• by Laurie Sullivan , Staff Writer, Yesterday

Generating higher returns on investments from generative artificial intelligence (GAI) may be a little more difficult to achieve than some executives initially thought, although many companies are throwing millions and sometimes billions of dollars into AI projects.

Researchers with MIT's NANDA project identified that many organizations have not seen measurable return on their AI initiatives, despite spending between $30 billion and $40 billion in enterprise GAI investments.

The July 2025 findings highlight in the "the GenAI Divide" that only 5% of organizations gain significant value from AI, while 95% see no measurable profit or loss impact.

According to MIT's Project NANDA, most enterprise AI investments yield zero returns due to core learning limitations in today’s systems.

Like the way early web protocols DNS and HTTP were required to scale, NANDA is creating the essential building blocks to support trillions of autonomous AI agents that can share information across companies to complete tasks like buying goods and services, and complete shipping requirements for those purchases.

MIT’s project does take into consideration safeguards for those companies not wanting to share information. The project calls it a core feature of the architecture, where security and privacy mechanisms allow companies to benefit from agent collaboration while keeping their sensitive data secure. 

Data shows more than 80% of organizations evaluated enterprise-grade system tools like OpenAI ChatGPT and Microsoft Copilot, and nearly 40% deployed them, but only 20% reached pilot stage and just 5% reached production. 

The research conducted from January to June 2025 revealed several patterns defining the GenAI Divide. Most fail due to delicate workflows, lack of contextual learning, and misalignment with day-to-day operations.

But the core barrier to scaling is not infrastructure, regulation, or talent, but rather learning. Most GAI systems do not, yet, retain much feedback for long periods of time, adapt to context, or continue to improve and retain this type of learning. This is changing. 

Google DeepMinds and Meta Platforms continue to train their respective AI models to improve cognitive abilities, which is part of a vision for personal superintelligence.

The research also showed divided investment patterns. Sales and marketing functions hold approximately 70% of AI budgets across organizations, yet back-office automation often yields better return on this type of investment.

Some 50% of AI budgets get swallowed by sales and marketing, but many cost savings the study documented came from back-office automation.

Business Process Outsourcing, the practice of a company hiring a third-party service provider to handle non-core business functions or operations, eliminates extra costs and save the companies between $2 million and $10 million annually in customer service and document processing, agency spend reduction of 30% decrease in external creative and content costs.

The report identified agentic AI protocols such as Model Context Protocol (MCP), Agent-to-Agent (A2A), and NANDA. All enable agent interoperability and coordination.

The ultimate goal for the web, per the study, a complete agentic ecosystem where autonomous systems can discover, negotiate, and coordinate across the entire internet infrastructure, fundamentally changing how advertising, ecommerce, and business processes operate.