I was intrigued by a recent MIT study that showed 95% of the companies that have invested in GenAI are getting zero return. This is a report that rattled the stock market, where investors started having second thoughts over whether the billions ploughed into GenAI are going to pay off.

But is it a shortcoming in the technology per se? 

No. 

What’s really holding GenAI back, according to research based on a review of more than 300 AI initiatives and survey responses from 153 senior leaders between January and June 2025, is this: Most enterprise AI applications don’t learn and don’t integrate well into workflows; we deploy them in the wrong areas; and our employees expect much more.

Some insights from the report that jumped out at me:

• Buying is twice as successful as building internally. 

• Tools like ChatGPT and Copilot are widely adopted, but these tools primarily enhance individual productivity, not P&L performance. Meanwhile, enterprise-grade systems are being quietly rejected. “Most fail due to brittle workflows, lack of contextual learning, and misalignment with day-to-day operations.”

• By contrast: The organisations and vendors succeeding are those aggressively solving for learning, memory, and workflow adaptation, while those failing are either building generic tools or trying to develop capabilities internally. “Winning startups build systems that learn from feedback (66% of executives want this), retain context (63% demand this), and customise deeply to specific workflows. They start at workflow edges with significant customisation, then scale into core processes.”

• Big firms lead in pilot volume but lag in scale-up. Humans are a large part of the reason. Take a look at this chart.

• Budgets favour visible, top-line functions over high-ROI back office functions. While less transformative front-office gains are visible and board-friendly, the back-office deployments often delivered faster payback periods and clearer cost reductions. For example, about half of GenAI budgets go to sales and marketing. This bias reflects easier metric attribution, not actual value, and keeps organisations focused on the wrong priorities. Sales and marketing dominate not only because of visibility but because outcomes can be measured easily. Metrics such as demo volume or e-mail response time align directly with board-level KPIs. Legal, procurement, and finance functions, in contrast, offer more subtle efficiencies. These include fewer compliance violations, streamlined workflows, or accelerated month-end processes — important but hard to surface in executive conversations or investor updates. 

• Most implementations don’t drive headcount reduction.

• For mission-critical work, 90% of users prefer humans because GenAI lacks memory and adaptability.  

• Tools that succeeded shared two traits: low configuration burden and immediate, visible value. Tools that struggled, on the other hand, involved “complex internal logic, opaque decision support, or optimisation based on proprietary heuristics.”

So, what does this mean for us in the media industry?

• We need to think through change management so users’ expectations are aligned with the capabilities of the tools that we are deploying.

• To move the needle for our businesses, we have to think through where we are using AI:

• How many of us are trying to apply AI in places such as newsrooms with “complex internal logic, opaque decision support, or optimisation based on proprietary heuristics”? This is not a recipe for success. What can we simplify to make this work better?

• Many applications that I have seen in the news business are aimed at incremental efficiencies — yet, the real gains appear to be from products that drive revenue, growth and strategic differentiation.

• Narrow, specific use cases are more likely to provide value than more general applications, e.g., a broad, company-wide rollout of Copilot.

• The technology will only get better. Agentic AI, which embeds persistent memory and iterative learning by design, directly addresses the problems the study mentioned and is already a reality. As AI evolves, we will use it more, not less.  

If you’d like to subscribe to my bi-weekly newsletter, INMA members can do so here.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .