Published on the 04/09/2025 | Written by Heather Wright
Behind the AI headlines lies a murky fact…
Despite the massive enterprise investment in genAI, 95 percent of enterprises are getting zero return, at least at the company financials level, with just one in 20 pilots delivering significant value.
That’s according to research from MIT’s Project Nanda (networked agents and decentralised AI) which says while more than 80 percent of organisations have explored or piloted genAI tools – and nearly 40 percent are reporting they’ve deployed them – the tools are being used primarily to enhance individual productivity, rather than tangible financial performance of the company, as measured in profit and loss.
“While experimentation is good, without a connection to the true business opportunity experiments inevitably fall short.”
At the same time The GenAI Divide: State of AI in Business says, enterprise-grade systems, both custom and vendor-sold, are getting short shrift, with 60 percent of organisations evaluating the tools, but only 20 percent reaching pilot stage and just five percent making it into production.
“Most fail due to brittle workflows, lack of contextual learning, and misalignment with day-to-day operations.”
To be clear, the 95 percent stat appears takes a very narrow view, looking at impact on P&L, and the five percent of integrated AI pilots apparently deemed successful are classed as ‘extracting millions in value’. The others, according to the report, saw little measurable impact on their profit and loss figures, or stalled in pilot phase.
The report says it looked at 300 publicly disclosed AI initiatives along with interviews from 52 organisations and survey responses from 153 senior leaders – though no further information is provided about those respondents or organisations.
It says for most, adoption is high, but disruption is low, with custom solutions stalling due to integration complexity and lack of fit with existing workflows. But it also suggests a thriving shadow AI economy. While just 40 percent of companies say they’ve purchased an official LLM subscription, employees from over 90 percent of the companies surveyed reported regular use of personal AI tools for work tasks.
“Behind the disappointing enterprise numbers lies a surprising reality: AI is already transforming work, just not through official channels.”
That shadow AI, is often returning better ROI than formal initiatives and highlighting the need for flexible, responsive tools.
“The organisations that recognise this pattern and build on it represent the future of enterprise AI adoption.
“Forward-thinking organisations are beginning to bridge this gap by learning from shadow usage and analysing which personal tools deliver value before procuring enterprise alternatives.”
The report comes hard on the heels of the underwhelming release of OpenAI’s GPT-5, which saw the company quickly bring back its older model for paying customers, in response to users frustration over the new iteration.
And despite the MIT pedigree, it has raised eyebrows. Kevin Werbach, Wharton professor and chair of legal studies and business ethics, is among those questioning it. He calls it ‘a great example of confirmation bias’, and, like others, has questioned where the 95 percent figure is drawn from – it’s referred to just twice in the report, but without any support for the claim.
“The fact that this report fails to demonstrate that most generative AI deployments fail, does not, of course, mean they are successful,” Werbach says.
“There are good reasons to wonder whether generative AI is creating returns to justify the massive level of investment. But it’s not going to be a black and white matter… Reality is inevitably messier, especially given the short time period since generative AI was introduced widely in enterprises.”
The MIT Nanda report also says while 50 percent of genAI budgets is going into sales and marketing, it is back-office automation that often yields better ROI.
“This bias reflects easier metric attribution, not actual value, and keeps organisations focused on the wrong priorities.”
While metrics such as demo volume or email response times can be measured easily and align directly with KPIs, use in legal, procurement and finance functions offers more subtle efficiencies such as fewer compliance violations, streamlined workflows or accelerated month-end processes – all valuable, but harder to quantify and justify in executive conversations or investor updates.
“The highest performing organisations report measureable savings from reduced BPO (business process outsourcing) spending and external agency use, particularly in back-office operations. Others cite improved customer retention and sales conversion through automated outreach and intelligent follow-up systems.”
Most implementations didn’t drive headcount reduction.
The report says the results suggest that learning-capable systems, when targeted at specific processes, can deliver real value, even without major organisational restructuring.
The report argues too, that the tools currently available lack the ability to adapt, remember and evolve for mission-critical work.
INSEAD professors Nathan Furr (professor of strategy) and Andrew Shipilov (professor of international management) say the report suggests many leaders are making the same mistake they made with digital transformation a decade ago – letting experimentation run wild in the hope of finding the ‘unicorn’ returns somewhere in the mess of experimentation.
It’s an approach the pair believe is doomed to failure, producing ‘a morass of unfocused, under-resourced teams that produce few scalable results’.
“While experimentation is good, without a connection to the true business opportunity – eg, transforming the core to serve existing and new customers – experiments inevitably fall short of hopes and expectations.
“It sounds obvious, but by framing AI as radical and disruptive we often lose sight of the connection to the most fundamental objective of business: to solve problems for customers.”
Instead, they say, businesses need to understand the AI moment in the larger arc of transformation, focus on AI’s potential to better serve customers, experiment with a focused set of opportunities to prove them, and then scale them up.
They argue that we are entering a world of data- and AI-driven decisions where humans provide oversight, but are not necessarily the core engine of the activity.
“This is a truly profound shift, and we’re only in the middle of it – in the end, it will take many years but it will lead to a fundamentally different kind of organisation.”
Understanding the bigger picture enables leaders to keep in focus that the point is to transform the business to use technology to better serve customers, faster, more easily and cheaply.
“All forms of AI (including genAI) are just a tool – one of many – that can help accomplish that. Just as the internet fundamentally changed how customers are served, but not why they’re served, AI adoption needs to be seen through this laser-like focus to succeed,” they say.
The pair say the real opportunity – one that they say will actually generate returns – is to look at the internal operations and the external customer journey and start with how you can create real value, in the near term, using AI tools, and focus experiments there.
Experiments should be connected to true value creations, as low cost as possible so you can have multiple cycles to learn and improve and be designed with a connection to how they can eventually scale to create value.
“As we enter the post-enthusiasm wave of AI, many leaders are in danger of misinterpreting the challenges implementing AI as a signal that AI cannot be used to create value,” the pair say.
“The truth is AI can create value and we are seeing significant process… but creating value always comes back to the initial moment of experiment design, when a team is able to see how a new tool can create value for customers, because no matter what new tools come in the future, the purpose of business will always remain the same: To solve important problems for customers.”
