Get in touch with us
We at accompio will be happy to help you.
25.09.2026
How can AI be implemented in companies? Henrik Wenck explains which processes are suitable, how employees are trained, and how the added value can be measured.
AI in the company They rarely realize their full potential through a single tool. What is crucial is how well an application fits the existing process, what data is available, and how the results are subsequently processed.
Many employees already use artificial intelligence, for example for text, analysis, or code. At the operational level, the potential often remains untapped. It is not enough to speed up a single workflow. Companies must consider the entire process and determine where AI makes sense, where classic automation is better suited, and when a human should decide.
Henrik Wenck, Senior VP at accompio AI, explains in the expert interview which processes are suitable for AI, what makes good AI training, and with which metrics the actual added value can be measured.
How do companies find suitable AI use cases? What role do data quality, employees, and measurable goals play? Henrik Wenck shows in the video how individual AI experiments can be turned into a structured approach for processes and teams.
00:06 Welcome from Dresden to our today’s accompio topic: AI in the company. We talk about process automation, training, and measurable added value. I’m Henrik Wenck. It’s nice to have you here. Please introduce yourself briefly and explain your role at accompio. Yes, gladly. Hi Luisa. I’m Henrik Wenck,
00:29 I have been working in the AI field for over ten years and hold the role of Senior VP at accompio AI. This is the AI unit of accompio and in our day-to-day work, we help companies find suitable AI use cases and implement them. This includes automation, and also empowering the employees of the respective companies to use AI. When we put this into practice, what would you say?,
01:05 In which areas does AI change the work routine in companies? So, I believe in general: AI is finding its way into more and more companies. A large number of companies already use or are starting to use AI. But I would distinguish here. On the one hand, there is a lot of individual use, i.e. employees who use AI accordingly,
01:34 to write texts, to have code written, or to analyze things. On the other hand, AI at the process level, at the corporate process level. I think there is still room for improvement there, and that’s where we’ll see a lot more in the next few years. Because if AI occurs at the individual level and you imagine the process, with different process steps, then AI currently still has problems,
02:05 To move from one process step to the next. And which activities are particularly suitable for being supported or even automated by AI? So, in very general terms, you can think of it like this: there is normal software development. And in software development, you have an input and an output, and then create a framework for how the input becomes the output.
02:34 So a developer then writes the rules down in code, and then that is processed accordingly. With AI, this mantra has changed. You have an input and an output, and based on examples, the AI model learns the rules. And suddenly, use cases were possible that were simply not possible before because the rulebooks were too complex.
02:59 And this includes Use Cases in normal text, i.e. natural language, as well as making it more specific: For example, if you have an IT support ticket and want to assign it to an output or department or person, that is a great use case for AI. Likewise, if you have questions and want a specific answer. The answer can also be based on internal knowledge
03:32 Its. So if you want to associate knowledge, AI is also great. What I also often see in practice is AI in data extraction. That is, a document comes in and automatically extracts things like the invoice amount, the date, etc. This is all based on unstructured data, and that can be mapped to an output. Now that we have that
04:01 Let’s take a step back and look at the process from the perspective of the individual steps: After all, quite often individual steps are already automated. What do I need to change in this process to create actual added value or an improvement for the entire process? In other words, I believe that first and foremost you should look at the process from the bird’s-eye perspective. You should look at what the process steps consist of and really start at the bottleneck.
04:33 And then I also have to ask myself the question: If I automate this point or make it more efficient now, am I actually just moving the problem around? For example, if I let someone write texts and then spend the same amount of time in the next step of reviewing them as if I were writing the texts myself, then I haven’t gained anything. Therefore, processes with AI need to be rethought, and you have to ask: Where do you start?
05:05 That’s what I would always recommend: in the problem where there is a bottleneck. Then you have to look at: What happens before that? What data goes in? Do I have the data quality I need to support this process and automate it? And then I also have to look at: What happens after that? So the output of the process or the process step can then be further processed directly, and if so, under what conditions?
05:38 So does the model have to exceed a certain threshold? So at what confidence level can the result be used further? And at what point should a human also look at it? That’s my next point regarding process automation. Where is AI currently reaching limits and where is human intervention essential? So, as I’ve already said, I believe that data quality plays a role.
06:11 it plays a huge role in very, very, very many AI use cases. Then of course there’s the probabilistic nature of AI. So, you can imagine this: if I have a process that has ten process steps and each of those process steps is processed with 95 % accuracy, then the errors in each individual process step multiply. And suddenly I’m no longer at 95 at the end of the process. %
06:43 Accuracy, but only after 60 %. The question is: Is it enough or can I, as I said at the beginning, also use normal software development to make certain processes run deterministically? So, that’s it. What we often see is the regulatory logic. Nothing is more frustrating than having implemented an AI use case that works well and then realizing, “Hey,
07:15 I really shouldn't have been allowed to implement that. And to the second question: What should humans continue to do? Hm, I think AI can do good work, but it's hard to take on responsibility. So the responsibility should lie with humans, and the same goes for relationship work. Whether that's internally or with customers externally.
07:46 You have just mentioned the responsibility that is essential. If we assume that AI currently handles routine tasks: What skills or what skills will then become more important for employees? So, I believe the difference between current work and previous work is that many work processes no longer start from scratch. You rarely start with a blank sheet of paper anymore.,
08:16 But instead, you let the first draft be generated by AI, and that simply requires a different skill set. I have to read the output accordingly, I have to be able to assess it, and I also have to be able to question it critically: Is what is being shown to me really the right thing? So, if I had to summarize it, I would say:
08:41 Process knowledge has continued to become even more important, because I need to understand the processes. I need to understand what takes time, how long it takes, and where errors can occur. And on the other hand, this critical evaluation. AI models are incredibly good at generating output that sounds good. And that’s exactly where I need the expert assessment to say: This sounds good, but it’s technically wrong. Let’s come back to the topic of AI training.
09:14 This kind of skill needs to be developed anew by an employee. What should an AI training program convey so that an employee truly develops these skills and doesn’t just learn how to use a tool? I believe the question already raises the problem head-on. So, many AI training programs are very general in nature, and I believe that an AI training program should focus on the daily work of the respective employee, because otherwise you end up incorporating content that is nice and entertaining,
09:48 but ultimately, not at all about the things you do on a daily basis. I think we should also focus on not just teaching rules about how to use certain software, but understanding the principles, because the software landscape is changing incredibly quickly right now. That way, you could say: Every few months there are new providers, or many companies are questioning that too: Which providers should I commit to in the long term? And
10:28 Exactly, once you understand the principles behind it, you can then transfer them to new tools. And then probably the third point. I would say: critical thinking or critical examination. Employees should not only be trained to produce output, but also to question the output. And now it is often the case that the expectations after such training
10:58 Well, yes, not being fully fulfilled and that what you have learned cannot necessarily be transferred and taken with you into everyday work. What would you say: What is the reason for this? Well, we have already seen this quite often in practice. I think it is because: You have had the great further training, you have learned a lot accordingly. You come into the office the next morning and there are hundreds of emails or 100 emails and calls that you have postponed accordingly for the day,
11:32 And because of the lack of time, you quickly get back into your own previous routine. And breaking out of that routine is actually the real challenge. In my opinion, you should give employees time and space to apply the newly learned skills. And just as importantly, the corporate culture should openly demonstrate this and expect employees to do the same.
12:03 apply what you have learned in your daily work. I would like to address the topic of productivity again. You already mentioned it earlier: If I implement activities with AI and then have to spend the same amount of time reviewing them as it took to create them myself, it’s not adding value. That’s not productive. Can you
12:28 Perhaps I could offer a few more tips on how to recognize whether an activity has really become more productive through AI? Yes, I believe that in AI it is important to always distinguish between two concepts. One is activity and productivity. AI tools are incredibly good at generating output. But that’s just a lot of stuff. That doesn’t really say anything about productivity.
12:56 Where productivity increases accordingly is when it can be measured. When I know in a measurable way: How did the process work before? Where am I now? So it’s about showing this comparability and this measurable effect. That’s one way to improve productivity; otherwise it’s just a gut feeling. Are there any specific metrics available to make this measurable or this measurable effect visible in the company?
13:30 There are definitely general metrics for this. So you can look at it accordingly: What is the current throughput of the respective process and how is it after the AI automation? What about the quality, i.e. how much rework do I have to do? How often do I have to look at this output as a member of staff and correct it accordingly? And beyond that
14:01 This is actually the real use of the AI solution. Because it’s useless to have the most brilliant AI solution in the company and no one uses it; you also have to measure it actively accordingly: How much is the AI solution used? And if a company now wants to start using AI in its daily work, how should it, in your experience, best proceed?
14:30 So, in our experience, it’s a good idea for them to first look at: Where are employees already using AI? And this is often more than most people expect. So, that’s the big issue with shadow AI. Then you should analyze the respective processes or the potential processes based on structured criteria. And then, in the final step, look at which employees
15:05 Especially when you have implemented the AI or the AI solution, you also need to empower the employees to use the solution and think beyond it. Exactly. Can an AI assessment uncover potential in such cases? Absolutely. That’s why it’s one of the services we also actively offer. We call it AI assessment.
15:33 We call this the AI audit and we sit down with companies to have interviews with the employees and the management to analyze the potential accordingly and then build a suitable roadmap on how the company can best leverage AI in the short, medium and long term. Organizationally, there are also a number of things to consider:,
16:03 Especially with regard to further education and work organization. We already briefly touched on this earlier: If AI is to become a permanent part of processes, not just sporadically, what specific changes do companies need to make? So, I believe that AI can no longer be thought of as something that exists outside of the current working world. And that means that the possibility of further education or training should no longer be limited to a one-time event.
16:32 This is a constant flow of further training that should be made available to your employees. You should create guidelines within the company or guidelines that give employees a sense of direction. So: What is allowed, what is not allowed, especially when it comes to regulatory requirements, in order to avoid risking large fines.
17:04 And on the other hand, it is necessary to transform the corporate culture for AI in a way that corresponds to the way AI is used. Many companies are also employing multipliers in teams who serve as contact points and show employees how AI can be used. And yes, I think we are now in a constant state of change. And that will look different month by month.
17:40 Finally, I have one more question from your perspective: What mistakes do companies most often make when using AI-based process automation? And can that be avoided? Yes, based on my own experience, I think that many companies still approach this process very unstructurally. They bring in tools, maybe have a training course, and hope that the topic will resolve itself accordingly.
18:11 What we see, especially with regard to process automation, is that it is most helpful to really get into these processes. To consider: How do I want to design my processes with AI? How do I empower my employees to work with AI in the long term and how do I measure the impact of AI? And not only before the process and after the process, but continuously after the process.
18:43 Because AI will also change again and again. Henrik, thank you for the informative conversation. It was a lot of fun for me. I hope you had a good time too. We’ll talk again at the next accompio focus topic. Have a good one.
In everyday work, AI manifests itself on two levels. At the individual level, employees use applications for writing, programming, or analyzing. This usage has already taken hold in many companies.
At a process level, the task is more demanding. A business process consists of several work steps, systems, and responsibilities. If AI only handles one of them, its result must reliably reach the next step. It is precisely at these transitions that media gaps, additional testing efforts, or new bottlenecks often arise.
The meaningful use of AI therefore begins with a process analysis. Only when input, desired output, and subsequent processing are clear can it be assessed whether AI, classic software logic, or a combination of both is the appropriate solution.
Tasks that have fixed rules that quickly become too complex are particularly interesting. AI models can learn from examples to determine which inputs lead to which results. This also allows for the processing of unstructured content such as texts or documents.
An IT support ticket can be assigned to a relevant department or person based on its content. A similar principle applies. AI Flow for the classification of incoming requests. The prerequisite is that categories, responsibilities, and further processing are clearly defined.
AI can link questions to internal knowledge and provide a preliminary draft of the answer. For it to work reliably, the knowledge sources must be up-to-date, accessible, and clearly defined in terms of expertise. Sensitive content also requires a controlled environment, such as a central repository. AI knowledge platform for companies.
Relevant information such as amount, date, or reference numbers can be extracted from invoices, forms, or other documents. AI Flow Inbox Extract converts such information into a structured form so that operational systems can process it further.
Anyone who speeds up a process only at one point is likely to be moving the problem to the next step. Henrik Wenck therefore recommends looking at the entire process from a bird’s-eye perspective.
The starting point is the point at which a process is currently stuck. It should be examined there whether automation really saves time or just creates additional testing effort.
An AI model can only work with the information available to it. Missing, inconsistent, or outdated data reduces the reliability of the result.
Before implementation, it must be determined what happens after the AI step. Can the result be directly transferred to another system? Does it need to be released? Who takes over a case when important details are missing?
Automated decision-making requires clear quality limits. A high confidence can allow for direct further processing. In the case of an uncertain result, a human should review and decide.
Not every process step requires AI. Clear rules can often be more reliably represented using conventional software. In practice, the greatest benefit often comes from a combination: AI processes variable content, while deterministic logic controls clearly defined processes.
AI works probabilistically. A result that sounds plausible can be technically incorrect. In long process chains, there is another effect: When several steps each have a certain probability of error, the deviations can increase over the entire process.
Regulatory requirements must also be clarified before implementation. A functioning AI use case is of little value if it processes data in a way that is not permitted within the company. Clear rules and a reliable AI governance vs. shadow AI Therefore, they are part of the technical planning.
„AI can do good work, but it’s hard to bear responsibility.“
Henrik Wenck, Senior VP at accompio AI
Responsibility and relationship management remain with the individual. This applies to internal decisions as well as to the interaction with customers. Employees must be able to prioritize results, to conduct a professional assessment, and to review the consequences of a decision.
An AI training is then helpful if it addresses the specific work routine of the participants. A general introduction can spark interest. For practical application, employees need examples, tasks, and risks from their own processes.
A continuous learning culture is more effective than a one-time training session. AI Academy for employees This creates a common foundation. Additionally, multipliers can serve as contact persons in individual teams and carry forward concrete deployment possibilities.
AI tools can generate a lot of output in a short time. This activity is not yet a sign of productivity. The usefulness of the tool is only determined when a company looks at the process before and after the introduction using the same criteria.
These metrics should be collected continuously. Models, data, and workflows change. Therefore, measuring immediately after implementation is not sufficient. Testing of artificial intelligence In addition to technical quality, stability and behavior under changing conditions also play a role.
The first step is an honest inventory. Companies should record which AI tools employees are already using, for which tasks they are used, and which problems they are intended to solve. Often, more usage is visible than expected.
Subsequently, potential processes are evaluated using uniform criteria. These include the expected benefits, technical feasibility, the data available, the testing effort, and any potential risks. KI Use Case Sprint helps to structure and prioritize ideas.
For a broader assessment, an AI assessment or AI audit is suitable. Interviews with employees and management show where there is already experience and which processes are particularly relevant. This results in a roadmap with short-, medium-, and long-term measures. AI Compass Health Check provides a structured starting point for this.
A common mistake is an unstructured approach. Companies buy a tool, offer training, and expect that usage will develop on its own. However, this lacks common goals, responsibilities, and measurable criteria.
A binding plan is more helpful. It describes which processes need to be changed, how employees can be empowered in the long term, and how the impact of the AI used can be continuously measured. This also includes guidelines for permitted applications, data, and releases. Since AI solutions change quickly, this plan must be reviewed regularly.
AI can, among other things, classify support tickets, assign requests to the appropriate teams, make internal knowledge discoverable, and extract data from documents. The usefulness depends on whether the result can be reliably incorporated into the next process step.
Recurring tasks with clearly defined input and output are particularly suitable, where unstructured data is processed. Before implementation, data quality, downstream steps, and the required testing effort must be evaluated.
A human review is particularly important when confidence is low, sensitive data is involved, or a decision has legal, financial, or personal consequences. Responsibility and professional approval remain with the human being.
Effective training combines the fundamentals with real work processes. It imparts process knowledge, transferable principles, safe handling of data, and the critical evaluation of AI results.
Relevant metrics include lead time, quality, rework, and actual usage. It is crucial to compare the process with the one before implementation and to continuously measure it after launch.
An AI assessment examines existing usage, processes, data, competencies, and risks. Interviews and a structured evaluation result in a prioritized roadmap for short-term, medium-term, and long-term AI initiatives.
Do you want to identify suitable AI use cases, streamline processes, and prepare employees for the new way of working? accompio supports you from the initial assessment to the roadmap, through to the technical implementation and empowerment of your teams.
Henrik Wenck has been working in the field of AI for more than ten years and works as a Senior VP at accompio AI. He helps companies identify suitable AI use cases, implement them, and empower employees to use AI productively.
