New insights from MIT Sloan Management Review break down the main types of artificial intelligence startups and how companies can benefit from working with them. Provides tips for using agent AI coding tools for knowledge work. And we explore why artificial intelligence is not ready to contribute to large-scale productivity improvements.
Learn about 6 types of AI startups
Venture capital firms are pushing for startups in their portfolios to incorporate generative AI into their product offerings and internal operations. There’s a good reason for this, according to MIT Digital Economy Initiative professor Thomas Davenport and Babson College professor Jeffrey Shea. Startups face no organizational or technological barriers to implementing enterprise AI.
However, the challenge is knowing when, where, and how to use AI and understanding who is developing the AI tools.
Davenport and Shay categorize AI startups into six types to provide clarity for prospective adopters.
- originator Build a basic model that will be commercially deployed.
- explorer Look out for future use cases like agent AI and quantum AI.
- infrastructure builder Create the data, application programming interfaces, and frameworks needed to use generative AI.
- enhancer Apply generic AI models to specific industries and problems.
- optimizer Transform the way you work with AI.
- experimenter Test AI without a budget, roadmap, or integrations. This is by far the largest group of companies.
Potential business customers want to ensure they are working with a company that matches their needs and expectations. For example, organizations with existing AI capabilities can turn to infrastructure builders to accelerate development, while organizations looking for quick wins on specific capabilities should seek out enhancers as partners.
Read “Explaining the 6 types of AI startups”
Don’t think AI will make you more productive
MIT Institute professor and Nobel laureate Daron Acemoglu argued in a podcast interview that the effects of technology throughout history have not been “predetermined,” expanding on ideas laid out in his 2023 book “Power and Progress.” “We have a lot of agency and a lot of choices in shaping the future of technology, and different futures correspond to different winners and losers,” he said.
AI is no exception. It is versatile enough to offer different versions of the future, each with different outcomes for the economy, industry, and workers, depending on the application.
But today, Acemoglu argued, organizations are choosing to use AI to: automation When the technology is actually in the form of information technology. This explains why AI does not improve productivity at a macroeconomic level. Automation benefits capital owners. Workers, on the other hand, will benefit from the ability to gain insights and make decisions from the information that AI systems aggregate.
It doesn’t help that existing architectures and economic models make it difficult to build AI tools in this way. In theory, electricians could use AI tools to understand why equipment isn’t working, or nurses could use AI tools to make recommendations for patient treatment. However, for that to happen, AI tools need to be task-optimized, reliable, and trained on domain-specific information. Achieving all of this takes time and capital.
What needs to change? Acemoglu believes organizations need to determine where they can address bottlenecks without overdoing it through automation. He proposed creating “pro-worker, pro-human technology” that would decentralize access to information and support front-line decision makers. This would allow the benefits of “technological improvements” to be shared more equitably, as was the case after World War II, when innovation led to widespread wage increases.
However, many of today’s innovators, including startups, choose to partner with capital owners in pursuit of acquisitions and wealth creation. “Maybe we should be more cautious about mergers and acquisitions,” Acemoglu said. “That can cause very different dynamics.”
Hear “AI isn’t improving productivity”
Lead an AI-driven organization
Face-to-face at MIT Sloan
Register now
Use agent AI coding tools for knowledge work
You might think that agentic AI coding tools are only for software developers, but MIT Sloan Professor Rama Ramakrishnan writes that tools like Claude Code can also be used for knowledge work. Three technical components of agent AI coding tools help explain why.
- multi-level reasoning Break down your tasks into a series of steps and create an action plan to accomplish them.
- adaptive execution Observe the results of each step and correct course if something isn’t right.
- Using tools It involves processes run by external systems, such as connecting to a database or running commands.
These technical components demonstrate that agent AI coding tools can not only read files, but also remember their contents, rerun the analysis when new source material is available, and perform multiple tasks at once. As a result, the tool is suitable for tasks such as conducting competitive intelligence, preparing for meetings, and creating different versions of marketing campaigns, writes Ramakrishnan. (Read this article for detailed guidance on creating and adjusting prompts.)
Ramakrishnan has a few words of caution for executives before they get to work. One is to ensure that coding tools only have access to trusted folders and files. Excluding documents from external sources reduces a security risk known as prompt injection, which can manipulate AI tools into misbehaving. Additionally, agentic AI coding tools are not 100% accurate and mistakes can occur, so users should review the tool’s suggested actions before approving them.
Read “AI Coding Tools for Knowledge Work: What Business Executives Need to Know”
This article is based on the following insights: MIT Sloan Management Reviewleading the discussion on advances in management practice among influential thought leaders in business and academia. This publication provides readers with evidence-based insights and guidance for innovating, operating, leading, and creating value in a world transformed by technology and large-scale social and environmental forces.
