AI TECHNOLOGY
At its heart, artificial intelligence differs from traditional software because it uses probability, not instructions, to get work done. This gives it enormous power to handle real world inputs, without requiring expensive reformatting and collection. But it also means that outputs are not automatically reliable.
We address this by engineering AI systems that surround the core AI models with guardrails. Think of AI like a smart but inexperienced worker. It can read fast, write fast and find things fast, but it needs supervision and clear direction, just like a green apprentice.
AI RISKS
Leaders should consider six main risks as they assess AI strategy and implementation:
- Security: AI systems are not inherently prone to cybersecurity risk than other cloud software. An AI model will not remember what it has worked on, contrary to popular misconception. However, AI has unique risks, because we are exposing our software to more of the real world than traditional software. The biggest concern amongst cybersecurity specialists is “prompt injection,” which means the AI is exposed to malicious instructions because of a document, website or other source that tells it to do unwanted things.
- Mitigation: Ensure the IT department is trained and aware of AI-specific risks and recommends and enforces policies to avoid them. This should not be a heavy lift but needs to be ongoing as new risks evolve.
- Agentic control: The power of AI agents is that they develop their own plans and execute them. But, just like a human, they can go off track. This can become problematic because it wastes time and resources. Also, whenever you give an agent access to tools and resources, it can do things you’d prefer it not, like delete or alter files.
- Mitigation: Ensure extensive testing, strict permissioning and periodic testing.
- Easy Button: Workers trust AI because it sounds confident and looks polished, but it can lead to over-reliance on a tool that still makes mistakes..
- Mitigation: Treat AI training the same way you’d treat safety training. Essential, non-negotiable and ongoing. Basic training on how to use AI well is not expensive and does not need to be extensive. Part of this training should be in-person groups, where coworkers learn from each other.
- Overwhelm: In an industry with long hours and stressful days, AI presents another challenge: burying people in information faster than they can process it.
- Mitigation: Train workers how to think about integrating AI into their work and how to effectively instruct AI. Long, overdone answers are not inevitable, and brief training can make them both aware of the problem and how to get the right number of inputs.
- Overautomation: More than one AI vendor is promising automation of key workflows. This sounds attractive, but there is a reason experienced workers do certain things. Good AI solutions maximize context and opportunity for workers to apply judgment and creativity where needed, while automating the supporting functions that make judgment possible.
EDITOR’S NOTE: For March/April and May/June, I am honored to cede my article space to SMACNA’s AI Leadership consultant, Hugh Seaton. Hugh has been providing valuable insights, webinars and thought leadership to our members over the last two years. Visit SMACNA’s Construction Technology & AI site (www.smacna.org/business-resources/business-management/construction-technology-ai) to learn more about the topic. This is part one of two.
