|
From Prompt to Agent: Why the Way You Use AI Is About to Change Completely
For most of the past three years, interacting with AI has meant one thing: typing a question and reading an answer. Even sophisticated users, those who write detailed prompts, use custom instructions, and maintain project contexts, are still fundamentally operating within a request-response model. You ask. It answers. You decide what to do next.
That model is ending. Not gradually. This week.
The launch of Copilot Cowork inside Microsoft 365 means that for the first time, an AI agent has been handed the keys to the most widely used enterprise software suite on the planet. A user can now give a single instruction and have an agent plan and execute a multi-step workflow across multiple applications simultaneously, without touching any of them manually. This is not a feature upgrade. It is an architectural shift in how professional work gets done.
The Three Layers of Agentic AI
It helps to think about agentic AI in three layers, each representing a different level of autonomy and a different set of organisational implications.
Layer 1: Assisted generation. This is where most organisations currently sit. AI helps with drafting, summarising, and generating content on request. The human remains in control of every step. Tools: standard Claude, ChatGPT, Copilot in its existing form.
Layer 2: Agentic execution with oversight. The AI takes a complex instruction, breaks it into steps, executes across multiple tools, and checks in at key decision points. The human sets the goal and reviews outputs at checkpoints. Tools: Claude Cowork (desktop), Copilot Cowork (enterprise), Claude Code, Cursor. This is where the frontier moved this week.
Layer 3: Scheduled autonomous operation. AI agents run on timers or triggers, operating in the background without human initiation. A weekly competitive intelligence brief assembled automatically. A Monday morning summary of the week's AI developments delivered before you arrive. This layer is not yet mainstream, but the architecture to support it is being built now.
Most UK organisations are operating at Layer 1 and treating it as sufficient. The problem is that competitors who move to Layer 2 this year will be operating at a structural productivity advantage that compounds over time.
The Governance Gap
The transition from assisted generation to agentic execution creates governance challenges that most organisations are unprepared for. When AI is assisting with a document, a human reviews every output. When AI is executing tasks across your enterprise data, email and calendar in the background, the oversight model has to be fundamentally different.
Microsoft recognised this in the Copilot Cowork launch. The announcement was accompanied by Agent 365, a separate platform designed for IT and security teams to monitor, govern and audit AI agents across an organisation. Microsoft's own chief marketing officer for AI noted that "AI agents are as subject to phishing attacks as people are. As soon as an AI agent has an email address, they get spam too, and they can respond to it."
The question for every leader deploying AI in 2026 is not "which tool do we use?" It is "what is our governance model for autonomous AI action inside our organisation?"
What to Do This Week
Map your Layer 2 opportunity. Identify three to five workflows in your organisation that involve multiple tools, multiple steps, and significant coordination time. These are your highest-value targets for agentic AI. Build the case now, before others make it for you.
Define your oversight model. What categories of action can an AI agent take without human approval? What requires a checkpoint? What is never delegated? These questions need answers at policy level, not just tool level.
Start with a personal context file. Before deploying agents at team or organisational level, build an AI context file that captures your role, priorities, communication style and recurring decisions. This is the foundation for agentic AI that produces outputs you would actually use. See this issue's Prompt Pocket for a prompt that builds this in one session.
The shift from prompt to agent is not a future scenario. It is a deployment schedule. The organisations that treat it as one will be ahead. The ones waiting to see how it plays out will be catching up.
|
