(2) The Agent Unlock: Why AI Needs Managers, Not Magicians
When a new graduate joins your team, you do not hand them a project, close your office door, and expect a perfect result by Friday. You brief them. You set milestones. You review early drafts. And yet with AI agents — tools now capable of persistent memory and independent action — most professionals do the opposite. They expect magic. They get disappointment. And they blame the technology.
The model matters, but deployment quality depends on organisational maturity.
The three tiers of human-AI interaction
At the first tier, AI is a search tool: one-shot queries, no memory, the human rebuilding context each time. At the second tier, it becomes a cowork tool — Claude Code, Cursor, Copilot — where the human remains present, instructing the AI to act on local files. Delegation surfaces, but the human must sit at the machine. The third tier is the agent unlock: a separate identity with persistent memory and independent channels, capable of working while you are away. This is where AI stops being a tool and becomes a report — someone you brief, trust, and verify.
The orchestration gap
Companies often deploy agents with expectations of autonomous work, then hit what platforms such as Coworker.ai and Google Agentspace call the “orchestration gap”: the distance between a capable model and a reliable business outcome. McKinsey’s State of AI report found that 65% of organisations regularly use generative AI, yet fewer than 15% have scaled beyond pilots. Model capability explains only part of that gap; organisational readiness explains much of the rest.
Agents arrive without organisational context
The failure is managerial, and the analogy is the new graduate. They arrive articulate and knowledgeable — much like a large language model. But they do not know your stakeholders, your unwritten rules, or the conversation you had about that project yesterday. As Andy Grove argued in High Output Management, the quality of a manager's output is the output of their team — which depends entirely on how well tasks are delegated, not how hard the manager works individually. An AI agent, like a new hire, arrives with high raw capability but low "task-relevant maturity" in your specific environment.
Ethan Mollick's research at Wharton confirms this. The strongest predictor of successful AI integration is not technical sophistication but "task decomposition" — the ability to break work into discrete, verifiable steps with clear checkpoints. The organisations winning with AI are the ones already good at briefing and feedback loops.
Fluency is easy to mistake for context
The cognitive trap is subtle. Because the model sounds confident and broadly informed, users assume it possesses contextual maturity. It does not. When ambiguous delegation produces plausible-sounding but misaligned output, the user concludes "AI doesn't work here" rather than "I briefed this poorly." The technology gets blamed for a management failure.
Better models will not remove the brief
There is a counterargument: that future models will infer intent better and shrink the managerial burden. This may hold for search and cowork use cases. But the point of an agent is independent action across time and context — and even a very smart employee still needs to know what "good" looks like in your organisation.
Enthusiasm can lead people to overestimate what unsupervised delegation can achieve. Managers who already break work down, set checkpoints, and give clear feedback have a more durable advantage: they know how to turn capability into dependable output.
AI will not democratise management skill. It will widen the gap between those who have it and those who do not.
Sources
-
https://cloud.google.com/blog/topics/generative-ai/google-agentspace-announcement
-
https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-in-2024
-
Grove, A.S. (1983) High Output Management. Random House.
-
Mollick, E. (2024) Co-Intelligence: Living and Working with AI. Wharton School research.