You are about to have digital subordinates. Are you ready to manage them?
Picture this: a call comes in about a fatal accident. There are few facts so far, the clock is already running, and the commission has to reconstruct the events, establish the causes and produce the paperwork.
As an experienced AI user, you ask a chatbot (GigaChat, DeepSeek and the like) to draft a plan. It answers — and stops. Then you pick the next prompt yourself, carry the data across, ask it to analyse a photograph or prepare interview questions…
The chatbot helps you and speeds things up, yet so far it resembles a diligent intern: it has completed the task and is waiting for the next one.
And this is where the AI agent comes on stage. Unlike a chatbot, an AI agent can do more than carry out separate instructions — it can reach the final goal on its own across many steps. It first builds a plan, requests materials, chooses tools, checks the results and drives the task up to the next human decision. In management terms this is no longer an intern (or a secretary) but a proactive digital specialist.
Over the past 4 years AI has made a great leap forward:
This is not merely a calendar of technologies, it is a scale of our readiness to use them.
Which is a good moment to consider your own level of AI skills and “which year you are living in”.
I have travelled that road myself: from prompts to assistants, agents and building my own tools. The new does not cancel the old: without a well-formulated task an agent will not produce a good result either. What changes is the scale of delegation.
I tested that difference on an end-to-end case: the investigation of a fatal accident. The same route can be walked through with a chain of prompts or handed over to a cloud agent, which at the time of the experiment was available free of charge and without a VPN. The result required is the same, but the work is organised differently.
| Chatbot and prompting | AI agent |
|---|---|
| The human chooses every next step | The agent follows a plan and asks for what is missing |
| Context has to be carried over and repeated | State is kept in files, tables and artefacts |
| A limited set of tools is invoked by a separate request | The agent connects the permitted tools itself |
| Verification has to be launched with a new prompt | Self-checking is built into the working cycle |
| The human is the operator of the process | The human is the methodologist and the decision controller |
In the agent scenario the system analysed photographs and interviews, checked the regulatory basis, built a timeline and a cause tree, found the gaps and prepared draft documents. It did not replace the commission and did not take legally binding decisions. But the specialist no longer had to drag the AI through every stage by hand.
The phrase “AI will not replace the expert” is true, yet no longer sufficient. Agents change the way expertise is applied: less attention goes into switching between steps, more into facts, causality and the quality of decisions.
In the past an HSE specialist could choose an expert career track and never manage people the way a manager does.
Now even a junior employee may acquire digital AI colleagues. They need goals, skills and tools; their authority has to be bounded and their work accepted. Which means the skill of managing AI agents is now needed by everyone.
Anyone who has tried agent mode on a real multi-step process usually stops asking: “What else should I write to the chatbot?” A different question arises: “Which process do I hand over to an agent next, and where do I put the time it frees up?”