TLDR: A new human-AI coaching system combines a cognitive coaching model with an LLM to proactively support novice entrepreneurs and their mentors. The system challenges novices’ assumptions, helps mentors plan emotionally attuned strategies, and improves meeting depth and focus. Key findings from a field deployment show it uncovers blind spots, adapts support based on dual contexts, empowers mentors to customize AI logic, surfaces emotional and cognitive root causes, and orchestrates more effective human-human collaboration.
Entrepreneurship is a challenging journey, especially for new founders. They often face complex, undefined problems, like identifying risks or making strategic decisions without clear guidelines. This can be overwhelming, and even experienced mentors find it hard to provide tailored support due to limited time and visibility into a novice’s thought process.
Traditional tools like business templates or task trackers offer some structure, but they often fall short. They require novices to already possess advanced thinking skills to identify their own risks, and they don’t adapt to the ever-changing nature of a startup. Similarly, many AI tools are passive, simply responding to user input without proactively challenging assumptions or facilitating deeper reflection. This creates a gap in supporting the dynamic, human-centric process of entrepreneurship coaching.
Introducing a Proactive Human-AI Coaching System
Researchers at Northwestern University have developed a novel human-AI coaching system designed to bridge this gap. This system combines a specialized understanding of entrepreneurial risks with a large language model (LLM) to proactively support both novice founders and their mentors. The core idea is to move beyond simple information gathering and instead scaffold critical thinking and collaboration.
How Does It Work?
The system operates through two main interfaces: a chatbot for novices and a dashboard for mentors.
For novices, before a coaching meeting, they interact with a chatbot. This chatbot doesn’t just ask generic questions; it uses the novice’s project context to generate personalized, diagnostic questions. For example, instead of asking “What is your current focus?”, it might ask “What specific aspect of this AI system are you currently focusing on?” based on the novice’s project. After gathering information, the system identifies potential design risks and prompts the novice to reflect on these risks, encouraging deeper thought about uncertainties and assumptions.
Following this, novices receive a dashboard summarizing their project information and the diagnosed risks. They can then prioritize specific risks for their upcoming meeting, giving them agency in shaping the discussion.
Mentors, on the other hand, access a parallel dashboard. This dashboard provides a summary of the novice’s project, including the risks they prioritized and even those they chose to omit. This offers mentors valuable insights into the novice’s mindset. Based on the novice’s input and the mentor’s own coaching goals, the system generates tailored coaching strategies and questions. Crucially, mentors can inspect and even modify the underlying risk framework of the system, ensuring its logic aligns with their evolving coaching needs without requiring any coding.
The system is built using Python, Streamlit for the interface, LangChain with the LLaMA 3–70B language model for understanding and generating text, and Firebase for data storage. Its intelligence is rooted in a cognitive coaching model that captures how expert mentors think about risks and strategies.
Real-World Impact: Key Findings
An exploratory study involving one mentor and eleven novice entrepreneurs in a university incubator revealed several key benefits:
- Proactive Challenge: The system successfully uncovered “blind spots” for novices, bringing attention to risks they hadn’t considered. It also helped them reorient priorities away from easier tasks towards more critical challenges. While some novices initially distrusted the AI’s challenges or expected direct answers like ChatGPT, they later appreciated its reflective approach.
- Adaptive Support: By combining novice input with mentor goals, the system provided highly personalized diagnoses and coaching strategies. This made the risks feel more “real” to novices and the mentor’s advice more actionable. However, the system’s effectiveness was sometimes limited by the quality of information novices provided.
- Mentor Empowerment: Mentors found the system’s diagnoses highly relevant and could easily modify the risk framework to better reflect real-world scenarios, enhancing trust and adaptability.
- Surfacing Root Causes: The system helped novices feel more comfortable disclosing struggles, revealing underlying emotional and cognitive blockers (like perfectionism or fear of failure) that mentors might not have otherwise discovered. This allowed mentors to adjust their strategies with greater sensitivity.
- Orchestrated Collaboration: The asynchronous preparation for both parties led to more focused, productive, and in-depth coaching meetings. Novices entered discussions with clearer priorities, and mentors were better prepared to offer targeted guidance. Initially, some novices misunderstood the system as a reactive AI assistant, but after experiencing its complementary role with the human mentor, they recognized its value in making the coaching journey more focused.
Also Read:
- AI’s Hidden Flaws: Uncovering Cognitive Biases in General-Purpose AI for Software Engineering
- Holistic Explainable AI: A New Framework for End-to-End AI Transparency
Looking Ahead
This research marks a significant step beyond static templates and reactive AI tools. It proposes a new model where AI acts as a diagnostic scaffold and facilitator of judgment, enhancing human relationships and shared reasoning in complex, ill-defined domains. The principles derived from this study, such as proactive challenging, dual-context adaptation, expert-governed AI logic, and surfacing root causes, have broad implications for fields like healthcare training, education, and knowledge work.
While the study was limited in scale, it provides valuable insights into how AI can augment, rather than replace, human-human collaboration, fostering deeper understanding and more effective interactions. For more details, you can read the full research paper here.
