Agentic Organization: How companies are reshaping collaboration and value creation and why organizational development is crucial now

AI Provides Answers – The Low-Threshold Entry Point
At the first stage, AI is primarily used as an intelligent answer generator. People ask questions, and AI provides answers: quickly, conveniently, and with a relatively low level of sophistication—but ready for immediate use.

Value contribution:

• Quick knowledge retention
• Efficiency in research and information retrieval
• No automation—efficiencies can be improved using templates

Limitations / challenges:

• No integration into workflows
• Quality depends heavily on the input and individual AI expertise
• Corporate knowledge is not systematically integrated
• The first “template governance” is merely a guideline and does not yet constitute a principle of control or design

First Targeted Steps: What Companies Need to Consider at This Starting Point:
Level 1 is a good starting point, but it does not yet lead to organizational progress. It establishes initial points of contact and reveals whether your company meets the basic prerequisites for implementation—without yet bringing about sustainable change or a noticeable improvement in competitiveness. This makes a clear approach, specific objectives, and an appropriate progress assessment all the more important. Pure “IT experiments” can be useful, but they should be consistently embedded within a coordinated purpose and framework. Otherwise, successes and lessons learned remain localized and cannot be replicated. In the worst-case scenario, the organization wastes time early on with an inefficient initial implementation and loses momentum in its AI ramp-up.

Specific areas of action for Level 1:

• Organizational: Train your employees in the use of generative AI and in how to select the right models for their projects. Teach them what constitutes efficient and effective prompting, and raise their awareness of data handling and the costs associated with using AI.

• Process-related: Even at this stage, it’s already worthwhile to provide prompt templates for various functional areas of your company and to encourage their use. Initial governance measures can boost efficiency, improve the quality of responses, and—most importantly—ensure consistency in those responses.

• Technical: Check whether publicly available AI models are suitable for your purposes. However, companies are generally required to opt for a isolated setup, if only for data protection reasons. Here, you can decide whether to use your own IT infrastructure for hosting or to use a service provider.

In the second stage, employees refine their interaction with AI. Content is not only generated but also deliberately controlled, adapted, and improved through prompts.

AI learns to understand your business—but only if your organization understands its own data

In the second stage, the organization and its employees refine their interaction with AI. Content is not only generated but also enriched, customized, and improved through the integration of your company’s data and documents. Through this integration, the AI can not only generate responses based on general, publicly available knowledge, but also address contexts specific to the company.

However, anyone who delegates Phase 2 to IT as a purely technical task misses the key levers: At its core, now is the moment to adopt a data strategy—not as an IT issue, but as a cross-functional corporate and management task. This is because the quality, structure, and availability of your data directly determine the performance of your AI systems from this point forward and are the key to specific results that only you can achieve by leveraging your organization’s knowledge. In other words: Your AI maturity level can never exceed the maturity level of your data. An organization that wants to operate AI at Level 2 but has data at “Level 0”—fragmented, undocumented, and scattered across silos—will fail.

Value contribution:

• Significant increase in specificity through systematically integrated “corporate knowledge”
• Improving Prompt Templates Through Domain-Specific Context
• Significantly increased access to knowledge, including from unstructured sources
• The first real competitive advantages, as AI results are no longer generic but rather company-specific and thus relevant to decision-making
• More efficient integration into workflows through access to document templates or initial integration with IT interfaces

Limitations / challenges:

• The efficiency and effectiveness of use continue to depend on individual usage
• No comprehensive integration into the process and IT landscape
• Data quality, accountability, and availability are often insufficient for AI to “understand” context directly without additional prompting
• The lack of a comprehensive, semantic information model means that AI (and your employees as well) do not have a clear, unambiguous understanding of terms to build upon

Enabler or Cost Driver: What Companies Need to Consider at This Crossroads:
This stage builds expertise and can deliver specific results for your company. This creates the potential not only to optimize employee efficiency but also to generate the first real competitive advantages. However, AI still requires active, individual control. Processes are not yet integrated and therefore essentially remain manual and localized.
In terms of AI maturity, companies are at a strategic crossroads here: If this step is successful, an additional AI setup—one more strongly focused on workflow integration—can be implemented, thereby creating a genuine lever for productivity and competitiveness. However, without clear objectives, without overarching management, and especially without a data strategy, there is also a risk that AI will become a driver of complexity and costs without delivering any tangible added value.

Data Strategy as a Leadership Task: At this stage, information and data management become critical to success—and thus a leadership task. Specifically, this means that data is not a byproduct of operational work, but rather an asset that must be actively managed. IT is a key stakeholder and contributor here, but anyone who treats data as a purely technical resource confined to IT will find that even the best AI models will fail when faced with poorly maintained, inconsistent, or incomplete data.

At this stage, a data strategy means at least three things:

1. Clarity on which data is strategically relevant —not all data is equally important. Which data is critical to the success of your prioritized AI use cases?
2. Binding responsibility for data quality and availability —not as a project, but as an ongoing operational discipline within a line function
3. A domain-specific information model that captures the specific essence of your business and makes it understandable to all users—both human and machine

Why Enterprise Search Alone Isn’t Enough: Many companies find themselves at this point without a strategic perspective on their data assets and assume that simply integrating their existing, operationally and technically oriented information and data assets with an enterprise search tool is enough. These tools can certainly deliver initial results on their own, such as faster access to existing documents.

However, this purely “bottom-up” approach alone is not sufficient for AI to truly “understand” an organization’s specific data. Enterprise Search finds documents. An information model creates meaning. The difference is fundamental:

• Bottom-up (Enterprise Search): The technology indexes whatever is available—regardless of quality, recency, relevance, or subject-matter context
• Top-down (information model): The organization itself defines—based on its subject-matter expertise—what its core information objects are, how they relate to one another, and what they mean

AI can only “extract” a precise and unambiguous meaning from a company-wide, domain-specific information model. This model must be created and maintained by subject matter experts—not by IT. It comprises three levels:

• Semantic level: Consistent terms, taxonomies, and definitions—so that “customer,” “account,” and “order” mean the same thing to AI when they refer to the same concept
• Contextual level: Metadata describing the source, recency, confidentiality level, and intended use—so that AI can assess how reliable a piece of information is
• Structural level: Relationships between units of information (customer → contract → product → transaction) — so that AI can establish connections, rather than just providing isolated facts

The good news is that this model doesn’t have to be fully in place on day one. Through functional ownership of their data, companies can quickly generate initial value based on specific use cases, even if the target state hasn’t yet been fully achieved. The journey itself delivers the first benefits of the goal.

The economic risk of “automating” poor data quality with AI: One aspect that is systematically underestimated in the discussion about AI integration is that the costs of poor data quality rise significantly as the level of automation increases. In a manually managed organization, people quietly correct data errors—they know from experience that a certain number “can’t be right” and adjust it accordingly. In an AI-integrated organization, those same errors aren’t corrected—they’re amplified. Every automated decision based on flawed data, in turn, generates more flawed data—and a vicious cycle can begin. The resulting costs are immense: incorrect analyses, flawed recommendations, and data-driven workflow interruptions are not prevented by AI; on the contrary, it causes these problems to escalate.

Conversely, every investment in data quality is a direct investment in AI performance. Companies that view data as an asset and invest accordingly will find that their AI systems, using the same technology, deliver significantly better results than competitors who focus on the technology and neglect the data.

Specific areas of action for Level 2:

• Organizationally: Don’t just think of information and data sources within the company in terms of their initial use; instead, aim for cross-functional reusability. Specifically, this means structuring and preparing information and data in such a way that everyone—colleagues in other departments as well as AI systems—is able to fundamentally understand their purpose, content, and status. This can be achieved by beginning to assign the generation, maintenance, and provision of information and documents to specific functional areas and by establishing initial standards. To achieve this, it has proven effective to establish a data governance organization outside of IT and to define corresponding data roles—such as data owners for governance aspects and data stewards at a more operational level—within your business units. In addition, the level of data literacy and culture must not be neglected. Until now, data was often “the responsibility” of individual departments or even teams. Now, data is becoming a cross-functional corporate asset. On the one hand, this means building an understanding that transcends previous functional boundaries. On the other hand, existing “claims of ownership” over data—which are a root cause of the infamous data silos—must be systematically dismantled.
• Process-oriented: Identify the most relevant information and data within your company or its functional units. Begin to conceptualize these “data treasures” within a life cycle model. This means that the creation, use, storage, and availability of data are specifically defined and assigned to dedicated responsibilities. On this basis, it is now possible to use integrated AI systems in an action-oriented manner for each of these life cycle steps, thereby establishing a comprehensive standard process. In addition, it is essential to standardize, comprehensively define, and measure data quality at every step and to continuously optimize it based on these criteria. It is not enough to speak generically of “data quality.” Define quality dimensions and establish a procedure that specifies quality expectations for data usage and continuously monitors derived metrics. It is also important to include metadata here. It is crucial for AI to recognize the context in which the data exists—creator, version, purpose, scope of validity, and other aspects are often not explicitly stated today because human users can draw on experience that is not accessible to AI.
• Technical: Identify the relevant information sources within your company and make them available to AI. There are various technical methods and approaches for this, which address, on the one hand, the way in which data becomes accessible to AI (including retrieval-augmented generation, vector databases, and knowledge graphs) and, on the other hand, address non-functional requirements such as process integration, latency, or data volumes. Depending on the technical solution, standard interfaces may already be available. However, you may need dedicated expertise to integrate your specific setup.

Value contribution:

• Significantly higher quality of results
• Repeatable work steps through templates
• Structured handling of AI

Limitations / challenges:

• Greater need for knowledge among users
• Results vary depending on competence
• No overarching process changes

What companies need to consider:
This stage builds up expertise, but does not yet lead to a transformation of processes or roles. The effect remains local and team-related.

AI agents are designed to support roles, tasks, and processes—transforming you from a passive supplier into an active sparring partner

At this stage, AI becomes “agent-like.” . In short: AI agents figure out for themselves what needs to be done to produce the desired result. These can be simple tasks, such as finding suitable meeting suggestions as a follow-up to meeting minutes, or more complex issues, such as analyzing typical quality problems based on historical data in distributed systems. Assistants thus take on defined subtasks, support more complex activities, and can handle various aspects of workflows end-to-end. In doing so, they increase the speed, consistency, and quality of work results—without requiring an employee to manually guide the AI through each step via prompts. Within the scope of its integration into the IT landscape and access permissions, the agent can actively pursue its own solutions.

Value contribution:

• Solutions no longer need to be specified step by step; instead, they can be identified and implemented “proactively” by AI agents themselves.
• Significant efficiency gains for certain process areas/steps
• Employees and teams can focus on defining goals and building on the results

Limitations / challenges:

• A lack of transparency regarding processes and workflows hinders the efficient professional development of agents
• Conflicts and overlapping tasks resulting from a lack of clarity regarding roles and responsibilities are not resolved by agents, but are merely reproduced automatically
• Requirements for data quality and availability are rising once again, and the often-lacking business relevance of data structures prevents the large-scale deployment of agents
• Access to and technical integration of agents with legacy systems is often limited
• Agents that are active and perceived as acting autonomously generate resistance and fear within the organization and are therefore not consistently utilized

The agents are starting to take effect: what companies need to keep in mind for this breakthrough:
Stage 3 is the visible breakthrough—but without (data) governance, clear roles, and process harmonization, a patchwork of tools emerges that becomes difficult to manage in the long term and cannot resolve existing inefficiencies. In certain areas, the tools can certainly improve efficiency locally. However, organizational and data-related silos severely hinder the impact of AI. One visible symptom is internal agent platforms where “everyone builds their own agent,” resulting in many overlapping or completely identical applications being deployed. The problem here is not that business users are taking the initiative themselves. An “unmanaged” approach, however, does not automate value creation but rather leads to complexity, conflicts of responsibility, and dysfunctional processes. An organizational structure tailored to the company’s capabilities is therefore a central prerequisite for enabling business users to work with and design agent functionality and workflows—based on a cross-functional data foundation—with clear functional ownership. Agent integration is therefore not an IT issue, but rather a management and organizational task. This perspective also prevents the risk of suddenly exposing your employees to this transformative technology. For many people, the “magic” of agents can seem threatening and alien. Without dedicated change management and appropriate empowerment, you’ll lose acceptance and, in the worst case, generate significant resistance and rejection. This has always been—and essentially remains—true for any (technical) change. However, seeing an agent suddenly “do your job”—especially amid widespread media attention—has the potential to greatly amplify the emotional impact. Communication and support are therefore clearly critical to success.

Specific areas of action for Level 3:

• Organizational: Establish and develop a process and data organization structured by business function, along with governance, as an indispensable prerequisite for ensuring that agents operate without overlap, are defined and managed by the appropriate business units, and are supplied with the correct data. Communicate successes and learning progress. Establish an empowerment program for your workforce. Also, specifically define and address how the freed-up capacity should be used in a concrete and value-adding way across the various areas. Allow space for concerns and fears, help your workforce gain confidence in the new technology at their own pace, and work together to shape the new reality of work.
• Process-oriented: Set up initial “agent teams” for your strategic use cases and core process areas, and roll them out step by step to different areas. Use this as a basis to determine which tasks/process steps can be “automated” to what extent, and what technical integration, data access, and agent-to-agent interaction must be enabled to achieve this. Identify the handoff points in the agent-based workflows to remaining manual or traditional processing. Assess which of these “speed bumps” can be bridged in the short term through appropriate process adjustments.
• Technical: Define a portfolio of technical skills, system access, and data access that you want to make available to agents. Integrate external sources and functionalities. To this end, establish a comprehensive catalog that not only provides your employees with centralized access but also enables technical integration via MCP and A2A protocols. Determine which technical platform is best suited for your existing and strategically planned setup.

Autonomous AI agents act independently—work models break away from sequential process logic

In Level 4, the integration of AI agents is no longer limited to individual use cases, processes, or functional areas. The organization itself is designed so that agent-based systems can operate autonomously, in parallel, and in a modular fashion —without being hindered by “human-in-the-loop interfaces” or sequential process steps. Humans and AI collaborate at different paces and from different perspectives; functionally defined agent governance and flexibly orchestratable capability models form the organizational platform for this.
Work is no longer organized primarily through fixed process chains, but rather through a clearly defined capability landscape. Capabilities describe the organization’s functionally managed competencies—including objectives, data foundations, quality criteria, decision-making logic, and the permissible degree of autonomy. AI agents dynamically orchestrate these capabilities in accordance with defined policies and guidelines. Traditional processes thus lose their role as a central control instrument. The previous process steps are now understood within the framework of capabilities as separate work modules with inputs and outputs. The former sequential logic of the processes is reviewed in this context. All aspects related purely to procedural organization can be eliminated. Aspects related to control and governance are translated into target inputs and capability participation.

Humans and AI deliberately operate at different speeds:
Agents function as a permanent execution layer and handle operational, analytical, and coordinative tasks at a high frequency. Humans focus on activities that cannot be automated—or at least not fully automated—or that are deliberately handled “manually.” These are primarily the responsibilities that keep the organization on track: defining the initial intent (direction) and continuously refining it, establishing and strategically developing functional areas, orchestration, monitoring, and risk and value management. Value creation does not arise from individual agents, but rather from the interplay between clearly defined intent and scaled agent-based execution.

Value contribution:

• Substantial productivity gains through 24/7 execution of agent-based capabilities without sequential human bottlenecks
• Structural competitive advantage, as operational tasks and elements of core value creation are largely automated and scalable
• Greater adaptability, as new requirements can be met through the flexible and spontaneous combination of capabilities.
• Significant potential for innovation, as the functional organizational structure allows new strategic directions to be quickly applied to existing capabilities and new areas to be easily designed and integrated

Limitations / challenges:

• A very high level of organizational maturity is required
• New Role and Responsibility Models Are Needed for the Capability Model
• Significant Changes in Leadership and Culture
• Critical Dependence on a Cross-Domain Semantic Database
• Clear and Consistent Capability and System Architectures

The Agent-Based Vision: What Companies Need to Consider for This Transformation:
Level 4 is not a scaling step from Level 3, but rather a qualitative shift in the organizational principle. Agents are not deployed here to make existing processes more efficient, but to execute organizational capabilities independently and collaboratively. Without a conscious departure from sequential process logic and a consistent focus on functional capabilities, AI agents are indeed potentially powerful tools —Level 3 is, for the time being, an extremely desirable scenario for the vast majority of companies—but they remain “trapped” in the sequential constraints of traditional process organizations.

Since operational execution is delegated to agents in an “agentic organization,” the management and strategic design of key deliverables become the focus of human activities. It is therefore crucial that agentic autonomy be limited not technically, but organizationally: policies, subject-matter ownership, escalation mechanisms, and transparent performance metrics remain central. The agentic organization relies on data-driven transparency and the ability to intervene. This is the only way to build trust—among both leaders and employees—and the only way to combine agentic speed with traceability and accountability. On the other hand, companies can and must use the freed-up capacity for strategic development and building new capabilities.

• Organizational: Expand or finalize the technical and functional organizational transformation begun in Phase 3, and incorporate potential or targeted levels of automation, clear input/output definitions, and “Capability Agreements” that define interface and interaction points—such as data quality and the like. Control and governance concepts, as well as corresponding roles, must be defined and established. New roles such as “Capability Owners,” Agent Stewards, and Agentic Governance Managers are conceivable—there are no established frameworks for these yet; nevertheless, existing standards can be derived by extending and adapting existing methodological frameworks from architecture, data management, and IT operations. Agent-based control and strategic capability development must be established as a management discipline for agent portfolios, prioritization, and benefit and risk management.
• Process-oriented: Gradually decouple your value streams from sequential end-to-end processes and transform them into capability flows. The key here is to create combination spaces from the former value streams that, on the one hand, condition or enforce specific capability sequences but, on the other hand, allow for degrees of freedom in which agents can act independently and spontaneously. Clear intervention and fallback points must be established between agent-driven execution and human intervention. From the agent’s perspective, this means establishing dedicated exit points at which agents request human intervention, while also providing opportunities from a governance perspective to halt and correct the agents’ execution.
• From a technical perspective: “Anyonestill discussing technologyatStage 4 hasn’t completed Stage 3.” Or —to put it another way—good news: the work on technological innovation is largely complete at Stage 3. The focus of the transformation remains at the organizational and process levels. On the technical side, the focus should be squarely on stabilization and “zero-incident” operations. This is not a practically attainable state, but rather a guideline for continuous improvement—for which, of course, agent-based capability models should also be utilized. However, it is important to emphasize that the technical foundation is even more critical to an organization’s functioning—which is already the case today. Additionally, care should be taken to ensure that the technical setup consistently supports transparency, management, and control mechanisms. It is also likely—and, as the organization grows, almost inevitable—that various technical solutions and platforms will have been implemented on the path to Level 4. To reduce complexity and, certainly, costs, it is advisable to develop and implement consolidation and standardization architectures.