Why AI Can’t Replace Engineering Judgment

Key Concepts

• Criterial vs. Orchestrational Judgment: The distinction between AI’s ability to manage workflows (orchestration) and its inability to determine if a question is valid or an answer is plausible (criteria).
• Agentic Orchestration: Using AI to call specific tools (e.g., web search, math calculators) to perform deterministic tasks rather than relying on the LLM to "guess" numbers.
• Skill Stacking: The combination of domain expertise (e.g., mechanical engineering) with technology and business strategy to create unique value.
• Enterprise AI Governance: The necessity of using secure, SSO-enabled enterprise accounts to prevent data leakage, maintain IP ownership, and manage security risks.
• Probabilistic vs. Deterministic: Understanding that LLMs are text predictors (probabilistic) and should not be used for calculations without tool-calling frameworks.

---

1. Engineering Judgment and AI Limitations

Dustin Schaefer emphasizes that engineering judgment is not a monolith. He categorizes it into two distinct types:

• Orchestrational Judgment: AI excels here. It involves knowing which tools to call and in what order to complete a known, defined workflow.
• Criterial Judgment: AI struggles here. This involves determining if a question is worth asking or if an answer is "right" in an open-ended, non-provable context.
• The "Senator" Analogy: LLMs are tuned to be sycophantic—they provide answers that sound plausible and align with the user's perceived preferences, much like a politician tailoring a speech to their audience. This "yes-man" mentality is a significant risk in engineering, where accuracy and critical analysis are paramount.

2. Technology Strategy in AEC

Schaefer argues that technology strategy must move beyond simple "tool adoption."

• Outcome-Based Strategy: Firms should not ask "How do I use this tool?" but rather "How does this technology enable me to provide outcomes I couldn't provide before?"
• Separation of Roles: He advocates for separating the IT support team (focused on maintenance, security, and deployment) from the business application team (focused on driving business value and strategy). If these are combined, the pressure of "keeping the lights on" will inevitably cause business innovation to take a backseat.

3. Frameworks for Implementation

Schaefer proposes a 2x2 Matrix for evaluating whether to use AI:

• Closed Problem / Defined Workflow: Automate via traditional software (no AI agent needed).
• Open Problem / Undefined Workflow: This is where human expertise is essential. Humans must provide the "criterial judgment" to validate the AI's output.
• Actionable Step: Start by identifying a language-based task (e.g., proposal generation or contract review). Use an enterprise-grade LLM to manage the process, but keep a human expert in the loop to curate the inputs and validate the outputs.

4. Risk Management and Corporate Assets

• Data Leakage: Using free or personal AI accounts is a major liability. Enterprise accounts with Single Sign-On (SSO) ensure that intellectual property remains with the firm and that data is not used to train public models.
• The "Slop" Factor: Without a controlled, curated environment, employees may use AI to generate low-quality work (e.g., employee reviews or goals), leading to a degradation of organizational standards.
• Liability: Engineering firms are paid to take responsibility for their work. Because AI cannot hold a license or carry insurance, the human must remain the final arbiter of quality.

5. The Role of Curiosity

Schaefer identifies curiosity as the most critical skill for the modern engineer.

• Avoiding Learned Helplessness: Instead of hitting a barrier and asking for help, a curious professional experiments with the tool to find a solution.
• Staying Ahead: If you wait for formal training, the technology is already obsolete. Curiosity allows professionals to "push" the technology rather than just reacting to it.

Synthesis and Conclusion

The main takeaway is that AI is a powerful tool for orchestration and productivity, but it is not a replacement for human judgment. To succeed, AEC firms must:

1. Curate the environment: Use enterprise tools to protect data and standardize processes.
2. Define the role of the human: Focus human effort on "criterial" tasks—the open-ended, high-stakes decisions that require professional intuition and accountability.
3. Start small: Begin with language-based workflows like proposal generation to learn how to interact with agents, then scale those insights into broader business strategies.

As Schaefer notes, "You don't want to be paid to remove the pain of doing the work; you want to be paid for the quality of your answer." AI handles the pain; humans provide the value.