Rise of agentic AI: How trust is the key to human-AI collaboration

Source: Capgemini Research Institute; Capgemini Research Institute report (116 pp.) · 2025-07

AI agents constitute one of the fastest-emerging technological trends (see page nine,“What are AI agents and what is agentic AI?”). Unlike traditional AI applications, which often act as tools focusing on narrow, predefined tasks, agentic AI can manage and execute end-to-end processes with less human intervention, evolving from tools to team members and marking a new era in enterprise productivity, efficiency, and growth.

Moreover, trust in AI agents is declining. Only 27% of organizations express trust in fully autonomous AI agents, from 43% 12 months ago. This is potentially a reflection of business reality taking hold after the initial enthusiasm and overconfidence in agentic AI capabilities. Ethical concerns around AI, such as data privacy, algorithmic bias, and the “AI black box,” are prevalent, but few organizations act Nonetheless, in 3 years, organizations expect to have AI agents as members within human-supervised teams. This means that AI agents must be seen not as tools, but as part of the team. Such human-AI agent collaboration enhances the value of AI agents, driving transparency, trust, and positive outcomes. But 61% of organizations report rising employee anxiety about the impact of AI agents on their employment prospects, and over half believe AI agents will displace more jobs than they create. Despite this, fewer organizations are prioritizing reskilling or workforce restructuring.

Introduction. This report offers an overview of the transformative potential of agentic AI and AI agents for large organizations across all industries and sectors. It is primarily intended for senior executives in data and AI functions, and those in strategy, finance, IT, innovation, and risk and compliance, among others, who are involved in the implementation of Gen AI or agentic AI at their organization.

The report will help business executives identify the potential of agentic AI, assess the level of trust and collaboration between humans and AI agents, and the risks and challenges of implementing and scaling agentic AI. The report draws on the comprehensive analysis of a survey of 1,500 leaders (director level and above) across 14 countries. Finally, it offers recommendations for business leaders to accelerate their organizations’ agentic AI journeys.

AI agents are programs/platforms/software that are connected to the business environment with a defined boundary, make decisions autonomously, and act to achieve specific goals with or without human intervention. With the latest advances in reasoning AI models, AI agents are able to break down tasks, “reason” through potential pathways to find solutions to the given problem, try those solutions, and present successful outcomes.

Examples of AI agents include: OpenAI’s Operator, Devin, Manus, Google Gemini Agent Mode, Runner H by H Company.

• Operator by OpenAI can handle a wide range of repetitive browser tasks such as filling out forms, ordering groceries, booking restaurant tables, etc., based on user preferences.1 • Devin, branded as the AI software engineer, is capable of automating software development tasks such as code generation, debugging, and deploying code with minimal human input.2 • Manus is a general AI agent that handles various tasks in work and life such as creating personalized travel itineraries, analyzing documents, conducting research, and generating content and codes.3 Examples of AI/Gen AI assistants include: Microsoft Copilot, Google Gemini, ChatGPT, Le Chat by Mistral AI. Agentic AI is a broader term and includes systems, platforms, practices, tools, and technologies that enable agents to function. Unless otherwise specified, the two terms are used interchangeably in this survey. Agentic systems have been around since well before the current AI boom and can be built using both AI and non-AI technologies. Recent advances in AI, language models, and reasoning capabilities have driven the rise of agentic AI systems.

In essence, AI agents can proactively manage and execute end-toend processes with less human intervention, whereas AI/Gen AI assistants often act as tools focused on discrete, predefined tasks based on explicit prompts.

AI agents are at the center of the global technology narrative in 2025. We at the Capgemini Research Institute flagged AI agents as a top technology trend for 2025.4 • The past year has marked a turning point in automated software agents, with breakthroughs in natural language processing (NLP) dramatically expanding their capabilities not only to answer questions, but to plan, collaborate, reason, execute tasks, and improve continuously.

• Figure 1 shows that the length of tasks (measured by how long it takes a human to complete them) for which AI agents can achieve an 80% success rate has been doubling roughly every 213 days – indicating exponential growth in model capabilities.5 • AI had already surpassed human performance across many tasks, including several previously challenging benchmarks, with only a few exceptions. Even in those areas, the performance gap between AI and humans is shrinking rapidly.6 • Tech giants including Google, Microsoft, Amazon Web Services (AWS), and OpenAI, and enterprise solution providers such as Salesforce, ServiceNow, and SAP are racing to roll out AI agent platforms, tools and frameworks. These range from orchestration protocols, plug-and-play agents that integrate directly with enterprise systems like CRM and ERP, to agent studios for building customized agents.

• As Gen AI models mature and competition intensifies, cost dynamics are also shifting.7 • The inference cost for a system performing at the level of GPT-3.5 dropped over 280-fold between November 2022 and October 2024 (see figure 1), while hardware costs are down 30% annually and energy efficiency has improved by 40% each year.8 • Open-source models are closing the gap with closed models – for instance, DeepSeek’s entry into the AI space sparked significant concern among investors in rival technologies, as it demonstrated impressive capabilities while reportedly using fewer resources and lower costs than traditional AI models.9 • At geopolitical level, advancements in agentic AI are not only a matter of innovation, but one of national capability and influence.

Method. E The above iterations can be repeated in different combinations and until the stated goals have been achieved. The agent utilizes its memory to maintain context, learn from previous iterations or past experiences, and improve its performance over time.

F Agent-to-Agent (A2A) protocols help collaboration between agents, while Model Context Protocol (MCP) aids in accessing external tools. MIT’s Project NANDA is an initiative to develop the foundations of the Internet of AI Agents.

Although AI agents could be set up to operate autonomously, they still function within the scope of execution. Critical situations will require human oversight or decision-making. For more details on these terms, please see the glossary in the appendix.

How do AI agents work?

A AI agents interact with their environment and receive triggers, inputs, or goals from users.

B By employing reasoning models, they can break down the goals into specific actions or steps and prioritize them.

C For task execution, AI agents may access internal data and enterprise tools systems (such as knowledge bases and customer relationship management [CRM]), external tools (such as web search, third-party databases), and interact with other AI agents, or request clarification from users.

D In a multi-agent system, an “orchestrator” agent can break down the larger problem into smaller pieces to be tackled using agents that leverage small language models, significantly reducing the cost and time needed to arrive at an outcome. Guardrails enforce ethical, operational, and safety standards.

Discussion. To put our economic value estimate in context, we benchmark it against estimates from several leading organizations:

• Goldman Sachs predicts that Gen AI will drive a 6.1% rise in US GDP over the next decade.12 By 2028, this translates to around $540 billion in the US.

• IDC forecasts that AI technologies overall will influence 3.5% of global GDP by 2030.13 By 2028, this implies an impact of around $1.9 trillion globally. This encompasses all AI technologies, including agentic AI.

• MIT research estimates that a combination of AI capabilities could automate around one-fifth (slightly over 20%) of value-added tasks.14 This compares well with our estimate that, 20% of all processes on average will be automated by AI agents at Level 3 or higher autonomy over the next three years.

While estimates of AI’s economic value vary significantly, there is general agreement that agentic AI will have significant impact.

Lynn Comp, Head of Global Sales and GTM (AI Center of Excellence) at Intel, says: “AI agents represent a shift from relying solely on LLMs [large language models] to using a variety of specialized models for different tasks. AI has been a mature field for decades, with various modes such as data analytics, machine learning [ML], and deep learning. Now, AI agents can integrate models to provide optimal solutions.”

Anna Kopp, Digital Lead Germany at Microsoft, says: “We have two types of agents: those built by specialists, which users can access in Copilot and those created by users themselves. We currently have multiple pre-built agents. Examples include agents for tech support, employee selfservice, sales and agents as writing coach, travel planner, etc.”

Discussion with Jason Gelman Director of Product Management – Vertex AI, Google Cloud What are your views on the adoption of AI agents and bridging the gap between experimentation and production?

“There are significant challenges in taking agentic AI to production but all major AI providers are converging on the development of reasoning-capable agents, signaling that the ecosystem is maturing. Google has released an Agent Development Kit, and there are major advancements in reasoning models and tool integration, such as Gemini 2.5 Pro and Flash, which now enable faster, more capable agents.”

What is your perspective on improving reliability and safety of AI models?

“While humans are about 97% accurate in dataextraction tasks, AI is now approaching a similar level. However, users expect AI to be nearly infallible. Techniques such as retrieval-augmented generation (RAG) and controlled-context windows reduce hallucinations by anchoring model responses in verified data.”

What do you think about the growing importance of open-source models in the industry?

“The industry is moving toward a mixed landscape, where both open and closed models coexist. Although open-source models offer flexibility, they often come with significant operational overheads, including hardware setup, driver management, and ongoing maintenance, making them impractical for many organizations. As models become smaller and more efficient, on-device AI will grow, further supporting open-source adoption in specific contexts. However, for cloud-based workloads, proprietary models still dominate in performance, reliability, and total cost of ownership [TCO].”

Capgemini Research Institute 2025 “This wave of transformation is not about adding AI agents to workflows – it’s about shifting from isolated use cases to end-to-end processes and rethinking workflows entirely.”

Discussion with Daniel Vassilev Co-Founder and Co-CEO, Relevance AI (a low-/no-code platform that builds AI agents and multi-agent teams) What is the potential of AI agents?

Conclusion. To harness the full potential of AI agents, organizations must move beyond the hype and work toward:

• Redesigning processes to deploy AI agents effectively and reimagining business models • Transforming the workforce and organizational structure to fully onboard a new, agentic workforce as team members • Striking the right balance between agent autonomy and human involvement • Strengthening data and technological foundations to scale AI agents • Ensuring AI agents operate within defined scope of execution,and remain traceable and explainable to earn trust • Developing and integrating ethical AI that address the risks posed by autonomous AI agents.