ECIS 2026 (2026) AI Processed Human Approved
HARNESSING AND REDUCING THE GENERATIVE AND WORKFLOW AUTONOMY OF GENERATIVE AI SYSTEMS:A MULTIPLE-CASE STUDY
This study examines how providers of generative artificial intelligence systems manage system autonomy post-deployment. Utilizing an exploratory multiple-case study methodology, the researchers analyze seven active systems deployed in business-to-business, business-to-consumer, and internal organizational contexts. The analysis integrates qualitative data from 14 semi-structured interviews and direct system testing to model autonomy management practices.
Problem
Generative artificial intelligence systems can autonomously generate content and orchestrate workflows, but managing this autonomy introduces risks like hallucinations, security vulnerabilities, and operational errors. Although literature emphasizes the necessity of managing AI systems, there is a lack of empirical research showing how system providers balance control and autonomy in live environments. This leaves practitioners with limited guidance on managing post-deployment risks without losing system utility.
Outcome
- The study conceptualizes AI autonomy as consisting of two dimensions: generative autonomy (content creation) and workflow autonomy (process and system interaction), which is qualitatively supported by the case analysis.
- The research identifies 20 distinct post-deployment practices categorized under either harnessing value-adding autonomy (such as real-time auditing and semantic resource enrichment) or reducing unproductive autonomy (such as architectural weaving and resource circuit breakers), which is qualitatively supported by the case analysis.
- The analysis identifies 'averting effects' where efforts to manage one dimension of autonomy require maintaining the other, which is qualitatively supported by the case analysis.
- The study identifies a 'potential' relationship where reducing workflow autonomy creates a structural foundation that allows providers to further restrict generative autonomy, which is qualitatively supported by the case analysis.
- The findings are limited by the exploratory, qualitative nature of the seven cases, which does not quantitatively measure the operational efficiency, safety, or economic impacts of the individual practices, as well as the limited temporal autonomy of current systems.
- The research identifies 20 distinct post-deployment practices categorized under either harnessing value-adding autonomy (such as real-time auditing and semantic resource enrichment) or reducing unproductive autonomy (such as architectural weaving and resource circuit breakers), which is qualitatively supported by the case analysis.
- The analysis identifies 'averting effects' where efforts to manage one dimension of autonomy require maintaining the other, which is qualitatively supported by the case analysis.
- The study identifies a 'potential' relationship where reducing workflow autonomy creates a structural foundation that allows providers to further restrict generative autonomy, which is qualitatively supported by the case analysis.
- The findings are limited by the exploratory, qualitative nature of the seven cases, which does not quantitatively measure the operational efficiency, safety, or economic impacts of the individual practices, as well as the limited temporal autonomy of current systems.
What it means for you
- CIO / IT Executive: On Monday morning, audit your organization's active GenAI integrations and mandate the immediate implementation of 'resource circuit breakers'—hard programmatic limits on API calls and database writes—to restrict workflow autonomy and prevent runaway system actions.
- IT Manager: On Monday morning, set up a 'real-time auditing' log pipeline for your team's customer-facing GenAI application to flag and review outputs where the system's confidence score drops below 80%, immediately routing those queries to human-in-the-loop agents.
- Business Strategist: On Monday morning, map your company's GenAI product features onto a 2x2 grid of Generative Autonomy vs. Workflow Autonomy to identify features where you can reduce risky workflow autonomy (e.g., auto-sending emails) without degrading the value of the generated content.
- Researcher: On Monday morning, draft a research proposal for a quantitative empirical study that measures the operational latency, safety rates, and economic impacts of 'architectural weaving' vs. 'semantic resource enrichment' across B2B GenAI deployments.
- Policymaker: On Monday morning, draft an update to your agency's AI procurement guidelines requiring vendors to prove they have 'averting effects' controls in place, specifically demonstrating how their systems constrain generative output risks when workflow autonomy is elevated.
Transcript
Host: Welcome back to A.I.S. Insights - Turning IS Research into Business Action. I'm your host, Aaron Ivan Sterling. Today, we're discussing a study from the European Conference on Information Systems, or ECIS 2026, titled "HARNESSING AND REDUCING THE GENERATIVE AND WORKFLOW AUTONOMY OF GENERATIVE AI SYSTEMS: A MULTIPLE-CASE STUDY". Joining me is our expert analyst, Ava Irene Solis. Ava, could you introduce us to what this study is about?
Expert: Thank you, Aaron. This study examines how providers of generative artificial intelligence systems manage system autonomy once those systems are deployed. The researchers used an exploratory multiple-case study methodology, looking at seven active systems deployed across business-to-business, business-to-consumer, and internal organizational contexts. By analyzing qualitative data from fourteen semi-structured interviews and direct system testing, they created a model of autonomy management practices.
Host: Let's discuss the real-world problem this study addresses. We know generative AI is becoming highly capable, but why is managing its autonomy such a challenge for organizations today?
Expert: The problem is that while generative AI systems can autonomously generate content and orchestrate workflows, this independence introduces significant post-deployment risks. Organizations face issues like content hallucinations, security vulnerabilities, tool misuse, and operational errors such as infinite loops. While there is plenty of literature on how to build these systems, there has been a lack of empirical research showing how system providers actually balance control and autonomy in live, production environments. This leaves business practitioners with limited guidance on how to manage these risks without reducing the system's overall utility.
Host: To address this gap, what was the researchers' approach?
Expert: The researchers conducted a qualitative multiple-case study of seven deployed generative AI systems that vary in their levels of autonomy. To capture practical, technical nuances, they gathered data from fourteen in-depth interviews with system creators, including founders, CEOs, and technical leads. They also combined this with first-hand system testing and archival data from industry events to understand the current practices used to manage these systems.
Host: What were the key findings regarding how we should think about AI autonomy?
Expert: First, the study conceptualizes AI autonomy as consisting of two distinct dimensions. There is generative autonomy, which is the system's independence in creating content, and workflow autonomy, which is the system's independence in how its components and external systems interact. The case analysis qualitatively supports this two-dimensional view.
Host: And how do providers actually manage these two dimensions in practice?
Expert: The research identifies twenty distinct post-deployment practices. These are categorized into either harnessing value-adding autonomy or reducing unproductive autonomy. For example, to harness generative autonomy, providers use practices like real-time auditing with automated AI judges and sequential agentic refinement. To harness workflow autonomy, they use semantic resource enrichment, which makes digital resources interpretable to the AI model.
Host: What about when they need to restrict autonomy to maintain control?
Expert: To reduce generative autonomy, they use practices like architectural weaving, which combines generative components with deterministic, rule-based services, and static epistemic bounding, like Retrieval-Augmented Generation, to limit the AI's knowledge space. To reduce workflow autonomy, they pre-structure the action space, use human-in-the-loop governance approval gates, or implement resource circuit breakers, which are automatic kill switches for budget, time, or looping.
Host: The study also mentions an interplay between these practices. What did they find there?
Expert: They identified "averting effects," where efforts to manage one dimension of autonomy require maintaining the other. For instance, reducing generative autonomy through autonomous epistemic bounding relies on keeping a certain level of workflow autonomy so the system can independently gather and pre-process context data. Conversely, they found a "potential" relationship where reducing workflow autonomy by pre-structuring the action space actually creates a foundation that allows providers to further restrict generative autonomy.
Host: Are there any specific limitations to these findings that practitioners should keep in mind?
Expert: Yes, the findings are qualified by the exploratory, qualitative nature of the seven cases. The study does not quantitatively measure the operational efficiency, safety, or economic impacts of the individual practices. It is also limited by the temporal autonomy of current systems, as today's deployed systems do not yet widely self-learn or rewrite their own logic over time without human intervention.
Host: Given these findings, why does this matter for business? What can practitioners actually do with this knowledge?
Expert: This study provides an empirically grounded framework for any organization trying to move generative AI from prototype to successful live operation. First, product teams can use these insights to deliberately design user experiences that accommodate autonomy safely. Second, AI operations teams can employ specific post-deployment monitoring practices, like real-time auditing and resource circuit breakers, to mitigate risks and control costs. Finally, managers can use the twenty practices for capability mapping, evaluating their current technical and governance structures to identify gaps and guide future AI strategic investments.
Host: That is highly practical advice for organizations looking to deploy AI responsibly. To summarize, managing generative AI requires a balanced approach that addresses both content creation and workflow execution, and this study provides a clear menu of practices to do just that. Ava, thank you for sharing your expertise with us today.
Expert: Thank you, Aaron.
Host: And thank you to our listeners. Tune in next time to A.I.S. Insights for more research-backed business strategies.
Expert: Thank you, Aaron. This study examines how providers of generative artificial intelligence systems manage system autonomy once those systems are deployed. The researchers used an exploratory multiple-case study methodology, looking at seven active systems deployed across business-to-business, business-to-consumer, and internal organizational contexts. By analyzing qualitative data from fourteen semi-structured interviews and direct system testing, they created a model of autonomy management practices.
Host: Let's discuss the real-world problem this study addresses. We know generative AI is becoming highly capable, but why is managing its autonomy such a challenge for organizations today?
Expert: The problem is that while generative AI systems can autonomously generate content and orchestrate workflows, this independence introduces significant post-deployment risks. Organizations face issues like content hallucinations, security vulnerabilities, tool misuse, and operational errors such as infinite loops. While there is plenty of literature on how to build these systems, there has been a lack of empirical research showing how system providers actually balance control and autonomy in live, production environments. This leaves business practitioners with limited guidance on how to manage these risks without reducing the system's overall utility.
Host: To address this gap, what was the researchers' approach?
Expert: The researchers conducted a qualitative multiple-case study of seven deployed generative AI systems that vary in their levels of autonomy. To capture practical, technical nuances, they gathered data from fourteen in-depth interviews with system creators, including founders, CEOs, and technical leads. They also combined this with first-hand system testing and archival data from industry events to understand the current practices used to manage these systems.
Host: What were the key findings regarding how we should think about AI autonomy?
Expert: First, the study conceptualizes AI autonomy as consisting of two distinct dimensions. There is generative autonomy, which is the system's independence in creating content, and workflow autonomy, which is the system's independence in how its components and external systems interact. The case analysis qualitatively supports this two-dimensional view.
Host: And how do providers actually manage these two dimensions in practice?
Expert: The research identifies twenty distinct post-deployment practices. These are categorized into either harnessing value-adding autonomy or reducing unproductive autonomy. For example, to harness generative autonomy, providers use practices like real-time auditing with automated AI judges and sequential agentic refinement. To harness workflow autonomy, they use semantic resource enrichment, which makes digital resources interpretable to the AI model.
Host: What about when they need to restrict autonomy to maintain control?
Expert: To reduce generative autonomy, they use practices like architectural weaving, which combines generative components with deterministic, rule-based services, and static epistemic bounding, like Retrieval-Augmented Generation, to limit the AI's knowledge space. To reduce workflow autonomy, they pre-structure the action space, use human-in-the-loop governance approval gates, or implement resource circuit breakers, which are automatic kill switches for budget, time, or looping.
Host: The study also mentions an interplay between these practices. What did they find there?
Expert: They identified "averting effects," where efforts to manage one dimension of autonomy require maintaining the other. For instance, reducing generative autonomy through autonomous epistemic bounding relies on keeping a certain level of workflow autonomy so the system can independently gather and pre-process context data. Conversely, they found a "potential" relationship where reducing workflow autonomy by pre-structuring the action space actually creates a foundation that allows providers to further restrict generative autonomy.
Host: Are there any specific limitations to these findings that practitioners should keep in mind?
Expert: Yes, the findings are qualified by the exploratory, qualitative nature of the seven cases. The study does not quantitatively measure the operational efficiency, safety, or economic impacts of the individual practices. It is also limited by the temporal autonomy of current systems, as today's deployed systems do not yet widely self-learn or rewrite their own logic over time without human intervention.
Host: Given these findings, why does this matter for business? What can practitioners actually do with this knowledge?
Expert: This study provides an empirically grounded framework for any organization trying to move generative AI from prototype to successful live operation. First, product teams can use these insights to deliberately design user experiences that accommodate autonomy safely. Second, AI operations teams can employ specific post-deployment monitoring practices, like real-time auditing and resource circuit breakers, to mitigate risks and control costs. Finally, managers can use the twenty practices for capability mapping, evaluating their current technical and governance structures to identify gaps and guide future AI strategic investments.
Host: That is highly practical advice for organizations looking to deploy AI responsibly. To summarize, managing generative AI requires a balanced approach that addresses both content creation and workflow execution, and this study provides a clear menu of practices to do just that. Ava, thank you for sharing your expertise with us today.
Expert: Thank you, Aaron.
Host: And thank you to our listeners. Tune in next time to A.I.S. Insights for more research-backed business strategies.