We keep up with the reading: studies from universities, research institutes and industry associations on AI agents, on collaboration between humans and AI, and on AI in small and medium-sized enterprises. Each entry briefly summarised, with a link to the original source.
Business
AI agents in organisations: first practical experience
The OECD asked experts from 25 organisations in 11 countries about their use of AI agents. None of the organisations deploys agents with unrestricted autonomy; a human reviews consequential steps before they are taken. Still open are standards for evaluating longer workflows, traceability in multi-agent processes and new security risks.
OECD · OECD Artificial Intelligence PapersOriginal source ↗
Sovereignty
Economists warn of Europe’s dependence in AI
A group of leading economists and AI experts, including ifo President Clemens Fuest, presents a strategy for Europe. According to the group, Europe holds around 5% of global AI data-centre capacity; the target is 15% by 2030. Added to this are secure supply chains and European expertise in testing and auditing AI.
ifo Institute · Press release on the Transformative AI strategyOriginal source ↗
Multi-agent
Larger groups of agents change collective behaviour
The study examines how groups of language-model agents behave depending on their size. The finding: in interaction, the biases of individual models can be amplified, new ones can emerge, or model preferences can even flip. The effects depend non-linearly on group size – so testing individual agents is not enough.
City St George's, University of London · IT University of Copenhagen · UPC Barcelona · PNAS 123(34)Original source ↗
Labour market
Companies tend to expect AI to lower salaries for career starters
In an ifo survey of more than 3,000 companies that use AI, around half expect falling salaries for employees with less than five years of professional experience; for experienced staff the figure is about 40%. For university graduates, by contrast, the firms see hardly any downward pressure, and in some cases even opportunities for higher salaries.
ifo Institute · ifo Business Survey (June 2026)Original source ↗
Labour market
One in four establishments uses generative AI
Just under 25% of establishments in Germany use generative AI – almost five times as many as in 2023. Of these, 90% use freely available applications and only 16% use specially trained models. Basis: the IAB Establishment Panel, covering around 15,000 establishments.
IAB · Institute for Employment Research · in GermanOriginal source ↗
Sovereignty
One third use AI at least once a week
58% of people in Germany use AI, and 48% of those in work use it on the job. 72% are concerned that Germany is too dependent on US providers when it comes to AI.
Bitkom Research · in GermanOriginal source ↗
ReflectionPreprint · not yet peer-reviewed
Self-correction needs control
The authors describe the self-correction of language models as a control loop: it only helps if a model fixes more errors than it introduces. In tests with seven models, only some improved – a “verify first, then correct” approach reduced new errors markedly.
arXivOriginal source ↗
Overview
AI Index Report 2026
Stanford University’s annual report on the state of AI – from research and technical performance through the economy to public opinion, with questions about environmental costs, transparency and distribution.
Stanford HAIOriginal source ↗
Business
Understanding and applying AI agents
With the “Agentic Level”, the study introduces a framework that companies can use to assess whether a system is an AI agent and what risks and potential it brings. Recommendation: start small, involve employees, establish clear governance.
Fraunhofer IAO · IAIS · HNFIZ · in GermanOriginal source ↗
SMEs
One in five SMEs uses AI
Around 20% of small and medium-sized enterprises – some 780,000 businesses – use AI, five times as many as in 2016–2018. Those with a digitalisation strategy use AI considerably more often (35% versus 19%). Data and data management are decisive.
KfW Research · Fokus Volkswirtschaft No. 533 · in GermanOriginal source ↗
Humans and AI
Human–AI teams decide better than either alone
Researchers at the Max Planck Institute show that combining human and machine judgements can outperform both partners – even when the AI on its own performs better. The key is to weight the judgements by their confidence.
Max Planck Institute for Neurobiology of Behavior · Patterns 7(2) · in GermanOriginal source ↗
Multi-agent
Why multi-agent systems fail
The researchers analysed more than 1,600 runs from seven widely used multi-agent frameworks and found 14 failure patterns in three groups: system design, coordination between agents, and verification of results.
UC Berkeley · arXivOriginal source ↗
Humans and AI
When humans and AI are better together
A meta-analysis of 370 results from 106 studies shows: in decision-making tasks, human–AI teams performed worse on average than the best AI alone – in creative tasks, by contrast, better than humans or AI on their own.
MIT · Nature Human BehaviourOriginal source ↗
Summaries by nasGO. In each case the linked original source is authoritative. Preprints have not yet been peer-reviewed.
