AI Agents

From answer machine to digital colleague

A chatbot answers. An AI agent acts [1]. The difference sounds subtle but fundamentally changes how companies use artificial intelligence. Where generative AI remains limited to single-turn interactions and content generation, AI agents go a decisive step further: they autonomously pursue goals, plan their steps, and solve problems continuously [1]. The transition from generative to agentic AI expands the spectrum to include autonomous goal pursuit, tool use, multi-step planning, and active environment interaction [2]. This shifts AI's role in the enterprise: from a tool you query to a colleague that completes tasks independently [5].

2026 marks the tipping point [1]. Gartner predicts that 40% of all enterprise applications will contain AI agents by year's end [4]. The market is growing rapidly: from $7.8 billion to over $52 billion by 2030 [4]. The question for companies is no longer whether, but how they deploy AI agents.

What exactly is an AI agent?

AI agents are programs that carry out tasks "autonomously and step by step" and make simple decisions [1]. Unlike traditional chatbots that respond to individual queries, agentic systems operate in loops: they perceive their environment, store relevant information in memory, plan next steps, and execute actions [7]. This Perception-Memory-Action pattern forms the core of every agent architecture [7].

Technically, modern AI agents are built on three layers: core components such as perception, memory, and tool use, a cognitive architecture with planning mechanisms, and learning paradigms that continuously improve behavior [7]. These systems access databases, call APIs, search the internet, and control other software [2]. They become particularly powerful when multiple specialized agents collaborate. Such multi-agent systems significantly outperform individual agents on complex tasks [13]. Between Q1 2024 and Q2 2025, Gartner recorded a 1,445% increase in inquiries about multi-agent systems [4].

Three collaboration patterns have emerged: In the chain topology, agents work sequentially, with each picking up where the previous one left off. In the star topology, a central agent coordinates specialized sub-agents. The mesh topology enables decentralized exchange between peer agents [7].

Where companies deploy AI agents

Use cases range from customer service to autonomous data processing.

Customer Service: AI agents can independently handle up to 80% of standard tickets [1]. Insurer Mapfre Insurance demonstrates how a hybrid model works: agents handle routine claims while humans manage sensitive customer communication [3].

Sales and Marketing: Agents research leads, qualify prospects, and create personalized offers [1]. They don't follow rigid rules but adapt their strategy to the specific context [1, 2]. Real-time KPI monitoring can also be automated through agentic systems [1].

Back Office and IT: Document verification, invoice management, and real-time KPI monitoring are among the strengths of agentic systems [1]. Successful companies like Toyota, HPE, and Dell use AI agents not to automate existing workflows but for fundamental process redesign [3]. Moderna even created a new leadership role that unifies HR and IT to manage "workforce planning regardless of whether it's a person or a technology" [3].

Vertical Specialist Agents: Instead of universal models, domain-specific agents for individual industries are increasingly emerging [5]. In healthcare, NVIDIA and GE Healthcare are collaborating on autonomous diagnostic imaging systems [5]. Hippocratic AI offers AI nurses for $10 per hour, compared to a median hourly wage of around $43 for human professionals [5].

Self-Healing Data Pipelines: Agents autonomously monitor the health of data pipelines and repair issues like schema drift using reinforcement learning [5].

Currently, 30 agentic AI systems are in production: 12 chat applications with agentic tools (including Claude Code, ChatGPT Agent, and Manus AI), 5 browser-based agents, and 13 enterprise workflow agents (including Microsoft Copilot Studio and ServiceNow Agent) [8]. 20 of these 30 systems already support the Model Context Protocol (MCP) for standardized tool integration [8].

The market in numbers

Adoption rates paint a clear picture: 79% of organizations report AI agent implementations, but only 34% have achieved full deployment [6]. 96% of IT leaders plan to expand their agent systems, and 88% of executives are increasing their AI budgets specifically because of agentic AI [6]. 80% of organizations surveyed by Capgemini plan integration within the next one to three years [5].

The projected return on investment averages 171%, reaching 192% for US companies [6]. 66% of companies report measurable productivity gains, and up to 70% cost reduction is achievable through workflow automation [6]. Analysts estimate the annual GDP contribution of AI agents at $2.6 to $4.4 trillion by 2030 [6]. Long-term, the market is expected to grow at a CAGR of 43.84% to $199 billion by 2034 [6]. 66.4% of organizations already use multi-agent system architectures [6].

What can go wrong

Impressive as the numbers sound: according to Gartner, over 40% of agentic AI projects will fail by 2027 [3]. Only 11% of companies actually use AI agents in production, while 38% are piloting and 42% are still developing their strategies [3]. 51% of AI-using organizations have already experienced negative consequences [6].

The main reason for failure: companies layer AI agents onto existing workflows. Intel puts it bluntly: "Don't simply pave the cow path" [3]. Success requires fundamental process redesign, not mere automation of existing processes [3].

A growing security gap compounds the problem. The AI Safety Index rates seven leading AI companies: Anthropic leads with a C+ grade (2.64 out of 5), followed by OpenAI with C (2.10) and Google DeepMind with C- (1.76) [11]. No company scores better than D on "Existential Safety" [11]. Of 30 agents deployed in production, 25 publish no internal safety results, and 133 of 240 safety and evaluation fields contain no information at all [8]. Agent complexity doubles roughly every seven months, yet safety measures are not keeping pace [12].

Prominent researchers warn of four specific risks with highly autonomous agents: security risks from unpredictable action combinations, cascading data breaches through cross-platform access, automation bias from human-like presentation, and error propagation across multiple action surfaces [10]. The researchers' recommendation: semi-autonomous systems with human involvement offer a more favorable risk-benefit profile than fully autonomous agents [10]. SIPRI adds another dimension: the largely unexplored risk of agent-to-agent interaction [9]. The 2010 Flash Crash serves as a historical parallel, when high-frequency trading algorithms caused a trillion-dollar market crash within minutes [9]. SIPRI therefore calls for secure testing environments, unique agent identifiers, and a "social contract" for interaction guidelines [9].

How to get started

The IHK recommends a pragmatic pilot approach: define a use case, test it, then scale [1]. The barriers to entry are low today because established tools already offer agent functionality [1]. Three factors determine success:

Ensure data quality. 48% of organizations cite data discoverability and 47% cite data reusability as key challenges in implementing AI agents [3]. Without clean, accessible data, even the best agents run into dead ends.

Rethink processes. Successful organizations don't automate old workflows. They redesign processes from the ground up, tailored to the capabilities of agentic systems [3]. Deloitte recommends microservices architectures with specialized agents close to data sources and FinOps for token-based cost management [3].

Ensure GDPR (DSGVO) compliance. For German mid-sized companies in particular, data protection is a critical success factor when adopting AI agents [1]. Those who factor in compliance early avoid costly rework later.

AI agents are no longer a distant future technology. They are already changing how companies work. The decisive question is not whether organizations will deploy them, but whether they are ready to fundamentally rethink their processes.

References

[1] Lerchl, M. (2025). "AI Agents and Agentic AI: Benefits for Companies." *IHK Regensburg für Oberpfalz / Kelheim*. https://www.ihk.de/regensburg/digitalisierung/kuenstliche-intelligenz/ki-agenten-und-ihr-nutzen-fuer-unternehmen-6706606

[2] Schneider, J. (2025). "Generative to Agentic AI: Survey, Conceptualization, and Challenges." *arXiv Preprint*. https://arxiv.org/abs/2504.18875

[3] Rowan, J., Mittal, N., Patwari, P. & Burns, E. (2025). "The Agentic Reality Check: Preparing for a Silicon-Based Workforce (Tech Trends 2026)." *Deloitte Insights*. https://www.deloitte.com/us/en/insights/topics/technology-management/tech-trends/2026/agentic-ai-strategy.html

[4] Chugani, V. (2026). "7 Agentic AI Trends to Watch in 2026." *MachineLearningMastery.com*. https://machinelearningmastery.com/7-agentic-ai-trends-to-watch-in-2026/

[5] Dilmegani, C. (2026). "10+ Agentic AI Trends and Examples for 2026." *AIMultiple*. https://aimultiple.com/agentic-ai-trends

[6] Arcade.dev Team (2025). "30 Agentic Framework Adoption Trends: Enterprise Investment, Market Growth, and Implementation Success Rates." *Arcade.dev Blog*. https://arcade.dev/blog/agentic-framework-adoption-trends

[7] Arunkumar V, Gangadharan G.R. & Buyya, R. (2026). "Agentic Artificial Intelligence: Architectures, Taxonomies, and Evaluation of LLM Agents." *arXiv Preprint*. https://arxiv.org/html/2601.12560v1

[8] Staufer, L. et al. (2026). "The 2025 AI Agent Index: Documenting Technical and Safety Features of Deployed Agentic AI Systems." *arXiv Preprint*. https://arxiv.org/html/2602.17753v1

[9] Boulanin, V., Blanchard, A. & Lopes da Silva, D. (2025). "Before It's Too Late: Why a World of Interacting AI Agents Demands New Safeguards." *SIPRI*. https://www.sipri.org/commentary/essay/2025/its-too-late-why-world-interacting-ai-agents-demands-new-safeguards

[10] Mitchell, M., Ghosh, A., Luccioni, A.S. & Pistilli, G. (2025). "Fully Autonomous AI Agents Should Not Be Developed." *arXiv Preprint*. https://arxiv.org/abs/2502.02649

[11] Future of Life Institute (2025). "2025 AI Safety Index (Summer)." https://futureoflife.org/ai-safety-index-summer-2025/

[12] International AI Safety Institute (2026). "International AI Safety Report 2026." https://internationalaisafetyreport.org/publication/international-ai-safety-report-2026

[13] Guo, Chen, Wang et al. (2024). "Large Language Model based Multi-Agents: A Survey of Progress and Challenges." *arXiv Preprint*. https://arxiv.org/abs/2402.01680