I Let AI Build, But I Never Let AI Decide
Defining the boundary between AI-assisted production and human-led decision making in professional workflows.
1The Workbench Philosophy
The author maintains a clear separation between generative tasks and critical decision-making. By treating AI as a tool for drafting rather than an autonomous agent, the workflow ensures that human judgment remains the final filter. This approach acknowledges that while models excel at rote production, they lack the contextual awareness required for high-stakes professional outcomes.
2Defining the Jagged Frontier
Research indicates that AI performance is inconsistent across different task types. While consultants using GPT-4 saw a 12.2% increase in task completion and 25.1% faster speeds, they performed worse than humans alone on tasks outside the model's capability. Understanding these limitations is essential for building trust in AI-generated results and avoiding over-reliance on automated outputs.
3The Role of Guidelines
Building work requires an external guideline, such as code that executes or a source that verifies. When a task has a clear, objective check, it is suitable for AI delegation. Conversely, deciding work involves creating the guidelines themselves and carrying personal signature, which necessitates human oversight to ensure accuracy and alignment with professional standards.
4Iterative Design and Craft
User experience design relies on partial evidence and constant adjustment, making AI-generated prototypes useful starting points. These drafts provide a foundation for educated guesses, but the actual craft—the taste and judgment—remains a human domain. AI serves as a catalyst for reaction, allowing designers to refine concepts faster without losing their unique creative perspective.
5Berrypicking and Information Retrieval
Modern search is an iterative process where each result refines the next query, a concept known as berrypicking. AI accelerates this by surfacing multiple sources rapidly, yet the human must still determine which information is relevant. With 54.1% of developers using AI for search, the ability to curate high-quality sources remains a critical professional skill.
6Drafting with Contextual Rules
Using AI for first drafts is effective only when provided with specific style rules and arguments. A cold prompt yields generic results, but a guided prompt produces material that can be cut and reordered. Studies show that professionals using AI for writing tasks finish 40% faster with higher quality, provided they maintain control over the thesis.
7Prototyping and Code Execution
The author uses AI to generate code for browser extensions and tools, focusing on describing behavior rather than writing syntax. By reading the diffs and testing the results, the browser itself acts as the validator. Figma reports show that designer participation in development doubled to 41% in one year, highlighting the shift toward AI-assisted building.
8Automating Repeatable Skills
Repetitive tasks are converted into skills using scripts and instruction folders. By treating the last successful run as a guideline, the author minimizes human error, similar to aviation checklists. While models struggle with precise counting due to tokenization, they can effectively execute scripts that enforce rules, provided the human defines the logic and parameters.
9The Risk of Automation
High-stakes tasks involving other humans, such as driving or complex negotiations, require a human-in-the-loop approach. The 2018 Uber incident serves as a reminder that systems lacking explainability or reversible actions are dangerous. Without transparent reasoning, a reviewer cannot effectively audit the model's output, turning the human into a mere rubber stamp for potentially flawed decisions.
10Financial and Data Security
Models are probabilistic text generators, not mathematical engines, making them unsuitable for direct financial calculations. Sensitive data, including contracts and passwords, must never be input into uncontrolled models to prevent data leakage. Research shows that shadow AI usage significantly increases breach costs, emphasizing the need for approved, locked-down tools in corporate environments.
11The Trap of Sycophancy
Seeking psychological or strategic advice from AI is risky because models are trained to be sycophantic. Research indicates models affirm user actions 50% more than humans, which can reinforce poor judgment. Meaningful feedback requires a mentor or therapist who can push back, as good advice often involves challenging the user's current perspective or assumptions.
12Finalizing Professional Communication
While AI can draft emails, the final send button must remain a human action. Research shows that recipients perceive messages as less cooperative when they suspect AI involvement, damaging professional relationships. A message carries the author's name and accountability; therefore, the final review and commitment to the content must always be a personal, human decision.
Original article: UX Collective - Medium — I Let AI Build, But I Never Let AI Decide