What it can do, where it goes wrong, and how to get started without wasting money

Generative AI is systems that create new content — text, images, audio or code — based on what you describe, rather than just analysing existing data.
ChatGPT, Claude and Copilot are examples. The difference from older AI is that you steer the system in plain language instead of with rules or code, which is exactly why it suddenly concerns everyone and not just developers.
Most of what is written about generative AI is either sales material or warnings. This page tries to be neither: it describes what the technology is good at, what it is bad at, and what I see working in practice at smaller companies.
I run several systems myself that are built on generative AI in live use, including Nyheter AI, where a newsroom of AI agents monitors 27 sources around the clock, and Mitt AI-team, which builds AI colleagues for small and medium-sized businesses. What is written here is therefore not taken from a report.
The names and features of the tools change quickly. This describes the position in November 2026, and the principles last longer than the product names.
Realistically
The practical conclusion from the two columns is simple: let the system prepare and let a person decide. An AI project easily goes wrong when that rule is broken — either by letting the system make decisions it cannot handle, or by adding so much control on top that the time saving disappeared.
Use
Quotes, customer replies, product texts, newsletters, minutes. The gain is not that the text becomes better than what you would have written, but that it takes twenty minutes instead of two hours to reach a version you can edit.
Being able to ask “what applies to warranty claims on this product?” and get an answer from your own material. This can be the most valuable use in a smaller company, and is also the one that requires the most care about what may be sent where.
Instead of one general chat window: a set of specialised assistants, each with its own task and its own way of working. This is how I built my own AI team, and it is the form that has proved to be genuinely used — a general tool easily becomes nobody's responsibility.
Illustrations, product images and material for social media. Quick to get started with, but worth thinking about what you want to stand for — a company that looks generic in its communication rarely gains by becoming even more generic.
Getting started
Start small. Choose one recurring task that takes time, that is not sensitive, and where a mistake is spotted straight away. Let two or three people use it for a few weeks before you buy licences for everyone — the licence cost per user is rarely where the money goes.
Decide early what may be sent where. Personal data and sensitive material belong in a service with the right agreement, not in a free account that someone set up on their own. It does not have to be complicated, but it does have to be written down — otherwise people guess, and they guess generously. There is more about this on the page about GDPR and the EU AI Act.
And expect that the first thing you choose will not be the thing you stick with. That is not a failure, it is simply how it goes. The point of starting small is precisely that changing direction should be cheap.
If you want help choosing where to start, that is exactly what an AI strategy is about, and the implementation page describes what happens once the choice has been made.
And if you are facing the choice of tool itself: Copilot or ChatGPT for the company?
Frequently asked questions
Helps you decide where generative AI is useful in your particular business, chooses and sets up the tools, makes sure there is a written rule for what may be sent where, and teaches staff to use it. With me that means advice, tool selection and building, among other things, AI agents and search across your own documents. The tools are replaced quickly, so the work starts from your business and not from a particular product.
If you would rather start with the big picture, an AI strategy can be the first step, and AI consulting for businesses describes what working together looks like.
The licence cost per user and month is rarely the problem — what costs is the time it takes to work out where it helps and to turn it into a way of working.
Start with a few users on one defined task before you buy licences for everyone.
Yes, and that is where the benefit can often be greatest. But it requires that you have decided what may be sent where.
Personal data and sensitive material should be handled in a service with the right agreement, not in a free account.
You don't, and that is precisely the point of letting a person decide. Choose areas of use where a mistake is spotted straight away, and avoid those where an error can pass unnoticed and cost money.
That depends on how it is introduced, and nobody can promise the answer in advance. The aim of what is described here is that tasks nobody had time for get done and that the tedious parts of a job become fewer — not that staff are removed.
That is an intention, not a guarantee, which is why you decide where and how it is used.
A free half-hour call. You get an honest answer, even if that answer is that you should wait.
Book a first meeting