From Prediction to Action

For several years, the use of AI in organizations was mostly used for classify data, predict results and options recommending. These systems were useful but they needed for a human to ask a question and then obtain the result. New LLMs and foundation models have changed this situation. When they are connected to planning tools, memory, outside tools and business software, they can help to finish a series of actions instead of just giving one single answer [1,2].

This change creates a new challenge for Information Systems (IS) research. An agentic system is not only a database, a dashboard or a tool that produces reports. It can observe a situation, select the next step, use an enterprise tool, prepare a decision and ask a human for approval. Because of this, the system becomes closer to a controlled actor inside the organization. For this reason, Agentic AI should be studied as a sociotechnical system. This means that its technical abilities, organizational decision processes, human responsibility and governance rules must be designed and evaluated together [3,4].

Research Problem

The practical question that can appear is not if institutions should give all their decisions to AI. That would not be a safe process and it is unrealistic. The more useful question is how companies can build agentic information systems that help with independent work at the same time, humans stay in control. Some research questions guide this article:

• How can we match the skills of Agentic AI with decision-making tasks in an institution?

• How can trust, clear information and human supervision change the level of power given to an AI agent?

• How can public sector organizations, especially in education, use Agentic AI without having a negative effect on responsibility, fairness, privacy or human decisions?

Contributions of the Paper

This article focuses on ideas and practical use. It intentionally does not include results from real data or experiments. Its main contributions are:

• A clearer explanation of Agentic AI for information systems research

• An improved Agency Decision Organization (ADO) model that connects the agent's skills with decision-making, putting plans into practice, feedback, trust and control

• A practical example in public education. It shows how an agentic system could be helping with school requests and organize office work, without removing away the control of managers

• A clear plan for rules and implementation for organizations that want to try Agentic AI in a safe way

Theoretical Foundations

Agentic AI in organizations connects different areas of research. Table 1 shows a summary of the main ideas used in this paper.

Table 1: Interdisciplinary Foundations for Agentic AI in Information Systems

Intelligent agents and bounded autonomy

In classical AI, an agent is can be presented as a system which has the ability to observe its environment and act on it [5]. Agentic AI enhances this concept. It utilize large language models to support reasoning and communication, while other tools allow the agent to access databases, documents, workflow systems, calendars, search tools and application programming interfaces. In this context, autonomy is not referring to complete freedom. It should be given as bounded autonomy. Which means that the agent can operate according to a clear goal, role and permission level. However, its actions are still limited by policies of a specific organization, technical boundaries and human supervision.

This difference is important. An agent that provide a recommendation for school maintenance is not similar to a human manager who is legally responsible for approving the budget. The agent can give help, organize and prepare actions but the power must be given clearly by the organization.

Organizational Decisions and Algorithmic Boundedness

Simon’s theory of bounded rationality demonstrates a concept of that people rarely make perfect decisions. They work with information that are incomplete and often make a "good enough" decision when they do not have much time [6,7]. Agentic AI also has bounded rationality but it has different limits, such as training data, prompt instructions, access to tools, memory design, how well it finds information, context length and safety rules. This article calls these limits algorithmic boundedness.

Algorithmic boundedness can be clarified as an AI agent can look confident while still being wrong, incomplete or not following the company's rules. Therefore, we should not judge agentic information systems just by  how  good  level  their  writing.  We  shouldjudge them by how easy it is to track their steps, the quality of their evidence, how easily they can fix mistakes and how they ask for human help when they are not sure.

IS Success, AI Capability and Organizational Value

The DeLone and McLean IS Success Model is remain useful until now, this is because it explains how the quality of system, information, service, alongside with system use, user satisfaction and net benefits are connected [10]. For Agentic AI, system quality should include more than the normal technical performance. It should also include behavioral stability, tools for safe use, summaries of clear reasoning, proper escalation to humans, data protection and follow the rules of organizations. Research on the capabilities of AI also shows that business importance not only coming from the technology [10,16].

Trust and Appropriate Reliance

Trust is a crucial factor because agentic systems can provide recommendations or take actions. Trust should not mean that everything is accepted without thinking. The goal is proper reliance, people should accept the agent when provide a good work and reject, check or change its decision when the situation seems to be requiring human judgment [11,17]. For Agentic AI, trust depends on its performance, how clear the process is, if it shares the same values and if the user can check the evidence and step in.

The Technological Backbone of Agentic AI

Agentic AI is not a single model. It is a structure that connects a model to data, tools, memory, rules and human control. The features below are very important for information systems.

Foundation Models as Reasoning and Communication Components

Foundation models learn from large amount of data and after that they can be used for many different tasks [1]. Resulting in organizations can use them to read policies, summarize cases, write explanations and talk with users. However, these models should not be dealt with as the main source of knowledge.

Planning, Reasoning and Acting

With the chain-of-thought prompting, researchers found out that intermediate reasoning can be beneficial for some complicated tasks, especially those with provided examples [18]. React expands the approach by combining reasoning traces with actions in order to make the model think, able to use a tool, to be able to see the results and, therefore, to know what action to take next [19]. For the organization, such an approach is good as the process will be more understandable. The agent will be able to demonstrate what information it uses and why it suggests doing something in particular.

Tool Use and System Integration

The use of tools is important since the decision-making in the organization relies on current documentation rather than the knowledge of the model. According to the authors of Schick et al. [20], it is possible to train the language model called Toolformer to use external tools in certain situations such as search engines, calculators, translators and calendars. It means that in information systems, the agent is able to integrate the following systems: case management systems, student information systems, procurement platforms, maintenance ticketing systems and policy repositories.

Memory, Reflection and Feedback

Agent memory helps maintain context across a task. Reflection mechanisms allow an agent to review past outcomes and adjust future behavior. An example on this case, in Shinn et al. [21] authors used a method called Reflexion, its model keeps feedback as words, not in updated weights, so can do better next time. But in an organization, this reflection process needs control. New lessons should not change important decision rules by themselves. Instead of that, they should be reviewed before they have any new updates on the approved prompts, rules or knowledge base.

The Agency-Decision-Organization (ADO) Model

explains how Agentic AI can become useful in an organization. The model is built around three main layers: agency, decision and organization. It also includes two important factors that influence all layers: trust and governance (Table 2).

Table 2: Agentic AI Capabilities and Organizational Interpretation

Agency Layer

The agency layer focuses on what the AI agent can actually do. Which consists of its ability to understand information, search and retrieve data, make plans, use tools and remember previous interactions and review its own output. In this layer, the main question is what kind of information the agent can access and what actions it is allowed to start by itself.

Decision Layer

The decision layer is where the collected information is changed into a recommendation or an action. It includes policy checking, risk scoring, option comparison, uncertainty detection and explanation generation.

Organization Layer

The organization layer covers implementing, people roles, accountability, feedback and performance evaluation. The organization layer is held accountable for determining whether the AI output actually improves the process. A good agent may fail due to lack of quality data, decision-making ability, escalation procedure or training programs in the organization.

Trust and Governance as Moderators

Because they influence all the three levels of it. Trust is considered an important aspect in this case because the user will not be able to agree to the advice given by the agent unless it is understandable and trustworthy. Governance, on the other hand, sets the rules  for  how  the agent should be used. It explains what the agent is allowed to do, how risk should be evaluated, how decisions must be documented and how problems should be handled when they occur. NIST describes trustworthy AI through several characteristics, including validity, safety, security, accountability, transparency, explainability, privacy and fairness [13]. In a similar direction, ISO/IEC 42001 [15] focuses on the management side by requiring organizations to build, maintain and improve processes for responsible AI use.

Based on this model, four propositions can be suggested for future research:

• P1: A higher level of agent autonomy could be a help for organizations coordinate work faster, at the same time, this is more likely when the task is clearly limited and the agent is using reliable tools

• P2: Agentic AI is not fully independent because it works within technical, data and policy limits. For this reason, detecting uncertainty and escalating difficult cases should be treated as important parts of system quality

• P3: The value of Agentic AI in organizations depends on more than the AI system itself. It also requires good data, redesigned processes, clear governance and proper training for users

• P4: Users can rely on the agent in an appropriate way when they are able to see the evidence, understand the recommendation and able to override the agent work whenever they feel it is necessary

Use Case: Agentic AI for Government Educational Institutions

Context and Problem: Government educational institutions and school directorates deal with several requests every day from different schools. Such requests may relate to issues like repairs, needs in classroom equipment, replacement of teachers, safety reports, policy inquiries, student support issues, transportation arrangements and due dates of reporting. Normally, such requests arrive from different sources: Emails, online forms, telephone conversations or hard copies. As a result, such information is often not centralized. Moreover, sometimes the requests may be delivered late, to the wrong person or repeatedly since some of the necessary information may be missing. Managers can also have difficulties in identifying the urgency of the particular case or its status.

A possible use of Agentic AI in this scenario is a School Support Service Agent. The agent will not make any decisions regarding grading students, admission of students, punishment of students or any other rights. Thus, the functions of the agent are purely administrative: information gathering, policy review, suggestion preparation, forwarding of the case to the right employee and record keeping.

Purpose of the School Support Service Agent

This system will be helpful for the public educational institution or regional school directorate to manage school requests. The aims are:

• Collecting school requests via email, online form or web portal and categorizing them based on their type

• Evaluating whether the request contains all the required information and requesting for further information from the school where something is missing

• Identifying the relevant policies, previous cases, budgetary guidelines, school information and service level agreements

• Prepare a suggested response for the responsible officer, including supporting evidence, policy links, priority level, possible risks and the next recommended step

• Sending less complicated and low-risk requests for routine processing but sending sensitive, urgent and high-cost requests to the relevant officer

• After approval, update the case-management system, informing the school about the decision and documenting the case

Workflow Explanation

• School request or trigger: A school submits a service request or an internal data trigger identifies a problem such as repeated maintenance delays

• Intake and validation: The agent checks identity, school code, request type, completeness and urgency

• Retrieval: The agent retrieves approved policy documents, budget rules, previous similar cases and relevant school data

• Risk and eligibility assessment: The agent checks whether the case is low risk, high cost, sensitive or legally constrained

• Draft recommendation: The agent prepares a human-readable recommendation with evidence, assumptions, confidence level and alternatives

• Human approval or escalation: A responsible officer approves, edits, rejects or escalates the case

• Execution: After approval, the agent updates systems, creates tickets, schedules tasks or sends official messages

• Audit and feedback: The system records the decision path, captures officer feedback and supports reporting (Figure 1, Table 3)

Table 3: Decision Rights and Safeguards for the School Support Service Agent

1. Figure 1: Workflow for the School Support Service Agent Use Case

Reason Why This is an Appropriate Use Case

The use case is appropriate since it deals with administrative coordination instead of autonomous decision-making regarding students. This approach is capable of increasing service consistency, eliminating repetitive routing processes and achieving better audibility. Also, the use case falls in line with human-centered guidelines for the education domain, which state that AI systems should increase the capabilities of humans and not eliminate their educational judgment [22]. When it comes to European countries, designers should pay attention because AI in the education system might fall into high-risk use cases according to Regulation [23].

Expected Benefits Without Claiming Results

The potential benefits should be expressed in terms of hypotheses or design goals and not as the results if there is no evaluation of the product yet. Some of the potential benefits could be increased speed of request intake, reduced number of lost requests, consistent policy implementation, better explanations provided to schools, greater managerial control and better auditability.

Governance and Implementation Roadmap

Phased Deployment: A safe implementation should move in stages. The agent should earn authority through evidence, not receive high autonomy at the beginning.

• Advisory phase: The agent writes the recommendations and explanations. All decisions are made by humans. This phase examines retrieval quality, explanation quality and user acceptance

• Supervised execution phase: The agent can carry out certain approved low-risk activities like creating a maintenance ticket or giving a status report while humans make decisions

• Bounded autonomy phase: The agent is able to carry out well-defined and routine activities within certain thresholds while sensitive cases are always reviewed by humans

Minimum Governance Controls

The   following   controls   should   exist   before deployment:

Clear process ownership: One responsible office controls the agent and its decision authority limits

• Data governance: Approved sources, data reduction, retention and access control

• Policy-grounded retrieval: The agent responds only using approved policy documents as guidance

• Human approval gates: Spending, student-sensitive cases, legal exceptions, safety issues and ambiguous cases require human decision

• Audit logging: All recommendations document input, data sources, methods, user editing, approval and action taken

• Monitoring and incident management: Problems related to errors, bias, privacy issues and unsafe recommendations have escalation procedures

• Periodic review: The prompts, rules, permissions and knowledge bases are reviewed by subject matter experts

Evaluation Metrics for a Pilot

The evaluation of a pilot should consider both performance and accountability. Possible metrics include:

• Operational metrics: Time to classify a request, time to first response, backlog size, completion time and routing accuracy

• Decision-quality metrics: Correctness of policy retrieval, quality of the evidence gathered, correction rate of officers and accuracy of escalation

• Trust and usability metrics: Staff satisfaction, perceived usefulness, perceived transparency and override patterns

• Governance metrics: Percentage of cases with complete audit logs, number of privacy violations, number of unresolved incidents, fairness tests and results of compliance reviews

• Education sector metrics: Satisfaction of schools, equality of service among schools, response consistency and reduced administrative questions

Main Risks and Possible Solutions

Agentic AI poses potential threats if the system is over-permitted or not supervised properly. The primary threats include wrong policy interpretation, privacy breach, unfair prioritization, over-dependence on the part of employees, lack of accountability and automating crucial education decisions. It can be addressed through limited permission, source-based search, human verification, constant supervision, privacy by design, role-based permissions, bias testing and clear communication with schools.

Implications for Information Systems Research

Agentic AI demands that IS researchers go beyond the standard questions of adoption. It is no longer sufficient to merely determine whether the system is accepted; researchers have to determine the degree of authority conferred to the system, its coordination with other systems, the production of evidence and accountability. The ADO model indicates that all these elements should be studied together.

Implications for Public-Sector Organizations

Government agencies have higher levels of accountability requirements compared to some private sector enterprises. They require explanations, data protection of citizens, fairness and compliance with law. That is why the use of Agentic AI by government institutions should start with support in coordination and documentation and not with decision-making on issues such as rights, benefits, admission, grading and disciplinary action. This is particularly applicable in education, where decisions can affect students’ futures.

Future Research Agenda

Future research should test the ADO model in real organizations and across different public-sector settings. The following research questions need to be addressed:

• Which organizational tasks are suitable for bounded autonomy and which should remain human-only?

• How do staff members calibrate trust when an agent provides evidence, confidence levels and explanations?

• Which governance controls reduce over-reliance without making the system too slow to be useful?

• How can agentic systems be evaluated for fairness and service equity across schools with different resources?

• What audit evidence is necessary for managers, regulators and citizens to understand agent-supported decisions?

Agentic AI can potentially play an integral part in future information systems since it will be able to integrate reasoning, tool use, workflow execution and feedback processes. Nevertheless, for its potential to be realized, careful development will be required. The intention here is not to exclude humans from participation in organizational decision-making but to create systems which will allow humans to cooperate in their work, have access to trustworthy information, behave consistently and have an audit trail. Within the framework of the ADO approach, agentic capability becomes an organizational advantage when combined with decision rights, trust calibration and governance. For instance, an example of School Support Service Agent can show how governmental education organizations can develop their systems starting from administrative help and retaining power in human hands.

Limitations

The current paper is a conceptual one and can be applied to some kind of design task. The current research does not provide any results obtained from experiments or field observations. Thus, all advantages that are mentioned must be viewed as hypotheses until proved by means of a pilot study and user testing. The proposed school assistance application can be considered as relatively safe; it should not be used for students’ assessment, admission, discipline or eligibility determination.

1. Bommasani, Rishi et al. “On the Opportunities and Risks of Foundation Models.” arXiv, vol. 2021, 2021. https://arxiv. org/abs/2108.07258.

2. Wang, L. et al. “A Survey on Large Language Model Based Autonomous Agents.” Frontiers of Computer Science, vol. 18, no. 6, 2024. https://doi.org/10.1007/s11704-024-402 31-1.

3. Joyce, K. and T.M. Cruz. “A Sociology of Artificial Intelligence: Inequalities, Power and Data Justice.” Socius: Sociological Research for a Dynamic World, vol. 2024, 2024. https://doi.org/10.1177/23780231241275393.

4. Alter, S. “Re-Conceptualizing Information System Usage as Usage by (Increasingly Automated) Work Systems.” Communications of the Association for Information Systems, vol. 54, no. 1, 2024, pp. 1094-1123. https://doi.org/10.17705/1CAIS.05446.

5. Russell, S. and P. Norvig. Artificial Intelligence: A Modern Approach. 4th Ed., Pearson, 2021.

6. Jennings, N.R. et al. “A Roadmap of Agent Research and Development.” Autonomous Agents and Multi-Agent Systems, vol. 1, no. 1, 1998, pp. 7-38. https://doi.org/10. 1023/A:1010090405266.

7. Simon, H. “A Behavioral Model of Rational Choice.” The Quarterly Journal of Economics, vol. 69, no. 1, 1955, pp. 99-118. https://doi.org/10.2307/1884852.

8. DeLone, W.H. and E.R. McLean. “The DeLone and McLean Model of Information Systems Success: A Ten-Year Update.” Journal of Management Information Systems, vol. 19, no. 4, 2003, pp. 9-30. https://doi.org/10.1080/0742 1222.2003.11045748.

9. Mikalef, P. and M. Gupta. “Artificial Intelligence Capability: Conceptualization, Measurement Calibration and Empirical Study on Its Impact on Organizational Creativity and Firm Performance.” Information & Management, vol. 58, no. 3, 2021. https://doi.org/10.1016/j.im.2021.103 434.

10. Lee, J.D. and K.A. See. “Trust in Automation: Designing for Appropriate Reliance.” Human Factors, vol. 46, no. 1, 2004, pp. 50-80. https://doi.org/10.1518/hfes.46.1.50_30392.

11. Glikson, E. and A.W. Woolley. “Human Trust in Artificial Intelligence: Review of Empirical Research.” Academy of Management Annals, vol. 14, no. 2, 2020, pp. 627-660. https://doi.org/10.5465/annals.2018.0057.

12. National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1, 2023. https://doi.org/10.6028/NIST.AI. 100-1.

13. OECD. OECD AI Principles: Updated Principles for Trustworthy AI. OECD.AI Policy Observatory, 2024. https://oecd.ai/en/ai-principles.

14. ISO/IEC. ISO/IEC 42001:2023: Information Technology-Artificial Intelligence-Management System. International Organization for Standardization, 2023. https://www.iso. org/standard/42001.

15. Simon, H.A. Administrative Behavior: A Study of Decision-Making Processes in Administrative Organizations. 4th ed., Free Press, 1997.

16. Enholm, I.M. et al. “Artificial Intelligence and Business Value: A Literature Review.” Information Systems Frontiers, vol. 24, no. 5, 2022, pp. 1709-1734. https://doi. org/10.1007/s10796-021-10186-w.

17. Parasuraman, R. and V. Riley. “Humans and Automation: Use, Misuse, Disuse, Abuse.” Human Factors, vol. 39, no. 2, 1997, pp. 230-253. https://doi.org/10.1518/00187209 7778543886.

18. Wei, J. et al. “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models.” Advances in Neural Information Processing Systems, vol. 35, 2022, pp. 24824-24837.

19. Yao, S. et al. “ReAct: Synergizing Reasoning and Acting in Language Models.” International Conference on Learning Representations, 2022. https://openreview.net/forum?id= WE_vluYUL-X.

20. Schick, T. et al. “Toolformer: Language Models Can Teach Themselves to Use Tools.” Advances in Neural Information Processing Systems, vol. 36, 2023.

21. Shinn, N. et al. “Reflexion: Language Agents with Verbal Reinforcement Learning.” Advances in Neural Information Processing Systems, vol. 36, 2023, pp. 8634-8652.UNESCO. Guidance for Generative AI in Education and Research. United Nations Educational, Scientific and Cultural Organization, 2023. https://unesdoc.unesco.org/ ark:/48223/pf0000386693.

22. European Parliament and Council of the European Union. Regulation (EU) 2024/1689 Laying Down Harmonised Rules on Artificial Intelligence. Official Journal of the European Union, 2024. https://eur-lex.europa.eu/ eli/reg/2024/1689/oj.