7 minutes to read With insights from... Tim Herfurth Lead Data Scientist tim.herfurth@zuehlke.com Implementing agentic AI systems: opportunities and risks Agentic AI systems can help users execute complex tasks that require planning and reasoning with a degree of autonomy. These systems can adapt to context and use tools to interact with other systems. Unlike traditional AI methods, they are able to adapt to their context and act with purpose. While they reduce the need for human involvement in routine manual tasks, these systems still operate under human oversight. This ensures accountability, alignment, and trust as they come to support increasingly complex workflows. With the ability to reason, make decisions, and work with reduced supervision, agentic AI promises to transform workflows, boost productivity, and collaborate with human teams. Gartner predicts that by 2028, 33% of enterprise software will include agentic AI, enabling 15% of business decisions to be made autonomously, up from less than 1% in 2024. But realizing this potential is going to take more than just assigning tasks to large language model (LLM) agents. It requires that careful thought is given to the design of the solution architecture, and to user and agent interactions, and it requires careful monitoring in order to reducing operational and ethical risks. Here we present a practical blueprint for implementing agentic AI solutions in a way that maximises value for your business, while ensuring reliability and control. 1. Define scope and establish a clear baseline First, implementation of agentic AI systems should be guided by a fundamental question: ‘What problem is the agentic system intended to solve?’. Too often, teams get caught up in the hype around new technologies and overlook the importance of clear problem framing. Before diving into solutions, it’s essential to frame the challenge by answering three key questions: What is the goal? What is the goal? Design the system with clear business objectives in mind. These need to be defined in concrete terms that reflect the intended value. For example, ‘automate customer support’ is too vague. A more specific goal would be ‘reduce average resolution time by 30%’ or ‘resolve 50% of tier-1 queries without human intervention’. What does success look like? What does success look like? You should measure progress and performance using well-defined key performance indicators (KPIs) and success metrics. These metrics might be task success rates, average response times, customer satisfaction scores, escalation rates, or operational cost reductions. Your metrics should be measurable, relevant, and tracked consistently over time. What are the constraints and boundaries? What are the constraints and boundaries? Consider the technical, ethical, and regulatory limits within which the system needs to operate. That might include ensuring fairness, avoiding unintended harm, maintaining data privacy, or restricting access to sensitive systems. Clearly documenting these boundaries helps guide design choices and ensures alignment with broader organisational principles and responsibilities. This scoping process ensures that your AI development remains measurable, auditable, and aligned with business goals. Start simple Agentic AI can be implemented with varying degrees of autonomy. Where possible, start with deterministic workflows that don’t require flexible problem-solving strategies. They are easier to understand, debug, and control, and serve as a baseline for evaluating more complex behaviours. Then gradually add autonomy to handle more complex or ambiguous situations. This staged approach helps manage risk and system complexity. 2. Structure the Agentic AI system Agentic AI solutions work better with a modular setup, where different parts handle different subtasks or process steps, rather than using one big, all-in-one system. Use an orchestrating main agent Agentic AI systems often have a hub-and-spoke architecture, with a central agent acting as a coordinator. This main agent interacts with users, plans solutions, and breaks down tasks into manageable subtasks. These subtasks are then handed off to specialised agents, each of which is responsible for handling a specific part. Once completed, the main agent brings all the results together. In essence, it functions like a project manager, delegating work and overseeing progress without micromanaging every step. This ‘divide-and-conquer' approach brings several benefits: it simplifies debugging, improves traceability, allows targeted upgrades, and enables agent specialisation. Incorporate natural language interfaces One of the transformative strengths of LLMs is their ability to understand and generate natural language. By integrating natural language interfaces, you can empower non-technical users to interact with agentic systems directly. This democratizes access and accelerates adoption across departments. A well-designed interface can turn a complex multi-agent system into a conversational partner that feels intuitive and collaborative. 3. Ensure control and safety Giving AI agents more flexibility and autonomy in solving tasks can be a great productivity booster, but it also introduces risks. You need to strike a balance between flexibility and controllability, and you need to adhere to responsible AI development standards. Combine autonomy with human oversight A high level of system autonomy may not always be appropriate, especially in sensitive domains like finance, healthcare, or law. In these contexts, to ensure safety, compliance, and trust, it’s essential that a balance is struck between automation and human oversight.Applying human-in-the-loop principles allows you to retain control over critical decisions. To maintain an adequate level of control, oversight and user trust, you might: Require human review for high-impact actions or publication of results.Design escalation paths for edge cases or low-confidence decisions.Apply rollback mechanisms to reverse unintended outcomes.Adhere to a responsible AI framework and align your design with emerging AI governance standards. 4. Continuously evaluate, monitor and improve As with any AI system, agentic AI solutions need to be rigorously evaluated during development to ensure they meet the expected performance standards. After deployment, these systems require ongoing monitoring and validation, as changes in user behaviour, data drift, and other external changes can impact performance over time. Best practices for evaluation An evaluation strategy for agentic AI systems should combine structured testing with real-world validation. It should include multiple layers: Automated testing: where feasible, test programmatically using curated test sets. This approach supports repeatability, faster iteration, and consistent quality checks throughout development.Manual testing: where automation isn’t feasible, get subject matter experts to review outputs using predefined scenarios, success criteria or questionnaires. Real-world testing: validate in stages, moving from internal users to pilot groups and full rollouts. Collect in-app feedback, run targeted surveys, and analyse logged interactions to identify issues and guide improvements. Enable continuous monitoring and improvement The environments in which agentic AI systems operate are dynamic – tools, data sources, APIs, and user behaviour change over time. These systems therefore need to be treated as living products, with built-in mechanisms for monitoring and adaptation. To support this, it's important to: Track usage and user satisfaction. Monitor engagement, satisfaction scores, and interaction patterns to understand performance from the user’s perspective.Implement observability. Use monitoring tools to track task duration, error rates, agent reasoning paths and system behaviour in real time.Establish feedback loops. Let users rate responses, flag issues, and suggest improvements. Use this input for prompt tuning, fine-tuning, and retraining workflows.Detect and address model drift. Monitor for shifts in data or behaviour that impact performance, and retrain models as needed.Maintain robust data foundations. Ensure agents have access to accurate, up-to-date, well-governed data. This includes maintaining clean data pipelines, enforcing access controls, and monitoring API/tool integrations. Conclusion: From hype to impact Agentic AI is more than just a technical innovation: it marks a shift toward systems with greater autonomy and problem-solving capabilities. AI agents can interpret objectives, plan actions, and adapt as they work toward their goals. Building solutions based on AI agents requires a carefully thought-out architecture, clear evaluation strategies, and continuous oversight to ensure reliability and alignment. Agentic AI is ultimately here to assist us. It should be designed to enhance human capabilities, while always allowing for full control, especially where risks can arise from its decision making. Talk to Zühlke about implementing agentic AI in your business At Zühlke, we work with organisations across a range of industries to design and implement agentic AI solutions that are aligned with your strategy, built for real-world conditions, and developed with responsibility at their core. If your organisation is exploring autonomous AI, talk to us today, and let us help you define your roadmap and deliver systems built for long-term success.