The artificial intelligence revolution was supposed to deliver massive productivity gains. Instead, organizations are scratching their heads, wondering why their $200 billion AI investment isn’t moving the needle on bottom-line performance. New research from MIT Sloan Management Review cuts straight to the heart of this paradox: companies are fundamentally misunderstanding what AI is and how to deploy it.
The problem isn’t the technology—it’s the mindset. Organizations are treating AI as automation technology when it’s actually information technology. This critical misclassification explains why we’re not seeing the productivity revolution that historical precedent suggests should be happening by now.
The Six Types of AI Startups: Know Your Partner
Before diving into the productivity problem, let’s establish the landscape. MIT researchers Thomas Davenport and Jeffrey Shay have identified six distinct categories of AI startups, each serving different organizational needs:
- Originators: Build foundational models (think OpenAI, Anthropic)
- Explorers: Push boundaries with agentic and quantum AI applications
- Infrastructure builders: Create the data pipelines and frameworks for AI deployment
- Enhancers: Apply general AI to specific industries or problems
- Optimizers: Use AI to transform internal operations
- Experimenters: Test AI without clear budgets or roadmaps (the largest group)
The Twitter discourse reflects this startup proliferation perfectly:
“90% of ‘AI startups’ are just: Take user input, Send to OpenAI API, Display response, Charge $29/month” — @trikcode
This brutal assessment highlights why choosing the right AI partner matters. Organizations with existing capabilities should target infrastructure builders to accelerate development. Those seeking quick wins need enhancers that understand their specific domain challenges.
The Automation Trap: Why AI Productivity Gains Are Invisible
Nobel laureate Daron Acemoglu delivers a sobering reality check: AI isn’t improving productivity at a macroeconomic level because we’re using it wrong. The core issue? Organizations default to automation applications that primarily benefit capital owners, not workers or overall productivity.
This mirrors historical technology adoption patterns. The Industrial Revolution initially created massive wealth inequality before productivity gains eventually spread throughout the economy. The difference? Those technologies were genuinely automating physical processes. AI’s strength lies in information processing and decision support—capabilities we’re systematically underutilizing.
Consider the missed opportunities: Electricians could use AI to diagnose equipment failures. Nurses could leverage it for treatment recommendations. Instead, we’re building chatbots and automating customer service—applications that reduce labor costs without improving overall system performance.
The Information Technology Revolution We’re Ignoring
The productivity puzzle becomes clearer when we examine AI through the information technology lens. Unlike automation that replaces human tasks, information technology should amplify human capabilities. The post-World War II boom demonstrated this principle: technological advancement that augmented worker capabilities led to widespread wage growth and productivity gains.
Current AI architectures and economic models make it difficult to build pro-worker, pro-human technologies. Creating domain-specific, reliable AI tools requires significant time and capital investment. Most organizations choose the path of least resistance: automation that delivers immediate cost savings rather than capability enhancement.
Acemoglu’s solution involves creating technologies that decentralize access to information and support front-line decision makers. This approach would distribute the gains from technological advancement more equitably, similar to the broad-based prosperity following major historical innovations.
Agentic AI: Beyond Coding Into Knowledge Work
MIT’s Rama Ramakrishnan reveals an overlooked opportunity: agentic AI coding tools aren’t just for developers. These tools possess three critical capabilities that make them powerful for knowledge work:
- Multistep reasoning: Breaking complex tasks into executable sequences
- Adaptive execution: Course-correcting based on intermediate results
- Tool integration: Connecting with external systems and databases
These capabilities enable applications like competitive intelligence gathering, meeting preparation, and campaign development. The tools can remember file contents, re-execute analyses with new data, and run multiple tasks simultaneously—exactly the kind of information processing amplification that should drive productivity gains.
The market is taking notice. Real estate prices in San Francisco reflect the wealth creation happening in AI:
“San Francisco’s median house price jumped to a record $2.15 million in March, up 18% from a year earlier as wealth generated by AI startups flooded the city, according to brokerage Compass” — @business
The Path Forward: Strategic AI Implementation
The solution requires fundamental strategic shifts. Organizations must identify where automation addresses genuine bottlenecks without overdoing it. More importantly, they need to invest in AI applications that enhance worker capabilities rather than replace them.
This means building domain-specific tools trained on relevant data, optimized for specific tasks, and designed for reliability. It requires patience with longer development cycles and higher upfront costs in exchange for sustainable productivity improvements.
Regulatory intervention may be necessary. Acemoglu suggests increased scrutiny of mergers and acquisitions to prevent AI innovation from concentrating benefits among capital owners. The goal: ensuring that AI’s productivity gains distribute broadly rather than accumulating in narrow segments.
Conclusion: Information Technology, Not Automation
The AI productivity paradox isn’t a technology failure—it’s a strategic failure. Organizations treating AI as automation technology will continue seeing limited returns. Those that embrace AI as information technology, building tools that amplify human capabilities and decentralize access to insights, will capture the genuine productivity revolution.
The choice isn’t predetermined. As Acemoglu emphasizes, we have agency in shaping technology’s future. The question is whether organizations will continue chasing quick automation wins or invest in the harder work of building productivity-enhancing information systems. The macroeconomic data suggests the automation approach isn’t working. It’s time to try the alternative.