Current AI systems are a world apart from earlier symbolic AI systems that used manually-constructed rules and could explain their reasoning. Instead of using expert-crafted rules, deep learning AIs train on data and learn decision criteria. Current AI technology is powerful but far from perfect.

What do AIs do well and what are their limitations? How can they complement human teams and how do they work?

The state of the art is changing quickly. AI Engineering (AIE) helps organizations to get from AI promises to reliable AI applications. It creates compound systems of interacting AI agents that work with people. The agents support sensemaking and reasoning and provide explanations.

• Sensemaking is collecting, organizing, summarizing, and extracting information. Sensemaking with large language models helps users answer questions based on information in document collections. However, since LLMs can mix information indiscriminately from different contexts, LLMs by themselves are too errorful for critical applications and require careful checking.
• Reasoning technology supports AI models about how the world works. It is needed for effective and accurate scheduling, design, manufacturing, construction, logistics and other complex activities.

This paper explains the research breakthroughs, the state of the art, and the whitespace for advancing AIE. It was invited as a follow-on paper to an earlier 2019 paper on explainable AI (XAI) (Gunning et al., 2019) that received a Frontiers of Science award at the 2025 ICBS conference in Beijing.

The featured image is intended to convey how the AI agents created using AIE work together with people in an organization. The image was created by Perplexity.

Paper

Stefik, M., Gunning, D., Choi, J., Miller, T., Stumpf, S., Yang, G.-Z. 2025. AI Engineering Primer. (This paper will appear in the Proceedings of the 2026 ICBS Conference.) Dropbox Link