At the 2023 MIT Digital Technology and Strategy Conference, ten startups showcased how they are leveraging artificial intelligence to drive digital transformation across diverse industries. The companies span a wide range of applications — from Cleanlab, which uses AI to find and fix errors in datasets, to Zapata, which focuses on industrial generative AI for solving enterprise-scale problems using both language and numerical models. Other notable startups include CoCoPIE, streamlining AI deployment on mobile and IoT devices; Hopara, creating affordable digital twins; and MontBlancAI, applying anomaly-detection algorithms to manufacturing.

The lineup highlights the breadth of AI's current impact: Covariance builds competitive intelligence profiles by combining first-party, third-party, and public data; Einblick uses generative AI to help users write code and build models collaboratively; serviceMob provides AI-driven analytics for customer service operations; TechNext uses machine learning to assess emerging technology investment opportunities; and Wise Systems improves last-mile delivery routing. Together, these startups illustrate how small, agile companies are tackling significant challenges in data quality, logistics, manufacturing, and business intelligence by building on the latest advances in AI and machine learning.

New research from MIT Sloan School of Management challenges the prevailing narrative that AI will replace human workers, arguing instead that AI is more likely to complement human capabilities. The paper, titled "The EPOCH of AI: Human-Machine Complementarities at Work," introduces a framework built around five groups of uniquely human capabilities: Empathy and Emotional Intelligence, Presence, Networking, and Connectedness, Opinion, Judgment, and Ethics, Creativity and Imagination, and Hope, Vision, and Leadership. These capabilities, collectively called EPOCH, represent areas where machines are fundamentally limited due to statistical shortcomings like biased or small datasets, extrapolation challenges, and moral dilemmas.

The researchers — Roberto Rigobon and Isabella Loaiza — developed three key metrics to evaluate tasks across occupations: the EPOCH index, a risk-of-substitution score, and a potential-for-augmentation score. Their findings show that human-intensive tasks have actually increased between 2016 and 2024, and newly added tasks in the O*NET labor dataset in 2024 exhibit higher EPOCH capability levels than previously existing tasks. The study emphasizes that augmentation goes beyond partial automation — it enables humans to accomplish things they couldn't do before, much like advanced microscopes opened up nano-scale work.

The authors stress the importance of investing in developing workers' EPOCH capabilities, deliberately avoiding the label "soft skills." As Rigobon notes, teaching empathy, hope, and creativity is far harder than teaching technical problem-solving — yet these are precisely the skills that will keep human workers complementary to, rather than replaceable by, AI.

MIT Sloan professor Rama Ramakrishnan outlines a three-step framework for identifying the right generative AI use cases in business. The approach begins with breaking jobs into discrete tasks, since not all tasks within a role are equally suited for LLM automation — preparing course materials can be partially automated, while moderating discussions cannot. Next, organizations should apply a generative AI cost equation that accounts not just for licensing fees but also the less obvious costs of achieving the required accuracy, building error-detection mechanisms, and managing reputational risk. If the total cost of AI automation is lower than the status quo, the third step is to build, launch, and rigorously evaluate pilots, choosing from vendor solutions, commercial models like GPT-4, or open-source options like Llama 3.

Ramakrishnan also offers several best practices for successful implementation: maintain a robust error-checking process with a human in the loop, consider narrow and targeted use cases that require smaller and cheaper models, cultivate AI talent outside traditional data science teams, and prioritize low-stakes, high-volume use cases for quick ROI and organizational learning. He emphasizes that tasks failing the cost equation today should be revisited periodically, as rapidly improving LLM capabilities are steadily driving down adoption costs. The overarching advice is to avoid paralysis by starting with easy wins and scaling from there.

MIT Sloan professor Rama Ramakrishnan outlines three practical methods businesses can use to leverage large language models (LLMs) like GPT-4, Llama, and Claude. The simplest approach is prompting, where users directly ask an LLM to perform tasks requiring only common sense and everyday knowledge — such as classifying product reviews to identify potential defects. When tasks require more current or proprietary information, companies can employ retrieval augmented generation (RAG), which pairs a prompt with relevant company documents or data, making it particularly effective for applications like customer service chatbots trained on specific return policies.

For more specialized domains involving technical jargon or complex knowledge — such as medical diagnosis or legal document analysis — instruction fine-tuning is the recommended approach. This method further trains an LLM with application-specific question/answer pairs, modifying the model itself to improve accuracy in niche areas. While more labor-intensive, companies can offset this effort through synthetic data generation, using LLMs to create training data. Ramakrishnan emphasizes that organizations don't need to choose a single method; instead, they should deploy a mix based on specific use cases, starting with prompting as the lowest-effort option and scaling up to fine-tuning when greater accuracy and domain expertise are required.

A study from the MIT Center for Collective Intelligence, published in Nature Human Behaviour, challenges a widely held assumption about AI-human collaboration. After analyzing over 100 studies, researchers found that human-AI combinations on average do not outperform the best human-only or AI-only systems — a finding that surprised even the researchers. The study reveals that combinations work best when each party handles what it does better than the other: humans excel at tasks requiring contextual understanding and emotional intelligence, while AI dominates repetitive, high-volume, or data-driven tasks. Notably, when humans already outperform AI on a given task, adding AI to the mix tends to yield the best collaborative results.

One bright spot for collaboration is generative AI and content creation tasks, where human-AI synergy showed significantly positive effects compared to decision-making tasks. The iterative, interactive nature of generative AI — where humans can draft, edit, and refine outputs in real time — makes it better suited for collaboration than earlier AI systems designed for narrow tasks. Rather than simply dividing subtasks between humans and AI, the researchers recommend redesigning entire workflows and adopting continuous improvement models. MIT Sloan professor Thomas W. Malone emphasized that the findings should not discourage collaboration efforts but instead push organizations toward more sophisticated, evidence-based approaches to deploying human-AI teams.