The Management of Intelligence
Why AI projects fail less because of prompts than because of delegation, control and judgment.
We are on the “Jagged Frontier.” The limit of what AI can do is not linear. It is jagged, unpredictable and constantly shifting. Tasks that yesterday required human intuition can be solved today. Tasks that seem logically simple can still cause models to hallucinate.
Many companies are still navigating that boundary without a map.
The compliance gap
The symptom: the gap between technical feasibility and operational implementation is growing. While the marginal cost of cognitive work falls, the need for control, approval, and traceability rises.
We are seeing a shadow IT phenomenon here at both software and process level. Employees use large language models to take shortcuts. The result is often impressive. The process is often a black box.
In traditional organizations, an employment contract defines what an employee is allowed to do. There is liability, holiday entitlement and protection against dismissal. In software there are service level agreements and deterministic code.
Generative AI does not fit neatly into either category. It is not a pure tool, because it prepares decisions. It is not a legal entity, because it is not liable.
This compliance gap leads to a reflex in many management teams: ban usage and lose productivity, or ignore usage and accept risk. Neither is a strategy.
The Prompt Engineer Fallacy
A common misconception dominates the current debate: the assumption that prompt engineering is the skill of the future.
That falls short.
The ability to write complex prompts becomes less valuable as models understand context, intent, and examples. Prompting remains useful, but it is not the real bottleneck.
Experiments from the Wharton School and work by Ethan Mollick show one thing above all: Generative AI strengthens product and management work when the goal, context and evaluation are clear.
Companies are currently investing heavily in technical training. They teach employees how to talk to the chatbot. That can help, but it does not solve the deeper problem.
The real problem is the lack of delegation skills.
Many people have forgotten, or never learned, how to formulate tasks so precisely that an external person can complete them without questions. In a human context, an employee compensates for unclear instructions through judgment and implicit knowledge. AI does not do that reliably. It follows the instruction, including flawed logic.
We are seeing a return of classic management virtues: clarity, structure, goal definition. Technology exposes bad briefings. An imprecise instruction given to a person leads to questions. An imprecise instruction given to AI leads to bad assumptions or hallucinations.
The Solution Shift: The Return of Taste
When the production of content, code, images and strategies becomes cheap and readily available, value shifts in the economic system.
Away from the economy of skill, the question of how well you can do something. Toward the economy of will: how well can you judge what should be done?
Judgment, or taste, becomes the scarce capability.
Today we often still pay seniority for experience in execution. Tomorrow, we will pay more for the ability to select. AI can design ten variants of a go-to-market strategy in seconds. The value lies not in the design alone, but in deciding which variant carries.
This shift also affects the liability problem.
AI cannot be sued like a company or a person. Therefore, it must not be the final authority in relevant decisions. Human-in-the-loop is not a phrase; it is a question of responsibility. The role of people changes nonetheless. They are no longer always the creator. More often, they become the guarantor.
The senior manager signs off on the AI's result and takes responsibility. That is precisely where the manager's economic contribution lies.
The goal is not to abolish humans. The goal is to move people from execution into rule-setting. People define the laws. AI executes within those laws.
In this economy, the fastest person doesn't win. The one with the best judgment wins.
The logic: Frameworks are the new prompts
How do you delegate to an entity that has a lot of knowledge but no common sense?
Use proven logic frameworks.
Management frameworks from the 20th century are currently experiencing a renaissance as architectural concepts for AI agents.
A prompt like “Write me a marketing strategy” is worthless. The result will be generic. A task based on a strict framework reduces the search space and forces the model into a traceable structure.
- RACI matrix as a control element: if we define the role of the digital employee as responsible, accountable, consulted, or informed, its output behavior changes. It better understands which role it should simulate and where its limits are.
- PESTEL analysis as a context filter: instead of allowing a market analysis to be written freely, the agent is forced through the grid of political, economic, social, technological, environmental, and legal factors. This keeps critical dimensions from disappearing.
- Minto pyramid for communication: if you instruct the agent to communicate according to Barbara Minto's Pyramid Principle, the output becomes decision-ready faster: key message first, arguments later.
These frameworks act as guardrails for neural networks. They reduce the AI's search space and enforce structure.
It is a return to fundamentals. Frameworks taught in business schools decades ago are often more useful than modern prompt-hack templates.
Next step: make decision-making competence measurable
The integration of digital employees is not just an IT project. It is also an HR and organizational process.
We should start interviewing the AI.
When we hire someone, we ask for work samples. We test problem-solving skills under pressure. This is exactly what should happen with digital employees.
- Define the job, not the prompt. Create a job description for the AI.
- Provide context. Give the AI the relevant decisions, documents, emails, strategy sheets, and rules.
- Evaluate the result. Give feedback on the logic, not just the text.
Companies that take this step cleanly link growth less tightly to manual admin effort. The advantage lies not in better AI alone, but in a better architecture of work.
The foundation is there. Now management has to turn it into a controllable system.