What AI actually delivers in German companies - a report by the Data & AI Leadership circle of appliedAI & Odgers

New foundation models arrive every week. Benchmarks are broken, demos impress. And yet: ask decision-makers at German companies which process has become measurably faster, cheaper, or better because of AI - and the answer is often evasive.

Not because AI doesn't work. But because most organisations are optimising the wrong things.

Together with Odgers, we have been bringing together AI leaders from German companies in our "Data & AI Leadership Circle" for several years. What this group is grappling with at the start of 2026 is a clear diagnosis: the technology is ready. The question is whether the organisations are.

The pace of AI development since 2024 is generating above all one thing: noise. New models, new tools, new benchmarks every week. For decision-makers in companies, the difficult question is no longer whether models are capable. It is which of the promised productivity gains actually materialise and where the real bottlenecks lie within their own organisation.

The performance, speed and accessibility of models have increased dramatically since 2024. Productivity gains in many companies, however, remain isolated, difficult to measure and operationally fragile. The bottleneck is not model quality - it lies in governance, process integration, security and change management.

appliedAI and Odgers Berndtson have been bringing together a group of AI leaders from German companies (the "Data & AI Leadership Circle") for several years to discuss the current challenges of AI adoption. To make this practical perspective visible, we are publishing the current state of the expert panel here.

This publication summarises eight core observations from this exchange. It is aimed at CxOs, board members and Heads of AI who want to build reliable, regulation-compliant and scalable competitive advantages from AI.

The performance gains of modern foundation models since GPT-4o are substantial. Coding, text, analysis and multimodal tasks are approaching or surpassing human expert performance on benchmarks. At the point where AI outputs permanently outperform manual processing, clinging to old ways of working itself becomes a risk. What matters now is no longer whether models are capable, but how quickly organisations can integrate them safely.

Core thesis: Companies that do not adapt their operating and decision-making model to the pace of model development by 2026 will fall structurally behind - regardless of the platform they choose.

Implications for management:

1. Model selection is tactical. Operating model, governance and integration are strategic.
2. Speed without control disproportionately increases risk — in security, reputation and compliance.

Tools like GitHub Copilot noticeably increase code output, especially for experienced engineers. The bottlenecks shift as a result:

1. QA, integration, testing and compliance continue to determine throughput time.
2. Auto-generated code increases the review and verification workload.
3. Productivity claims of 10 to 30 percent are not substantiated as long as they are based on lines of code or subjective assessments.

Core thesis: AI increases local efficiency, for example in coding. However, the main effort lies in integration into the existing or adapted process environment. Experts estimate this share of total effort at up to 70 percent. The tangible productivity gains therefore apply only to a smaller part of the overall effort.

Further observations:

1. AI is more effective in greenfield environments than in legacy environments.
2. Teams with existing structural or quality problems benefit below average.
3. Testing and quality assurance become more important, not less.
4. Long-term risk: competence erosion through excessive delegation to AI.

A great many companies are experimenting with AI. Only a few are achieving substantial business value. There is a systematic gap between technical capability and operational value contribution. Where the leap succeeds, the impact can be considerable - in individual cases reaching triple-digit millions per year.

Observations from practice:

1. Purely technical or purely strategic excellence is not sufficient.
2. For companies using AI not only for assistance tasks but for process automation, the bottleneck is rarely the AI or technology resource. It lies in the business unit. Based on the experts' experience, each AI resource invested requires roughly two process or domain experts to achieve a process improvement effect.
3. End-to-end responsibility for an automation project (idea, operations, audit) is rarely clearly anchored.

Core thesis: AI adoption is primarily a transformation project, not a technology project.

Agents significantly extend classical business process automation. At the same time, regulation and liability limit the degree of autonomy. Leading companies are therefore working toward a target picture in which humans and agents map a continuous process via standardised handover points.

1. Bot-to-bot communication
2. Generative process design (e.g. Pega, Regrello)

Core thesis: Process architecture is being redesigned right now - with explicit human oversight as a condition, not an option.

Uncoordinated AI initiatives are increasing rapidly.

1. Demand for the latest tools is high among technology-enthusiastic parts of the workforce. For strategic and security reasons, companies often initially grant access only to a small group within a sandbox.
2. Employees then independently install tools, plugins and local models.
3. These parallel developments lead to duplicated effort, security vulnerabilities and brand risks.
4. Increasing autonomy through agents, plugins and complex workflows intensifies the risk.

Core thesis: None of the panel companies make new technology freely available for unguided experimentation. Without a central register, clear approval processes and defined accountabilities, AI shifts from a lever for efficiency to a governance problem.

What is needed:

1. a central AI portfolio or PMO
2. clear ownership per AI application and agent
3. binding rules for plugins, local installations and external tools

Internal AI platforms are vulnerable to subtle manipulation.

1. Changes to system prompts can permanently shift model behaviour.
2. Missing audit trails make root cause analysis nearly impossible.

Core thesis: Prompt security is cybersecurity, not a UX detail.

Best practices:

1. protected system prompts
2. role and access management
3. monitoring, versioning and audit trails

The debate around cloud and data sovereignty is highly politicised. In practice, a more differentiated picture emerges.

1. Heavily regulated industries need regional storage or on-premises approaches.
2. While attention to AI is high, geopolitical dependencies in collaboration and communication software are largely accepted.

Core thesis: Sovereignty cannot be pinned to the AI stack alone. It must be considered across the entire application landscape.

In the experts' assessment, AI amplifies existing performance differences. This creates concrete tasks:

1. Expectations and targets must be redefined fairly.
2. Metrics must be chosen so that they do not create false incentives - for example through pure output measurement.
3. Entry-level positions must offer development paths so that judgement capacity is preserved within the company.

In Germany, early involvement of the works council is a critical success factor.

Core thesis: AI productivity without a qualification and performance strategy is not sustainable.

Many companies are at a turning point at the start of 2026. The technology is ready. The open question is whether organisations are prepared to consistently transform their decision-making, process and governance logic alongside it.

AI accelerates the realisation of ideas, for example in prototyping. In the medium term, it puts existing business models in question. Process improvement and automation with AI are necessary for this, but not sufficient. The experts agree that this step must be taken. What matters is the sequence: first revise the process - then decide on the right AI tool.

Competitive advantage does not come from the best model, but from its controlled integration into processes and accountabilities.

Ready to accelerate your AI transformation?

Learn more about the appliedAI Companion Partnership Program - and how a structured AI maturity assessment gives your transformation the clarity, direction, and momentum it needs.