Insights from a discussion with Marcel Pesch, CEO and founder (Academy4AI), and Holger Rosemann, Managing Partner (XQI Interim Partner GmbH)
The Introduction
AI agents are currently the topic that is discussed the most, yet with the least precision. In a session lasting around 90 minutes with experienced interim executives, we sought to clarify the terminology, openly discuss the business model and not shy away from the uncomfortable issues – liability, data leakage and loss of control.
Marcel Pesch, CEO and founder of Academy4AI, as well as winner of the E-Learning Award and a member of the German Federal Association for AI, says he runs his company with over 30 AI agents and three permanent staff members. This set-up makes him an interesting interviewee – not because it serves as a model for everyone, but because it highlights the questions that every interim manager will have to answer over the next two years.
The discussion was moderated by Holger Rosemann, Managing Partner at XQI Interim Partner GmbH.
This article summarises what we consider to be the six most relevant insights.
1. Automation, assistant, agent: three terms that are constantly confused
Most of what is marketed as an ‘agent’ these days isn’t one at all. Pesch draws a clear line.
An automation is a fixed, recurring process. Someone fills in a contact form, and the data is automatically fed into the CRM. Useful, but: ‘It’s not an agent, it’s an automation.’
An assistant is an input-output system with no connection to other systems. You prompt it, receive a reply, perhaps rename the chat, and treat it as a personal marketing assistant. That, too, is not an agent.
According to Pesch’s definition, an agent requires three components:
- A Large Language Model as a thinking engine – whether it’s ChatGPT, Copilot, Claude or Perplexity is of secondary importance.
- A memory, i.e. a knowledge database containing product documentation, customer pain points, handouts and papers. Pesch comments: “Ultimately, this always influences the quality of the output.” This point is regularly underestimated in practice.
- At least two tool integrations: “I always need integration with tools – at least two tools.”
It is only through integration that a qualitative leap is achieved – not in terms of speed, but in terms of autonomy: “You have a certain degree of decision-making power over the agent itself.”
Why this is relevant for clients: Anyone who promises a client an ‘agent’ and delivers automation loses credibility. Conversely, anyone who builds a genuine agent without deliberately restricting decision-making authority will face the problems outlined in section
2. Two examples of mistakes that are worth hearing at least once
Pesch speaks openly about two situations from within his own organisation.
The unsolicited post. An agent was connected to Langdock and Meta and published content on their own initiative – “because we simply didn’t take that approach of, ‘hey, please get human approval first’.”
The credit card. In a test, an agent was given a prepaid credit card with a balance of 200 euros; the task was to book a cheap flight from Düsseldorf to Bologna, costing no more than 200 euros. The agent completed the task – and then attempted to make another booking elsewhere without having been given a further instruction.
The second case is the more instructive one. Not because an agent did something forbidden, but because they repeated an action they’d learnt once without any further prompting. Pesch’s conclusion: this is precisely where “security mechanisms should also be implemented”.
In practical terms, for a client, this means: any tool integration with financial or legal implications – payment, ordering, publication, acceptance of a contract – requires a strict technical limit, not just a prompt instruction.
3. ‘Human-in-the-loop’ is not just a buzzword, but an architectural decision
Pesch strongly recommends not allowing agents to operate externally without authorisation: “It’s always important here to keep ‘human in the loop’ – to ensure the agent doesn’t actually send the post out itself. It can do so, but I wouldn’t recommend it, not even here.” Even with our own newsletter agent, the rule is: “We still take a look at what’s going out.”
Furthermore, he describes a second, technical control principle – the ‘Sheriff’ method, a ‘classic dual- and triple-check principle’. The key trick is that the control agents deliberately run on different models to the producing agent. One verification agent runs on Claude, one on Gemini, and one on Perplexity. The logic behind this is obvious – a model that has the same systematic bias as the generator won’t spot its errors.
Despite this control architecture, Pesch remains realistic: “I’d still say there are three to four iterations for every output. You always need to double-check, always.”
That is the most honest statement of the entire conversation – and it puts the usual productivity promises into a healthy perspective.
4. The business model question: Those who bill by the day are penalising themselves for efficiency
This was the point at which the discussion in the room became most lively – and at which there was no ready-made answer.
Holger Rosemann put the dilemma as follows: “Anyone who continues to bill by the day is actually penalising themselves for being efficient.” If an interim manager, with the support of agents, delivers the same service in a third of the time, the fee simply decreases if the daily rate model remains unchanged.
Three possible solutions were outlined during the discussion:
Split billing. Pesch refers to an example where the service contract distinguishes between human and agent-based services – in the order of 70 per cent human workforce and 30 per cent agent-based workforce. This is an initial pragmatic approach which does not resolve the issue, but makes it negotiable.
Billing based on replacement costs. Instead of billing for the time actually required, the question is: how much time would it have taken if a human had been assigned to the task? Given the time savings of 80 to 90 per cent mentioned by Pesch, this represents a substantial difference – and it must be specified in the contract, not discussed retrospectively.
Outcome-based models. One participant outlined a contract for services with a share of the savings achieved over the first two years. He himself highlighted the downside: “But it entails a completely different risk profile for the provider if that target is not met.”
The most important objection, in our view, came from the audience and applied to all three models: “It’s all about measurability.” Without clear documentation of the baseline at the start of the engagement, any savings cannot be proven later on. Anyone considering performance-related remuneration must establish the baseline in week one – including the parameters which, if they change, will trigger an adjustment to the agreement.
And an underestimated cost factor: Tokens. One participant reported annual token costs of €300,000, and rising. Drawing on his own experience, Pesch described how a voice agent on the website drove up costs – after switching to a chat-only solution, they fell by “easily 70 or 80 per cent”. His recommendation to interim managers: explicitly include token costs for agent-based work in the invoice.
5. Law and liability: Technology is moving faster than the law
The most awkward question of the evening came from a participant and remained unanswered: if an interim manager works with seven agents, does this affect their professional indemnity cover? Are they liable sevenfold?
Pesch’s assessment of the legal situation is matter-of-fact: “There is currently no case law on this, nor on the taxation of AI agents or the liability of AI agents.” The EU AI Act sets out an initial framework – AI competence as a requirement, labelling obligations – but: “Nevertheless, the technology is advancing much, much faster than the legislation.”
The practical implication is not to wait and see, but to establish internal company rules. Pesch reports having drawn up a comprehensive Gen-AI handbook for a major corporation. The rationale is provided by a cited study: according to this, more than half of employees in medium-sized companies use AI secretly. Anyone who fails to set guardrails does not have an AI-free organisation – they have an uncontrolled one.
For Interim Managers, this means two things: firstly, discussing their own liability situation with their insurer before deploying agents on assignments. Secondly: specifying in the engagement contract which decisions are prepared by agents and which are the responsibility of humans.
6. Data sovereignty: The GDPR alone is not enough
One participant raised a point that is often overlooked in many AI discussions: the US Cloud Act can effectively undermine the GDPR – data can flow to the US despite European contractual arrangements.
Pesch has drawn conclusions from this for his own company: “We no longer have any US providers in our portfolio.” Work is carried out via a European platform with a data processing agreement; for larger applications, he recommends German hosting. In the medium term, he expects to see more local server installations on in-house hardware.
A notable observation came from Holger Rosemann regarding the actual use of models in Europe, where Chinese models are said to account for a significant proportion – prompting a dry comment from one participant: “We don’t know where the data goes there either.”
For Interim Managers who regularly work with confidential client data, this is no academic question. The minimum requirement is to know which model underpins the tool being used and where it is running.
Conclusion: Four key takeaways
Firstly: The difference between automation and an agent lies in decision-making authority – and decision-making authority requires limits.
Secondly: ‘Human-in-the-loop’ is not a convenience feature, but a form of risk management. Three to four verification loops per output remain a realistic target.
Thirdly: The daily rate model is coming under pressure. Anyone who does not actively consider contract drafting will lose the efficiency gains to the client – without receiving anything in return.
Fourthly: The legal issues remain unresolved. This is no reason to wait, but it is a reason to clarify liability and data flows contractually before the first agent goes live.
Or, as Marcel Pesch put it: “In the end, it’s always the human who wins.”
This article summarises a dialogue between experts. The views expressed regarding the law, liability and taxation are the opinions of the participants in the discussion and do not constitute legal advice.
