Design Systems That Guide AI Instead of Just Asking Questions.

Summary

MIT SMR argues that professionals must move beyond simply prompting chatbots. According to the authors, this involves configuring agentic systems by defining context, capabilities, and orientation. This skill is termed 'intelligence steering.' The goal is to uncover insights that limited human questions might miss.

Why it matters

Agentic AI is shifting from chatbot Q&A to continuous, multi-step analytical workflows. In this scenario, the bottleneck is not model capability, but the ability to frame what the system should focus on. The article urges managers to view AI configuration not as an IT handoff, but as a core professional discipline. This approach could reshape how analytics, strategy, and discovery work are organized.

Highlights

• Three pillars of steering an agent: context (accessible data), capabilities (what it can do), and orientation (what it focuses on).
• Running multiple agents with different orientations on the same dataset enables comparing conflicting interpretations rather than just obtaining a single answer.
• The authors ground their framework in two published studies rather than vendor marketing: one on AI-assisted abductive theorizing, the other on organizations as algorithmic compilations.

Critical take

The piece reads more like a conceptual manifesto than a practical guide. While the three-pillar framework is intuitive, the article offers no concrete workflows, failure examples, or evidence that 'steering' outperforms well-structured multi-turn prompts in real analytical tasks. Labeling this skill 'intelligence steering' risks rebranding what experienced data scientists have done for years as a new executive competency. The 19-minute duration is largely spent on framing rather than content.