The GenAI Divide: State of AI in Business 2025 (AI column)

I read something interesting yesterday. report About the state of affairs in the implementation of AI in the industry from the project team MIT NANDA Aditya Challapally, Chris Pease, Ramesh Raskar and Pradyumna Chari (July 2025). Project NANDA stands for Networked Agents and Decentralized Architecture and is an MIT research initiative aimed at building the infrastructure for distributed intelligence of autonomous agent systems. The study was conducted from January to June 2025 Preliminary report on the implementation of GenAI (generative AI) in business.

The report itself is 26 The authors of the report show a gap between widespread interest in generative AI and the real business effect of its implementation. The authors even came up with the beautiful name “GenAI Divide” to describe this effect, which in numbers looks like this.

1. The gap in results Despite global investment of $30–40 billion, ~95Percentage of organizations did not receive measurable returns from GenAI projects, and the remaining 5% were able to make money from it. The authors explain this not by the quality of models or regulation, but by different approaches to implementing technologies.. This is encouraging because it can be corrected. 2. High experimentation, low ROI AI tools (ChatGPT, GitHub Copilot, ...) They are widely distributed - more 80Most companies have experimented with them, and 40% have already deployed such solutions for employees. However, these tools have largely increased individual productivity and have had little impact on the organization’s key financial performance. In other words, generative AI is actively being tried out in workflows, but has not yet transformed business on a large scale. 3. Failures of Corporate Pilots This is how corporate projects look like. 60% -> 20% -> 5% where 60% evaluated corpus GenAI solutions (customist), 20% reached the pilot, and 5Percent went to sale. As a result, most pilots remain in place without scaling. The main reasons for this are maladaptive. ("fragile") workflows, lack of ability to take feedback and context into account, and poor integration of the solution into the day-to-day operations of the company. 4. The phenomenon of GenAI Divide The study identified four characteristic patterns separating successful and unsuccessful organizations. (GenAI Divide): Limited industry impact - only 2 from 8 Sectors are experiencing structural changes from AI, or rather it is "technology" and "media & telecom" Large corporations launch many pilots, but lag behind in scaling solutions Budgets are mostly spent on visible front-office functions (sales) at the expense of back-office processes that often have better returns on investment Projects with external technology partners are twice as successful as internal projects.

The main barrier to scaling GenAI systems has been the inability to learn: most solutions do not retain context, adapt and improve as they are used. However, a small group of leaders managed to bridge this gap with a different approach – from the outset they put on customization for processes and close integration of AI tools into work, and efficiency was measured by business results rather than technical model metrics.

The authors conducted it in the first half of the year 2025 Over the years, several data sources were used.

1. The team carried out a systematic review 300+ publicly announced AI initiatives to identify overall trends and project outcomes
2. The team conducted 52 Structured interviews with representatives of various organizations implementing GenAI (quality insights from the fields)
3. The team analyzed the questionnaires 153 senior managers (C-level et al.), gathered at four major industry conferences. Such a multifaceted design (case analysis, interviews and surveys) It allowed us to look at the issues of using GenAI from different angles.

#AI #Engineering #Software #Metrics #Leadership #ML #Metrics