--- title: "Agentic AI: From Control to Autonomy" date: 2026-06-12 description: Agentic AI explained in fourteen minutes: what happens when generative AI is put in a loop, and the strategic question that follows — where to transfer agency, and at what scale. author: Mario Thomas canonical: https://mariothomas.com/videos/agentic-ai-explainer/ --- Agentic AI is the most talked-about technology in boardrooms this year, and for good reason. For the first time, we're not talking about AI that assists people, we're talking about AI that makes decisions on its own, at scale, without waiting for approval at every step. Here's the simplest, true description of it you'll hear. Agentic AI is generative AI in a loop. In the next few minutes, I'll show you exactly what that means, and the strategic choice it forces. For Boards, this creates a governance challenge of a genuinely new kind. When AI can make millions of decisions a second, rather than hundreds a day, operate around the clock rather than business hours, and coordinate whole workflows without anyone orchestrating it, the question is one Boards have never had to ask before. Where should we consciously transfer decision-making authority to a machine? This is not a technical question. It's a strategic one. Most conversations about agentic AI start with the technology, the models, the architectures, the capabilities, but that really is the wrong starting point. The technology isn't the constraint. The constraint is understanding what you're actually delegating, and having the frameworks to govern that delegation. So, I want to give you a way of thinking about agentic AI as the transfer of agency, but not accountability. It's how work changes as you progressively delegate decisions to AI, where autonomy creates genuine advantage, and what makes the delegation safe rather than reckless. Because the organisations that win with agentic AI won't be the ones with the most sophisticated technology, they'll be the ones that choose consciously where to transfer agency and where to keep human judgement. Before we get to agentic AI itself, we need a way of describing how any work gets done. By humans, by machines, or both. Every piece of work runs on the same eight components. Knowledge. What you know and how you reason about problems. Perception. Is how you gather and understand what's happening around you. Planning. Is how you break an objective into steps and decide what to do in what order. Action. Is how you actually execute the doing. Memory. What you retain and build on, rather than starting fresh each time. Adaptation. How you improve, based on what worked and what didn't. Orchestration. How you coordinate when many things must happen at once. And lastly, governance. How you keep the whole thing safe, compliant, and aligned with your values. These eight exist in every work pattern. What changes, and this is the entire story, is who or what performs each one. That's what tells you where authority is being transferred. Let me show you. Stage one is human-centred work. How your organisation has run for decades. Humans performing all eight components. Knowledge is human cognition. And hard-won expertise. Rules of thumb. Standard operating procedures. Perception is critical reading. Manual data gathering. Stakeholder interviews. Planning is checklists. Mental models. Heuristics. Built through experience. Action is people executing in applications. Email, ERP systems, and spreadsheets. Clicking and typing. Memory lives in people's heads. Notebooks. Shared drives. And the institution itself. Adaptation happens through after-action reviews. Mentoring and training. Orchestration is project managers. Coordination and meetings. And governance is policies. Peer review. Sign-off. Compliance. And audit. This is the baseline. Every component depends entirely on human capability. Humans are the loop. Now watch what happens when we introduce AI. Stage two is where AI arrives as an on-demand assistant. What most organisations are living right now with ChatGPT, Claude, and their peers. And I say AI model deliberately because this is bigger than large language models. Small specialised models, vision models, domain models, all belong here. But notice what hasn't changed. Humans still make every decision. So knowledge. The model is a copilot for ideas and drafts. The judgement stays yours. In perception, you curate the inputs. The model summarises and extracts. In planning, you outline. The model suggests options. You choose every move. For action, the model drafts and you execute. Nothing happens without you. And in memory, session context is kept only. And if you close the chat and it forgets, unless you take notes, you lose the chat. Adaptation. You just adjust your prompts. The learning is yours, not the system's. In orchestration, there isn't any. You're the integration. Copy-pasting between tools. And from a governance perspective, your oversight, your policies and whatever audit trail you create by hand is yours. Useful genuinely. But it's AI in the loop one pass at a time. With you between every step. You are still the bottleneck. Stage three is where sophisticated AI usage emerges. And where the loop becomes real. The AI gets invoked repeatedly, iteratively. But you hold the gate on every turn. In knowledge, models are invoked again and again with templated prompts against objectives and success criteria that you set. In perception, human and model parse documents and feeds together. Structured ingestion, light retrieval, with your approval at the key points. In planning, the model proposes a plan and even critiques its own work. You review, revise and accept. Action is semi-automated and workflows that are semi-automated appear here. Tool calls with your approval and humans execute anything irreversible. In memory, session memory plus retrieval over curated knowledge base happens and case logs start building institutional memory. In adaptation, your feedback tunes the prompts and policies. Occasionally, the model tunes itself. In orchestration, you are the conductor. Sequencing models and tools. Keeping gates between the steps. And in governance, review and approve gates. Logging, restricted permissions, escalation criteria and playbooks still belong to you. This is iterative AI. Powerful and well-governed, but you're still turning the handle. Your time and your attention remain the constraint on scale. Which brings us to the shift that matters. Stage four is agentic AI. The machine runs the loop itself. Not delegation of a task. Delegation of the whole decision-making cycle. And look at what changes on every row. In knowledge, the system selects its own models. A large model for reasoning. A specialist for a narrow task. A vision model for images. Without asking you which to use. In perception, event-driven ingestion from APIs, databases and sensors. The system decides what information it needs and goes and gets it. In planning, autonomous planning with self-reflection and dynamic replanning. The system asks itself, is this good enough? And adjusts without you. In action, direct execution through APIs and tools. Transactions created, tickets raised, activity scheduled, under scoped permissions. But with no approval sought per action. Memory, a working scratchpad. Plus a long-term memory that persists across runs. It remembers what it learned and applies it next time. In adaptation, automated evaluation and feedback loops improving the system's own behaviour. Governed but machine-driven. In orchestration, multiple agents working together as planners and executors and critics. Coordinating between themselves in parallel with no human in the hand-offs. And governance, policy engines, risk tiers, approval thresholds, sandboxes, immutable audit logs, kill switches. You built the guardrails and the system operates autonomously inside them. That last row is the one to remember. In Stage 4, governance is not something you do to the work afterwards. It's the thing you build before the work begins. Here's the critical insight for a Board. Agentic AI is not magic. And it's not a new kind of intelligence. Look across the table. Stages 2, 3 and 4 all use the same family of models. What changes is not the technology. It's the locus of control. In Stage 1, humans are the loop. In Stage 2, AI joins the loop one pass at a time with a human between every step. In Stage 3, the loop is real and a human holds the gate. In Stage 4, the machine runs the loop. So the strategic question is not whether agentic AI is impressive. It demonstrably is. The question is where handing over these eight components creates advantage and where human judgement in the loop remains essential. Where does scale outweigh expertise? Where is the high volume commodity work that benefits from continuous autonomous operation? And where are the irreversible decisions that demand a human being accountable? That is the choice agentic AI forces you to make consciously. And the eight components are the framework for making it systematically, row by row, rather than all at once. Everything in this video comes from the companion article, Understanding Agentic AI, which you'll find at mariothomas.com, along with everything else I write on AI and emerging technology for Boards and senior leaders. If you take one thing away, take this. Agentic AI is generative AI in a loop and the decision in front of you is not whether the loop is clever. It's where you're willing to put it and who answers for what it does. The transfer of agency is conscious, a conscious decision you make, and you keep accountability where it has always lived. Thanks for watching.