classify. Think of a spam filter that learns from the emails you mark as spam. Or a system that estimates how much you'll sell next month. The answer is a number, a category or a yes or no.

A model like that is usually trained on one task, often with your own data. If you want something else, you train a new model.

Generative AI: creating

Generative AI is a recent branch of the same tree. It also learns patterns from examples, but uses them to make something new. Not a yes or no, but a whole email. Not a category, but a twenty-page report in five lines.

The big practical difference: you don't have to train it yourself. One model can handle a huge range of tasks, and you tell it in plain language what you want. That instruction is called a prompt.

And ordinary automation?

There's also automation without AI: fixed rules that someone has designed. When an order comes in, send a confirmation. That works well as long as everything is predictable. Generative AI helps exactly where things become unpredictable: free text, an email that's different every time, a PDF with a different layout. In practice they often work together. The table further down shows them side by side.

Without formulas: generative AI has seen an enormous number of examples, and based on those it makes something that probably fits your question.

1. 01Training. The model is shown enormous amounts of text, images or code. From that it learns patterns: how sentences flow, what a quote looks like, what a cat looks like.
2. 02Your instruction. You give a question or instruction, possibly with a document or an example attached.
3. 03Piece by piece. A language model keeps predicting what most likely comes next, word by word, until the answer is complete. An image model works differently under the bonnet, but the idea is the same: it makes something that fits the patterns it has learned.

The models that produce text are called language models, or LLMs (large language models). GPT, Claude, Gemini and Llama are examples. How a language model works, with tokens and a context window, is explained in what is an LLM?

Two things to remember:

• It predicts, it doesn't know anything for certain. That's why a wrong answer sounds just as convincing as a right one.
• It doesn't know your business. It knows a lot about the world, but nothing about your customers, prices or agreements. Whatever it needs to know, you give it. That's also the key to better results.

Mostly work with language. And in an office, almost everything is language: emails, quotes, reports, contracts, customer questions. A few examples from everyday office work:

• Emails and quotes. A first draft in your tone, with the right details in it. You read, adjust and send.
• Summaries. Of a meeting, a long report or a customer file. What's in it, what was agreed, who does what.
• Customer questions. A reply prepared with the order information or the policy attached, so a colleague only has to check it.
• Reading documents. Pulling data from invoices, orders, CVs or contracts and putting it into the right system, instead of retyping it.
• Translation. Product texts, manuals or emails, in several languages at once.
• Answers from your own knowledge. An assistant that answers questions from your handbooks and procedures, with the source included. That's called RAG.
• Code and formulas. An Excel formula, a small script or a first version of an internal tool.

See the pattern? Generative AI helps in three places: where something comes in and someone retypes it, where someone searches for information across three systems, and where someone makes a document that's almost the same every time. How to find that work in your own organization is covered in which processes can you automate.

Want to know how many businesses already use it, and for what? The numbers and sources are in AI in the Netherlands in numbers.

Almost every organization starts the same way. Someone tries ChatGPT, a colleague follows, and after a few months more and more people use it. Each in their own way. That's a fine start, but it's phase 1 of 4.

In the organizations we help, now more than 160, we see four phases of AI adoption.

Phase 1: Prompting

People use ChatGPT or Copilot, each on their own. Rewriting an email, summarizing something, working out an idea. It saves individuals time, but the work itself doesn't change. And often nobody knows exactly which data ends up where.

Phase 2: Assistants

You build your own GPTs and skills that know your work: your tone, your products, your way of working. An assistant that writes quotes the way you do. Everyone works with it in the same way. But you still start the conversation, and you take the next step. More in build your own AI assistant.

Phase 3: Workflows

Automations that run by themselves. An order arrives by email, AI reads it, the order is ready in the ERP and a colleague approves it. Nobody has to think about it any more. This is where the work itself changes.

Phase 4: Agents

Digital colleagues that get a goal and take the steps in your systems themselves: reading a customer question, looking up the order, preparing a reply and a solution. Always with a human in the loop, who approves where it matters. Also read what is an AI agent?

Most organizations are in phase 1. The value is in phases 3 and 4, because only there does AI really take over work in your processes. More on that in the 4 phases of AI adoption.

Generative AI is powerful, but not flawless. Sort out these five things before you roll it out widely.

• It can make things up. A source that doesn't exist, a number that comes from nowhere. That's called a hallucination. You limit it by having AI work with your own documents with the source included, and by checking what matters. More in AI hallucinations.
• Data and privacy. Business versions from the large providers don't train on your data by default and work with a data processing agreement. Free consumer versions don't offer that certainty. And personal data always falls under the GDPR. Read using ChatGPT safely under the GDPR and using AI safely.
• The EU AI Act. For most office applications it's mainly about care and transparency. The law does ask you to work on AI literacy: your people know enough to use AI responsibly. See the EU AI Act for SMEs and AI literacy.
• A human stays responsible. AI prepares, a human decides. How to arrange that without creating extra work is in human in the loop.
• Shadow AI. If the organization doesn't arrange anything, employees arrange it themselves, with a private account or a free tool. A good business alternative and a few clear agreements work better than a ban.

Start with the work, not the tool. A question we hear a lot: which AI tool should we choose? The honest answer is that you only know once you know which work you want to change.

1. 01Choose the work, not the tool. Where is information retyped, searched for or the same document made over and over? That's the first principle of our APO method: process first, technology later.
2. 02Know where you stand. With the free AI self-assessment you see in two minutes where you are and what the next step is.
3. 03Teach your team to prompt well. A good prompt is a good work instruction. Start with the guide to prompting, or bring your team to the same level in one half day with an AI training on your own work.
4. 04Make agreements. Which tools are allowed, which data can go in and where a human checks. It doesn't have to be a thick document.
5. 05Work out the numbers for one process. How long does it take now, with how many people and how often per month? Then you know what AI could deliver there.

Rather not do that last step yourself? In the AI Quick Scan we spend one day with you on site. Within a week you have a baseline, a business case per opportunity, a roadmap and honest package advice. The Quick Scan costs € 6,000, excluding VAT.

After that, we build phases 3 and 4 together in the AI Partnership. We only start when the business case pays back at least twice in the first year.