AI that genuinely works.

Start with a well-scoped use case that involves high manual effort and clearly measurable value — not with a large, vague innovation project. The most reliable early wins usually lie in the back office, for instance in document processing or knowledge assistants, as well as in customer service. The key is to define KPIs upfront, so that after a few months you can judge whether scaling is worthwhile.

The costs go beyond pure tool licences, because onboarding, customisation and training all need to be budgeted for. For smaller companies, monthly tool costs often range from 50 to 200 euros plus a one-off setup, while for mid-sized companies they tend to run into four figures per month. Deliberately build in a buffer — around a third of companies using AI report costs significantly higher than expected.

Research shows that AI changes tasks far more often than it replaces entire jobs — in Germany, only a small percentage of jobs are estimated to be fully automatable. In practice, roles shift towards steering, reviewing and refining AI outputs. The concern is nonetheless real, which is why honest communication, training and bringing your team along from the very start are decisive.

Not necessarily: rule-based automation requires no historical data, and for many knowledge and language applications, pre-trained models plus your own documents via Retrieval-Augmented Generation (RAG) are sufficient. Often the real problem is not a lack of data, but that existing knowledge is scattered, unstructured and locked in individuals' heads. The rule of thumb here: a small, clean dataset beats a large one full of inconsistencies.

In our experience, integration into existing systems is the biggest practical hurdle — more important than the intelligence of the model itself. What usually succeeds is a process-oriented approach: embedding AI where your people already work, for instance in the CRM or ticketing system, rather than creating an isolated bolt-on tool. Clean interfaces, clear permissions and secure access to production systems are the most demanding part.

This uncontrolled use of personal tools — often called shadow AI — is widespread and a genuine data protection and security risk, as confidential data can leak out. Bans alone rarely help; it is far more effective to provide a secure, governed alternative that your teams are happy to use. Combined with clear policies and training, you thereby turn a risk into a controlled productivity gain.

In the free or Plus version you should not enter personal or confidential business data, since there is no data processing agreement in place and you risk a GDPR breach. Enterprise and API offerings provide a data processing agreement and exclude the use of your data for training. A residual legal risk under the US Cloud Act nonetheless remains — for highly sensitive data, a European, sovereign solution is the safer route.

Prohibited practices and the AI literacy obligation under Article 4 have applied since February 2025, and the obligations for general-purpose AI models since August 2025. The central high-risk obligations have been deferred via the Digital Omnibus package, but are expected to come into force in due course — so preparation should already be under way. In concrete terms: classify systems by risk category, document them, ensure human oversight and train your staff; fines run up to 35 million euros or 7 percent of global annual turnover.

GDPR-compliant means that the formal requirements are met: the data processing agreement, the legal basis, transparency and documentation are all in order. GDPR-secure goes further and means that the infrastructure technically permits no access by third-country authorities — so the US Cloud Act cannot take effect in the first place. With a US provider you can achieve compliance, yet a residual legal risk remains; true security requires a solution governed entirely by European law, because data residency alone is not data sovereignty.

For standard tasks, buying is usually faster, cheaper and lower-risk — purchased and partner-led projects are in practice far more successful than purely in-house builds. Building in-house pays off where AI is a genuine differentiator or where sensitive data demands internal control. The pragmatic middle ground: buy standard models and infrastructure, and build only the company-specific logic, data and integrations yourself.

RAG keeps your knowledge external and updatable and is the preferred choice for most enterprise scenarios, because it is more secure, more scalable and more cost-effective — especially with frequently changing data. Fine-Tuning embeds knowledge and behaviour directly into the model and is suited to stable, highly specialised domains or lean specialist models. A hybrid path is well proven: first prove the value with RAG, then fine-tune selectively for the most valuable use cases.

The transition from pilot to production is the most critical moment — a large share of AI projects are discontinued after the pilot phase, usually due to poor data quality, unclear business value or a missing scaling strategy. Successful scaling needs an AI operating model: defined accountabilities, a clean data and integration foundation, governance built in, and KPIs set early. Only what demonstrably creates value gets scaled — not every technically interesting idea.