Every large enterprise now has access to roughly the same foundation models, the same agent frameworks and the same cloud platforms. That creates an uncomfortable paradox for anyone who promises that agentic AI is a source of competitive advantage: if every competitor can buy the same agents, where does the advantage come from?

My answer, and the thread running through my doctoral research, is that the advantage does not sit in the technology. It sits in the cognitive transformation of the enterprise: the way an organisation rebuilds how it senses, decides and learns once autonomous agents become part of its daily work.

My position: at the intersection of technology and business

I work between two worlds that often talk past each other. My background is (among others) in Enterprise Architecture for Data and AI, where the questions are about platforms, integration and what systems can actually do. My research sits in strategic management and organisational theory, where the questions are about value, capabilities and competitive advantage.

Engineers tend to ask whether agents can do a task. Executives ask whether it pays off. The most important questions sit in between: how a technical capability becomes an organisational one, and why some firms turn the same technology into lasting advantage while others end up with expensive pilots.

That is where I position my research, and it rests on three convictions:

1. Technology alone does not create advantage. What competitors can buy, they can also copy.
2. Organisations are not just environments for AI. They are actors with their own routines, power structures, learning and resistance, and they shape what AI becomes.
3. The enterprise itself is the unit of transformation. Not the model, not the use case, but the organisation as a cognitive system.

The cognitive transformation of the enterprise

For decades, enterprise automation followed a script. Someone programmed every workflow and every exception in advance, and the system executed it. Agentic AI breaks with that logic: agents plan towards goals, break down tasks, decide under uncertainty and adapt as conditions change. Acharya et al. (2025) describe this as a qualitative leap rather than an incremental one.

I call the organisational side of this shift cognitive transformation. Every enterprise is, in a sense, a cognitive system. It perceives its markets, reasons about options, makes decisions, remembers what worked and learns from what did not. Until now, almost all of that cognition happened in people’s heads and in meetings. Automation only executed the results.

When autonomous agents enter the picture, cognition itself starts to be shared between humans and machines. The table below sums up what changes.

	Rule-based automation	Cognitive autonomy
Logic	Predefined workflows	Goals, planning and adaptation
Role of the system	Tool that executes	Colleague that decides within limits
Where cognition sits	In individual people	Distributed across humans and agents
What must be managed	Processes and exceptions	Roles, boundaries, trust and accountability

Multi-Agent systems make this especially visible. Instead of one generalist assistant, several narrowly scoped agents coordinate like a team of specialists. In enterprise benchmarks, Shu et al. (2024) report goal success rates up to 70% higher for multi-agent collaboration than for single agents, and industry teams report that agents perform best when each has a clearly defined role (Crawford et al., 2024).

The theory of agents is not new; Wooldridge and Jennings formalised it in 1995. What changed is the substrate. Large language models finally made those ideas practical, and the question moved from “can agents work?” to “how does an enterprise reorganise itself around them?” That is a socio-technical question, not a technical one (Ali et al., 2024).

The commoditisation problem

Strategy research has a well-tested lens for this: the Resource-Based View. Barney (1991) argued that sustained advantage comes only from resources that are valuable, rare, inimitable and non-substitutable (VRIN).

Apply that test to agentic AI and most of it fails at the second hurdle. A foundation model licensed through an API is valuable, but it is not rare. An open-source orchestration framework is useful, but anyone can copy it. The algorithms are, at first glance, generic.

There is a second problem. The Resource-Based View tells us which resources create advantage, but says little about howthey are put to work (Melville et al., 2004). For agentic AI, that “how” is exactly where the value lives. So if advantage exists, it must sit somewhere other than the technology itself: in the data, the processes, the people and the organisational context that the agents are woven into.

Why today’s adoption models fall short

The research on enterprise AI adoption is large, but fragmented. Each stream looks at one level of the problem and assumes the others away.

• Technology-Organisation-Environment (TOE) predicts whether firms adopt. Paiva’s (2024) meta-analysis of 12 studies (N = 3,398) found seven of eight factors significant, with organisational readiness and vendor partnerships weighing most. But TOE does not explain how adoption turns into capability, let alone advantage.
• UTAUT explains why individuals accept a technology through performance expectancy, effort expectancy, social influence and facilitating conditions (Venkatesh et al., 2003). It was built for tools that wait for instructions, not for systems that act on their own.
• Agent research in computer science is technically sophisticated, but tends to treat the organisation as a mere operating environment rather than an actor with its own learning, politics and resistance.

Even the evidence disagrees in places. Paiva’s meta-analysis found complexity to be the only non-significant adoption factor, while AL-khatib (2023) found it to be a clear barrier. Venkatesh et al. (2016) themselves have warned that the many incremental extensions of UTAUT have slowed rather than advanced theory, and called for multi-level frameworks instead.

What is missing is a bridge: an account of how individual people working with agents add up to organisational capabilities that competitors cannot easily copy.

The Enterprise Cognitive Autonomy framework

My proposal is to stop choosing one lens and combine four, each covering a level the others miss. The Enterprise Cognitive Autonomy (ECA) framework integrates them into one multi-level model.

Lens	Level	Factors	Question it answers
Resource-Based View (Barney, 1991)	Organisation	Tangible, intangible, human and organisational resources	What does the firm bring to the table?
UTAUT (Venkatesh et al., 2003)	Individual	Performance expectancy, effort expectancy, facilitating conditions, social influence	Will people actually use and trust the agents?
Microfoundations (Barney & Felin, 2013)	Behaviour	Process-based mechanisms, individual-level foundations	How do daily human-agent interactions add up to organisational capability?
Situated AI (Kemp, 2024)	Embedding	Grounding, bounding, recasting	How does generic AI become firm-specific and hard to copy?

Two adjustments matter. First, I extend the human resources category to include AI agents themselves. If agents take defined roles in a Multi-Agent system, they behave less like infrastructure and more like Drucker’s knowledge workers. Second, I treat social influence and facilitating conditions as organisational attributes, since decisions about Multi-Agent systems are strategic rather than individual choices.

The core claim is simple: resource-rich firms can still fail if people do not accept the agents, and enthusiastic users can still fail to create advantage if their work never becomes an organisational routine. Only alignment across all four levels turns an AI investment into a genuine cognitive transformation of the enterprise.

Grounding, bounding, recasting

The missing link between generic AI and firm-specific advantage comes from Situated AI. The idea goes back to Rodney Brooks (1991) and Lucy Suchman: intelligence does not live in abstract reasoning alone, but emerges from interaction with a specific environment. Kemp (2024) turned this into a theory of competitive advantage built on three organisational activities.

1. Grounding anchors agents in the firm’s own knowledge, decision criteria and strategic goals. Two companies can run the same model; only one has its twenty years of process knowledge, exception handling and customer history behind it.
2. Bounding protects what has been built through contracts, access controls and keeping the people whose tacit knowledge sustains the capability. The model is public; the configured system around it should not be.
3. Recasting continuously adapts agents as markets, processes and regulation change, so the capability stays relevant instead of freezing into a snapshot.

In my framework, I apply these activities to the inside of a Multi-Agent system. Individual agents are grounded one by one. The agent network as a whole is bounded. And interdependent agents are recast together, because changing one specialist agent changes how the whole team behaves.

This is how commoditised AI can meet the VRIN criteria after all. Not because the algorithm is rare, but because the way it is embedded is. Whether that embedding also produces measurable advantage over competitors is exactly what my empirical study sets out to explore.

What this means for enterprise leaders

If the argument holds, the strategic question shifts from “which model should we buy?” to “what do we build around it that others cannot?” A few implications follow.

• Invest in context, not just compute. Proprietary data, curated knowledge bases and documented decision logic are what grounding runs on. Infrastructure is necessary, but it is the easiest part to copy.
• Treat agents as team members with roles. Narrow, well-defined agent roles perform better, and they also make governance, accountability and oversight tractable.
• Design governance in, not on. Accountability for autonomous decisions cannot be added after go-live; the window to shape it narrows as agents become embedded in operations.
• Take acceptance seriously. Shao et al. (2025) found that workers wanted more human agency than experts judged necessary for 47.5% of tasks. Ignoring that gap is a fast route to rejection or under-use.
• Turn individual wins into routines. A handful of power users does not make an organisational capability. Processes that capture and share what works between humans and agents do.
• Plan for the J-curve. Productivity often dips during transition. Firms that abandon the effort too early never see the benefits (Bornet & Wirtz, 2025).
• Keep recasting. A situated capability decays if it is not adapted. Budget for continuous refinement, not a one-off project.

Open questions, and what comes next

The ECA framework is a proposition, not yet a finding. Several debates remain open in the literature, and I expect some of them to reshape it.

• Multi-Agent or Centaur? Borghoff et al. (2025) contrast loosely coupled autonomous agents with deeply integrated human-AI “Centaurian” systems. Which architecture preserves human judgement best is unresolved.
• Augment or automate? Raisch and Krakowski (2021) describe an automation-augmentation paradox that organisations cannot simply resolve by choosing a side.
• Capabilities or business models? Climent et al. (2024) argue that AI advantage may stem more from business model design and data network effects than from organisational capabilities.
• Is the evidence good enough? Most gains reported for Multi-Agent systems come from controlled benchmarks, and far fewer from real enterprise deployments.

To test the framework, I am running a qualitative multiple-case study with large German enterprises that have started or completed their shift towards Multi-Agent systems. I interview technical leaders, AI architects, process owners and end users, and compare how each group makes sense of the transformation across cases.

My bet, as a researcher standing between technology and business, is this: in a few years, the winners will not be the firms with the most agents, but the ones that learned to think differently with them. Cognitive transformation is an organisational achievement, not a software release.

If you work on agentic AI in a large enterprise in Germany and recognise these questions from practice, I would be glad to hear from you.

References

• Acharya, D. B., et al. (2025). Agentic AI: Autonomous intelligence for complex goals. IEEE Access.
• Ali, M. A., et al. (2024). Cognitive systems and interoperability in the enterprise: A systematic literature review.
• AL-khatib, A. W. (2023). Drivers of generative AI adoption in retail.
• Barney, J. (1991). Firm resources and sustained competitive advantage. Journal of Management, 17(1).
• Barney, J., & Felin, T. (2013). What are microfoundations? Academy of Management Perspectives, 27(2).
• Borghoff, U. M., et al. (2025). Human-artificial interaction in the age of agentic AI.
• Bornet, P., & Wirtz, J. (2025). Agentic Artificial Intelligence.
• Brooks, R. A. (1991). Intelligence without representation. Artificial Intelligence, 47.
• Climent, R. C., Haftor, D. M., & Staniewski, M. W. (2024). AI-enabled business models for competitive advantage.
• Crawford, N., et al. (2024). BMW Agents: A framework for task automation through multi-agent collaboration.
• Kemp, A. (2024). Competitive advantage through artificial intelligence: Toward a theory of situated AI. Academy of Management Review.
• Melville, N., Kraemer, K., & Gurbaxani, V. (2004). Information technology and organizational performance. MIS Quarterly, 28(2).
• Paiva, J. (2024). AI adoption through the TOE framework: A meta-analysis.
• Raisch, S., & Krakowski, S. (2021). Artificial intelligence and management: The automation-augmentation paradox. Academy of Management Review, 46(1).
• Shao, Y., et al. (2025). Future of work with AI agents.
• Shu, R., et al. (2024). Towards effective GenAI multi-agent collaboration.
• Venkatesh, V., et al. (2003). User acceptance of information technology: Toward a unified view. MIS Quarterly, 27(3).
• Venkatesh, V., Thong, J. Y. L., & Xu, X. (2016). Unified theory of acceptance and use of technology: A synthesis and the road ahead. JAIS, 17(5).
• Wooldridge, M., & Jennings, N. R. (1995). Intelligent agents: Theory and practice. The Knowledge Engineering Review, 10(2).