What has actually changed

A few years ago, writing the first draft of a letter, meeting summary or report could take an hour. Today a large language model (LLM) produces that draft in seconds. A large language model is a program trained on a very large amount of text so that it can write, summarise, translate and answer questions in everyday language. That is the biggest change at work: the cost of producing a draft has dropped close to zero.

Other changes came with it. You can paste a long document and get its gist in five minutes. You can turn a messy list into a table, or make a stiff letter sound friendlier. Some apps now also have AI agents: AI that, besides answering, carries out steps, such as opening files, filling in forms, sending messages or changing data, one step at a time.

So what stays the same? Almost everything that makes your work valuable. Someone still has to know what a document is for and who will read it. Someone still has to know the context that was never written down: promises to a client, a manager's habits, office rules, what was tried last month. Someone still has to judge whether the result is right and fit to send. And the one who signs that letter is still a person.

The honest picture is roughly this. A draft that used to take an hour is now ready in minutes. The work of deciding, checking and taking responsibility has not shrunk. Because drafts arrive faster and in greater numbers, checking now takes the biggest share of your job. The people who gain most from AI know exactly what result they are after and check it carefully. How fast you type a prompt hardly matters.

Why does the responsibility to check not move to the machine? There are three simple reasons. First, the model bears no consequences. If a figure in a report is wrong, the one who is reprimanded, loses a client's trust or has to apologise is you and your team. Second, the model does not know what it does not know. It writes a wrong sentence in the same confident tone as a right one. Third, the model does not see the whole situation. It sees only what you type or paste, so it cannot know that a sentence will offend a long-time partner or break an agreement that was never written down.

With agents, this question of responsibility matters more. When AI only writes a draft, a mistake stays on the screen until you copy it. When an agent carries out steps, a mistake can become an action at once: a message sent, a file overwritten, data changed. That is why many teams, ours included, apply the same rule to agents as to a new team member: it may prepare the work, but steps that are hard to undo wait for the person who is responsible.

One more thing that has not changed: an agent can report that its work is done, while that report is itself only text it wrote. How to tell a report from evidence is covered in the lesson "When AI goes wrong in hurried hands" in this module.