AI agents for the management team: what leaders need to know and decide now

AI agent or chatbot?

A chatbot gives answers, while an agent does things and works towards a goal on its own.

The difference is in the doing: an agent plans steps and uses systems. The UK audit regulator, the Financial Reporting Council, describes agentic AI as “artificial intelligence systems that can orchestrate and execute multiple components and/or tasks toward a goal, with some degree of autonomy”. In practice, that could be an agent that checks and forwards invoices, handles customer questions, or drafts a standard schedule that a planner then reviews.

The first question for a management team is: do we have the right culture and organisation for a good innovation and AI strategy? Then: which tasks no longer need to be done by people, and which examples can we learn from?

How good are AI agents really?

On tests, agents are improving fast, but they still fail roughly one in three attempts.

According to the 2026 AI Index Report from Stanford HAI, the success rate of agents on the OSWorld benchmark, which uses real tasks on a computer, rose from 12% to around 66%. That still leaves roughly one failure in every three attempts. For a management team, that is the heart of the matter. A test score does not tell you how an agent performs in your own processes; you have to measure that yourself, and it will not work without oversight. To see what happens when AI runs unchecked, read AI bloopers: artificial intelligence’s most memorable mistakes.

Why is it a management issue?

AI has become an everyday tool, and the bottleneck is often the organisation, not the employees.

In the EU, 19.95% of enterprises used AI in 2025, rising to 55.03% among large enterprises, according to Eurostat. Of the companies that use AI, 31.05% use it for business administration or management. In the Netherlands, one in six companies used AI in 2025, double the share two years earlier, according to Statistics Netherlands (CBS).

In Microsoft’s 2026 Work Trend Index, which surveyed ten markets including the US, the UK, Germany, France and the Netherlands, only 26% of AI users say their leadership is clearly and consistently aligned on AI. According to the same report, organisational factors such as culture, manager support and talent practices are linked to more than twice as much of the AI impact that users report themselves as individual mindset and behaviour. That is a statistical association, not a proven effect. That makes it a management team issue. An agent deployed by one department touches the processes, data and customers of the whole organisation.

Five management team decisions

The management team decides where agents go, what they may do, who is accountable, what people need to know and how to measure it.

1. Where do we use agents, and where not?

First ask every department to map its processes, from processing invoices to preparing reports, and to decide per task: can an agent do this, or does it belong with people? Then start with work that recurs often, is well defined and has an outcome you can check. Think of processing documents, answering standard questions or moving data between systems. Also agree where you do not want agents, for example in decisions about people.

There is a legal line here too. The EU AI Act classes AI used to recruit and manage workers, and AI that decides whether someone gets access to essential services such as credit, as high-risk, according to the European Commission. From 2 December 2027, strict requirements apply to those systems, including appropriate human oversight. Organisations that use them have to arrange that oversight themselves.

2. How much may an agent do on its own?

This is the most important decision. Make it deliberately and write it down, so that it does not happen by default. Distinguish between three levels: an agent that only reads and summarises, an agent that makes proposals a person approves, and an agent that acts on its own. For each application, record which level applies and when an agent may move up a level.

3. Who is accountable for mistakes?

You cannot hold an agent to account. You can hold a person to account. Regulators are clear about this. When it published its guidance on generative and agentic AI in audit in March 2026, the Financial Reporting Council stated that “the human auditor is always accountable” and that the guidance does not change that (FRC). Any organisation can apply that principle: give every agent a named owner.

4. What do our people need to know?

Since 2 February 2025, organisations that deploy AI in the EU must support the AI literacy of their people. Since 2 August 2026, national authorities have supervised this, according to the European Commission. There is no mandatory training or certificate; you may keep an internal record of what you have done. Among companies that considered AI but did not use it, knowledge is also the most common obstacle. Of those EU enterprises, 70.89% cited a lack of relevant expertise, according to Eurostat. And not everyone admits to using AI already. We wrote about that phenomenon in How to handle AI shame?

5. How do we know it works?

Agree in advance what an agent should deliver: shorter lead times, fewer errors, more time for other work. Also measure what goes wrong and how often a person has to step in. If an agent needs a lot of corrections, weigh the time those take against what it delivers.

What should customers notice?

If your agents talk to customers, those customers must know they are interacting with a machine.

That is a transparency duty under the AI Act, and it has applied since August 2026. The European Commission gives the chatbot as an example. So decide how you make that clear to customers.

And in the long term?

Other new technology is likely to follow agents, so what you learn now is something you will need again.

Agents are one development among several that will change business. That is why I advise executives to give their people room now to try and experiment with agents, on work that recurs often and is easy to check. Enthusiasm is welcome, but keep a critical eye on what an agent really delivers and how often a person has to step in.

For a broader view of what else is coming, see Beyond AI: five mega-trends that will reshape business in the next decade. How you as a manager get people moving for a change like this is the subject of the keynote on leadership and management.

Frequently asked questions

Are AI agents and agentic AI the same?

In practice, both terms are used for the same thing. Agentic AI is the umbrella term for systems that carry out multiple tasks towards a goal on their own; an AI agent is one such system.

May an AI agent decide on its own?

That depends on the kind of decision. For high-risk applications, such as AI in recruitment or managing workers, the EU AI Act requires appropriate human oversight from 2 December 2027, among other things. For other applications you set the level yourself, but the organisation remains accountable.

Do we need an AI officer or committee?

The law does not require it. The European Commission states that no specific governance structure is mandated for AI literacy. A permanent owner for each agent is still sensible.

How do we start small?

Choose one well-defined task, let an agent only make proposals at first, measure the results for a few months, and only then decide on more autonomy.