Guide
The ROI Problem
AI is now everywhere and a hot topic in every company. In fact, however, only 60–80% of companies that use AI see a real return on investment, according to studies by McKinsey in its “State of AI” report from late 2025. But how can your average company use AI strategically to save real time without simply increasing cloud overhead unnecessarily? As a digital agency, we’ve been working with ML and AI long before GPT-3 became the talk of the town. Now we’re offering a glimpse into how we work with AI and automate processes long term.
But first, let’s clearly distinguish between traditional automation and AI-driven automation. While the basic idea behind both approaches is the automation of repetitive tasks, there is nevertheless a crucial difference. In the traditional approach, the process operates solely on the deterministic “if-then” principle. The same result can only be achieved under identical conditions.
AI automation, on the other hand, uses machine learning to process unstructured data, learn from context, and make decisions independently.
But let’s get right to the key advantages and challenges. The competitive advantage in AI lies not in the model itself, but in the system built around it. Anyone can rent the same model, switch between them, and choose the one they want. The real advantage only emerges when the foundation is solid.
The most sensible way to break this down is into four building blocks: a knowledge layer that seamlessly integrates data and makes it searchable; An orchestration layer that directly recognizes intent, refines the query, selects the right source and model, retrieves context, and verifies itself through evaluations; An agent layer that handles multi-step tasks while adhering to permissions and rules. And an enterprise memory that records what works and what doesn’t.
In reality, this simply means that developing Google Apps Scripts and using expensive or new AI models is no longer and should no longer be the unique selling point of process optimization agencies. A process that requires over 150K tokens per system prompt, breaks down weekly, and has to be triggered manually is not automation It’s an expensive experiment.
FUF Insight
How to Build AI Automations the Right Way
To avoid exactly that, it is crucial to set up a system that rests on a solid foundation. The model is interchangeable infrastructure; the advantage lies in the runtime, more specifically in intent detection, retrieval, tool selection, memory, evaluations, and model routing. Prompting is not a permanent skill but rather moves into the background: One agent writes the prompt, while another executes it. The key issue is the gap between capability and the operating model: Agents increase capability, but if permissions, review loops, and accountability remain trapped in the old operating model, shadow automation results—rapid execution without accountability.
This leads to a fixed sequence: first, connect the knowledge layer; then, establish governance before scaling; next, evaluate the system in production; and finally, scale the agents. An agent is a tool that acts on behalf of an employee; therefore, responsibility is not reduced but shifted to the person who deployed, authorized, and supervised it.