Generative AI & Computer Vision: Current Capabilities & Limitations

The Double-Edged Sword of Generative AI

As we've seen, generative AI can produce human-like text, summarize vast documents, and answer complex questions. These capabilities have obvious applications in your field, from drafting project reports to analyzing complex contract clauses. However, their probabilistic nature is a double-edged sword. Unlike a calculator, which is deterministic, an LLM provides the most statistically likely answer, not necessarily the factually correct one.

This leads to the most widely discussed limitation: hallucinations.

While hallucinations are a significant risk, especially in high-stakes environments like construction, it's also important to have a practical perspective on managing them. The technology is improving, and there are design patterns that can help.

Challenges in the Construction Context

The general limitations of AI become magnified when applied to a domain as complex and safety-critical as Engineering, Procurement, and Construction. An academic paper focusing specifically on this intersection provides a clear-eyed view of the hurdles.

The paper you just read highlights a comprehensive set of challenges. For your role as a consultant, these can be grouped into several key areas you must assess for any proposed AI project.

Let's break these down from a strategic viewpoint:

• Domain & Technology Challenges:

  • Knowledge Gap: A generic model like ChatGPT doesn't understand the specific definitions of "liquidated damages" or "force majeure" within a particular contract, nor can it interpret a P&ID diagram. It requires extensive fine-tuning on domain-specific data.
  • Accuracy & Hallucinations: An AI suggesting an incorrect material specification or misinterpreting a critical path could have catastrophic financial and safety consequences. Human-in-the-loop validation is not optional; it's a core requirement.
  • Generalizability: A model trained on residential construction projects in one country may not generalize to an industrial EPC project in another due to different regulations, terminology, and physical constraints.

• Adoption & Ethical Challenges:

  • Cost: The high computational cost to train, fine-tune, and run large models is a significant barrier, especially for startups.
  • Liability: If an AI-generated design fails, who is liable? The software vendor? The engineering firm that used it? The consultant who recommended it? This is a major unresolved legal and regulatory issue.
  • Data Privacy: EPC projects involve highly sensitive, proprietary data (e.g., costs, schedules, intellectual property). Using third-party AI services raises significant data security and IP concerns.

These points are especially salient in your area of interest: contract management. While AI promises huge benefits, the risks are equally large.

The Hard Economics of Computer Vision

Computer Vision (CV) is another transformative AI technology for the EPC sector, promising to automate tasks like job site safety monitoring, progress tracking against BIM models, and quality inspection. However, like generative AI, the gap between what is technically possible and what is economically viable is vast.

A recent groundbreaking study from MIT provides a framework for this economic analysis. It argues that simply identifying tasks "exposed" to AI automation is not enough; we must analyze the cost-effectiveness.

The key takeaway from the MIT paper is stark: at current costs, only 23% of worker wages paid for vision tasks would be attractive for a firm to automate.

Why is this number so low? The reason lies in the high upfront and ongoing costs of developing, deploying, and maintaining a production-ready AI system. The paper breaks these down into:

• Fixed Costs: Engineering time for implementation and maintenance. These are substantial, often running into hundreds of thousands or even millions of dollars for a robust system.
• Performance-Dependent Costs: The cost of acquiring and labeling the vast amounts of high-quality data needed to achieve the required accuracy. Cost grows exponentially as you push for higher accuracy.
• Scale-Dependent Costs: The cost of running the model (inference).

This economic reality leads to a crucial strategic insight for any consultant advising a startup: scale is paramount. The MIT paper finds that while it's rarely economical for a single firm to build a custom CV solution, the economics become much more attractive when a solution can be deployed across an entire industry—the "AI-as-a-service" model. This makes the "build vs. buy" decision one of the most important you will help a startup navigate.

The Consultant's Strategic Viewpoint

So, how do you reconcile the immense potential of these technologies with their significant limitations and costs? This is where your strategic role becomes vital.

1. Distinguish Between Building and Using AI: The MIT paper highlights the high cost of building custom AI systems. However, as Andrew Ng points out, the cost of using pre-existing models via APIs is falling dramatically. A key part of your consulting will be to guide startups away from reinventing the wheel and toward leveraging these powerful, cost-effective APIs wherever possible. Focus their limited resources on the "last mile" problem: integrating these APIs into a valuable business workflow.

2. Champion the Human-in-the-Loop: Given the limitations of accuracy and the high stakes in EPC, you should never frame AI as a full replacement for human experts. Instead, position it as a powerful augmentation tool.

3. Define the Right Roles for Human and Machine: A useful mental model is to separate the "what" and "why" from the "how."

As a consultant, your job, alongside the startup's leadership, is to define the what (the business goal, e.g., "reduce time spent on initial project scheduling by 50%") and the why (the purpose, e.g., "to allow our senior planners to focus on complex risk mitigation"). The AI's job is the how (the tactical execution, e.g., "generate a draft schedule from these design documents").

Conclusion

In this lesson, we have moved beyond the theoretical capabilities of AI to a pragmatic assessment of its real-world business limitations. This critical perspective is the foundation of effective AI strategy.

Key Takeaways:

• Generative AI is a Probabilistic Tool: Its power is tempered by significant risks like hallucinations, domain knowledge gaps, and ethical concerns. In high-stakes fields like EPC, human oversight is non-negotiable.
• Economics Often Trumps Technology: For Computer Vision, the high upfront cost of development and data acquisition means that most automation is not yet economically viable at the individual firm level.
• Scale is a Key Driver of Viability: Business models like "AI-as-a-service" that aggregate demand across many firms are often more practical than custom-built solutions, especially for startups.
• The Strategist's Role is to Balance: Your value as a consultant lies in navigating the tension between technological capability and business reality, guiding clients toward practical, value-driven applications of AI.

In our next lesson, we will synthesize everything we've learned so far. You will develop a structured framework for matching specific AI techniques—Predictive ML, Generative AI, Computer Vision, and NLP—to specific business problems. This will be the first major component of your personal consulting toolkit.