--- title: "Agentic AI" date: 2025-10-01 description: Agentic AI, stripped of hype: generative models inside goal-directed loops that can use tools and act. Why delegated authority belongs on the Board agenda. author: Mario Thomas canonical: https://mariothomas.com/signals/agentic-ai/ --- ## What is Agentic AI? {#what-it-is data-toc="What it is"} Much of today's agentic AI places a generative model inside a goal-directed loop and gives the system authority to plan, use tools, evaluate progress, and decide what to do next. Used as a conventional assistant, a generative model produces an output and waits for a person to decide whether it is good enough, what to fix, and what should happen next. An agentic system is authorised to make more of those intermediate choices itself: it plans against a goal, invokes tools and APIs under scoped permissions, critiques the output, and iterates until the work meets the condition it was set. The organisation still defines the objective and the boundaries; the human no longer directs every step. The models underneath are familiar. What changes is the architecture around them and the authority it is given. The pieces assembled quickly. Agentic systems can now operate real software directly, an open standard has emerged for connecting them to enterprise tools and data, and major platforms have moved from research demonstrations into commercial products. [Gartner predicts](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027) that more than **40%** of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls. That divergence between commercial momentum and project attrition is the signal. Organisations can spend heavily on agentic capability and still destroy value if they delegate judgement without deciding where that delegation belongs. ## Why it matters to Boards {#board-implications data-toc="Board implications"} An agentic deployment is not merely a tool purchase; it is a delegation of decision authority. When a human drives the loop, they make countless micro-decisions: whether the work is complete, whether it answers the question, which source to check next, and when to stop. When the system drives the loop, it exercises more of those judgements autonomously, at a scale and speed no management reporting line can observe action by action. The Board cannot inspect every intermediate action, but it remains accountable for the governance system that determines what the agent may do, when it must escalate, how its actions are recorded, and who owns the consequences. This is the shift I have been writing about since [The Board in the Machine](/blog/the-board-in-the-machine/): organisations moving from overseeing decisions made by people to governing decisions increasingly shaped or executed by systems, while accountability remains with the humans who delegated the authority. The second reason for Board attention is attrition risk. Gartner attributes its forecast of widespread cancellations to escalating costs, unclear business value, and inadequate risk controls. Agentic pilots also inherit familiar problems from earlier AI programmes, including weak workflow integration and poor measurement, then add failure modes of their own because autonomous operation can take actions a copilot cannot. The organisations positioned to avoid that attrition will treat agentic AI as bounded delegation: failure must be recoverable, success criteria must exist before the spend, and escalation and audit mechanisms must be tested as part of the pilot. Management should also establish whether autonomous loops are already operating informally or outside approved controls. Their existence should be discovered through an inventory, not assumed either way. ## Questions Boards are asking {#questions data-toc="Questions"} ### Is agentic AI new, or rebranded automation? The degree of delegated control is the important change; the underlying components are not all new. Traditional automation executes steps defined in advance, while a conventional copilot generates output for a person to evaluate. An agentic system can make more of the judgements in between: whether the output meets the objective, what to improve, which tool to invoke, and when the work is done. That is a meaningful transfer of decision authority, even though the models, software interfaces, and automation techniques beneath it have longer histories. ### Where should we allow autonomous operation first? Where the work is high-volume, success criteria are clear, and failure is recoverable: routine enquiries with established resolution patterns, document processing that iterates to a quality threshold, internal workflows with bounded blast radius. Hold back the domains where proprietary expertise, ethical nuance, or high-stakes consequences dominate: strategic decisions, complex negotiations, anything customer-affecting where an error is not correctable. The strategic filter is simple to state: delegate where scale outweighs expertise, and keep humans in the loop where expertise outweighs scale. ### What governance must be in place before we say yes? Treat it like any other delegation of authority. The system needs defined operating boundaries, meaning what it may decide independently and under which permissions; escalation triggers for the edge cases; a tamper-evident activity record that reconstructs the information supplied, tools invoked, actions taken, approvals obtained, and outcomes produced; and intervention controls that scale from pausing a run or narrowing permissions to a full stop, each with a named human owner. Every agent also needs a distinct identity: an accountable sponsor, time-bounded permissions, no more authority than the task requires, and a separate approval before it can commit spend or communicate externally. The governance should enable safe delegation rather than prevent it: just enough structure to demonstrate good-faith oversight without destroying the speed that justifies the delegation in the first place. ### How do we avoid being in Gartner's 40%? Gartner attributes its forecast to escalating costs, unclear business value, and inadequate risk controls. Define the business outcome first, redesign the surrounding process to receive the automation, validate escalation and audit mechanisms as part of the pilot, and compare autonomous performance with the human baseline. A bounded pilot that shows the delegation is both safe and valuable is worth more than several demonstrations that prove only that the technology can run. ### Are agents already operating inside our organisation? Possibly, and management should find out. Consumer and enterprise tools increasingly include capabilities that can plan and act across multiple steps, but the presence of unapproved deployments is an empirical question rather than something to assume. A time-limited disclosure exercise, designed to identify use without punishing the people who disclose it, can turn an invisible exposure into an inventory the organisation can govern and may also surface useful applications already discovered by the workforce. ### What should management be able to show us this quarter? Three things. An honest inventory of where autonomous iteration already operates, sanctioned or not. A delegation map: which processes are candidates for agency transfer because scale beats expertise, and which are excluded because expertise beats scale. And for any proposed pilot, the success metric, the boundary conditions, and the governance mechanisms it is designed to validate. If management can show only vendor decks and enthusiasm, that is the finding. ## References - **GitHub** (March 2023): [Significant-Gravitas/AutoGPT: AutoGPT is the vision of accessible AI for everyone, to use and to build on. Our mission is to provide the tools, so that](https://github.com/Significant-Gravitas/AutoGPT). AutoGPT is the vision of accessible AI for everyone, to use and to build on. Our mission is to provide the tools, so that you can focus on what matters. - Significant-Gravitas/AutoGPT - **Anthropic** (22 October 2024): [Introducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku](https://www.anthropic.com/news/3-5-models-and-computer-use). A refreshed, more powerful Claude 3.5 Sonnet, Claude 3.5 Haiku, and a new experimental AI capability: computer use. - **Salesforce** (29 October 2024): [Salesforce’s Agentforce Is Here: Trusted, Autonomous AI Agents to Scale Your Workforce](https://www.salesforce.com/news/press-releases/2024/10/29/agentforce-general-availability-announcement/). Agentforce delivers fully customizable autonomous AI agents that can connect to any enterprise data and take action across sales, service, marketing, and - **Anthropic** (25 November 2024): [Introducing the Model Context Protocol](https://www.anthropic.com/news/model-context-protocol). Anthropic open-sources MCP, an open standard for connecting AI assistants to the systems where data lives: the tool-connection layer agentic systems depend on. - **OpenAI** (23 January 2025): [Introducing Operator](https://openai.com/index/introducing-operator/). OpenAI's research preview of Operator, an agent that uses its own browser to perform tasks on the web: the consumer face of the agentic shift. - **Gartner** (25 June 2025): [Gartner: Over 40% of Agentic AI Projects Will Be Canceled by End 2027](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027). The prediction that over 40% of agentic AI projects will be cancelled by end-2027, largely through impatience with early returns. - **European Commission** (2026): [AI Act](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai). The Commission's implementation page: the Act becomes generally applicable on 2 August 2026, with staggered high-risk deadlines to 2028.