Hamburg – November 02, 2026 -- Two-thirds of companies in Germany, Austria and Switzerland move fewer than a quarter of their AI pilot projects into regular operations, according to a study by Lünendonk & Hossenfelder in cooperation with Sopra Steria. Only 1% have integrated AI agents across more than one core business process.

Sopra Steria concludes that strict internal governance, data quality and compliance structures, often seen as brakes on innovation, are what allow AI agents to scale.

Only 19% run AI agents productively in individual task areas

Of the 141 companies that answered the question, 66% bring fewer than 25% of AI pilots into production. Some 58% still test AI agents in pilot projects and 18% are exploring the technology. Another 19% use agents productively in individual task areas, 1% run them across several central business processes, and 4% have not addressed the topic.

The study cites missing technical prerequisites and data connections, security and data protection concerns, a shortage of specialist staff, and poorly coordinated parallel AI projects as the causes. The survey covered 180 IT chiefs, IT department heads and business-unit managers at large mid-sized companies and corporations. Telephone interviews ran from April to June 2026.

65% keep humans in charge of high-risk decisions

In high-risk processes, 65% of respondents let agents give recommendations only, with a human making the decision. Over the next two years, 98% plan to invest in data quality and data integration. Another 94% each prioritize platforms and tools for building and orchestrating AI agents, and IT infrastructure.

Agents are mostly expected to complement existing software, not replace it: 73% consider this very likely. A further 91% rate it likely or very likely that agents will act as a higher-level control layer orchestrating existing systems. That requires access to many applications and datasets, which raises demands on access rights, approvals and traceability, Sopra Steria says.

Banks and insurers already have the governance templates

Martin Stolberg, Head of AI at Sopra Steria in Germany and Austria, argues that the AI model matters less than who is accountable for an agent's decisions and how well it integrates into the organization. According to Stolberg, many pilots stall because nobody has defined which data an agent may use, what it may decide alone and who reviews its output.

Without fixed rules, a company must clarify data access, liability and review from scratch for each new agent. With them, the check becomes routine. Two of the hurdles named in the study, security and data protection concerns and uncoordinated parallel projects, are essentially governance questions.

Banks, insurers and public authorities built such structures under supervisory and legislative requirements. Sopra Steria says these structures do not transfer automatically. Companies must adapt them for self-acting systems through tightly limited access rights, approval tiers and gapless logs.

96% expect efficiency gains, but only from broader autonomy

In the survey, 96% expect AI agents to deliver higher efficiency and lower costs, and 90% expect more speed. Sopra Steria says these effects materialize only when agents work across many processes and are allowed to decide themselves. It expects companies to grant that latitude once they can trace and control agent behavior.

Cosmo Consult, HyPlus, Materna, Reply and Woodmark were additional study partners.