In many medium-sized companies, the first steps toward adopting artificial intelligence (AI) follow a similar pattern: workshops, AI task forces, or strategy sessions generate numerous ideas. What happens next determines long-term success. The challenge rarely lies in coming up with ideas, but rather in selecting the right ones and putting them into practice.
Companies often jump right into projects that are strategically promising but time-consuming to implement: new data-driven business models, complex automation spanning multiple systems, or company-wide platform solutions. Such projects require time, resources, and often external support. In small and medium-sized enterprises (SMEs), these projects tend to drag on, internal capacity is lacking, and initial results are slow to materialize. We discussed alternative approaches with Leon Rocca, AI Solution Consultant at AlgebraX, a member of the Baden-Württemberg AI Alliance: What does an AI project in an SME need to be successfully adopted within the company, and where does the time go in such a project?
An Interview with Leon Rocca:
Mr. Rocca, how do you see this playing out in practice: What happens after the ideation phase, and what often causes implementation to fail?
Guidelines Based on Real-World Experience
Where Time Goes in AI Projects
10–30%The technical implementation is often the smallest part. Most of the time is spent understanding the problem and analyzing the data.
Hatching: typical range in the project.
What initiatives should companies start with?
With a problem that can be clearly defined. A solution doesn’t have to be perfect from the start. You can begin with limited time, limited funds, incomplete data, and without extensive subject matter expertise. If the problem is clear and there’s a strategy behind it, a solution will help employees even if it doesn’t answer every question at first.
Why isn't the focus of an AI project on the technology?
Because 90 percent of an AI project often has nothing to do with AI at all. First, you have to lay the groundwork: identify the actual problem and talk to the people who experience it every day. That’s the only way you’ll even come up with the idea of solving a problem with AI. That’s where most of the time is spent, along with preparing the data. The technical implementation takes relatively little time, partly because development moves so quickly these days thanks to AI.
When do employees adopt an AI solution?
If they realize early on that the solution is taking work off their hands. That’s why it’s important to reach a quality threshold right from the start—one where employees realize, “This is making my job easier.” If that threshold isn’t met, they might use the solution for a few days and then give up because it isn’t helping them.
On top of that, the solution must start where employees are already working. To achieve this, we integrate AI workflows into existing systems—such as Outlook or Gmail, Google Workspace, SharePoint, Copilot, or Open WebUI—as well as into custom systems and company-specific databases. Employees shouldn’t have to copy anything into an additional tool or formulate anything themselves. The solution must come to the employees, not the other way around. That is a crucial factor.
To that end, I recommend adopting a mental model right from the start: Am I making this person’s life easier with what we’re doing? Ideally, this works out so that employees are happy to use the solution because it benefits them personally in their day-to-day work.
Do you rely on low-code—that is, tools that don't require extensive programming—for AI projects? Why?
It’s important to me that a project can be replicated. We want to teach people how to use AI on their own. I don’t want to give companies the impression that we’re the experts who’ll lock ourselves away in our little office for a month and build something for them. Anyone who needs an AI solution should eventually be able to build it themselves. We also want to be able to hand off projects. For a quick proof of concept and a smooth handoff to the customer, low-code tools like n8n are ideal. In the Microsoft ecosystem, Copilot Studio also works well for this. If the project becomes more in-depth or complex, you may need to resort to custom software development, which can be done quite quickly these days.
What role does data protection play in AI projects at small and medium-sized businesses, including from the customers' perspective?
A big one. Before getting started, it should be clarified where the data is located, which language models will be used, and whether customers consent to the processing. It must also be determined whether and to what extent the data must be anonymized before processing.
What does an AI application in a business need to ensure that it continues to provide reliable support even months later?
It needs to be maintained. Simply letting an application that relies on corporate knowledge run on its own for years on end doesn’t work. The more data that comes in, the harder it becomes to find the right information, and eventually, the quality of the responses declines. That’s why, at regular intervals—for example, quarterly or semiannually—someone with expertise should review the database and remove anything that’s outdated or incorrect. It’s the same with any software: the world changes, and new data is added. A project like this requires a regular improvement cycle and must be actively managed.
How did that play out specifically in your project with etg24?
AlgebraX is already collaborating with etg24, a provider of a cloud platform for digital real estate management. When etg24’s customer volume increased and the employees’ knowledge was not documented in a way that could be easily accessed, they reached out to us. Together with the IDEV Institute at Karlsruhe University of Applied Sciences, we developed an AI agent that reads incoming customer service inquiries and suggests a response. Employees review the responses, adjust them as needed, and send them out themselves. The agent learns from the responses that are sent.
We started with an outdated manual and 412 completed customer communications. About 60 to 70 percent of our time was spent understanding the problem and talking with employees, about 20 percent on data and verification, and only about 10 percent on the technical aspects.
At first, the employees were skeptical because a previous AI initiative for automated documentation had done little to lighten their workload. That’s why the suggested response appears directly in their email inbox; often, all that was needed was to tweak a single sentence. The database is located in Germany and is GDPR-compliant; all customers have consented to the processing of their data. After a brief handoff, the lead developer migrated the solution to the company’s own systems.
And what does an AI project in an SME need in order to be accepted?
Based on my experience, there are four key things. First, a clearly identified problem that you address together with the people who experience it every day. Second, a solution that employees recognize from the start as something that helps them—otherwise, they’ll quickly give up. Third, a solution that integrates into the employees’ existing workflows. And fourth, the realization that an AI solution never solves 100 percent of the problem. It doesn’t have to be the perfect solution at the push of a button.
About Leon Rocca
Leon Rocca is an AI Solution Consultant and COO/CFO at AlgebraX. He studied International Management (B.A.) and Technology Entrepreneurship (M.A.) at Karlsruhe University of Applied Sciences, with a focus on AI agent systems.