Earlier this year, a report from MIT’s NANDA research initiative landed with a thud on LinkedIn. The finding getting most of the attention: despite somewhere between $30 and $40 billion in enterprise investment in generative AI, only around 5% of AI pilot projects are reaching production at scale with measurable business impact.
LinkedIn responded predictably. Pages of profound observations followed, most of them – I’ll be diplomatic here – somewhat repetitive and suspiciously well-formatted.
I don’t want to add to that pile. But the underlying finding is worth talking about properly, because it confirms something a lot of practitioners have been sensing for a while.
We’re now more than halfway through the financial year. If your organisation committed to doing something meaningful with AI tools in FY2025-26, November is a reasonable moment to ask honestly how that’s going.
What the MIT report actually found
The report – The GenAI Divide: State of AI in Business 2025 – is worth reading rather than relying on secondhand summaries. It’s based on a review of over 300 publicly disclosed AI initiatives, 52 structured interviews and surveys of 153 senior leaders, conducted between January and June 2025. The sample is modest, and the authors acknowledge limitations – so treat it as a useful signal, not a definitive verdict.
The headline figure (5% of pilots reaching production with real P&L impact) is striking, but the more interesting finding is why. The researchers didn’t find that organisations were using the wrong AI models, or that infrastructure was the problem, or even that talent was the bottleneck.
The core issue they identified was a learning gap. Most AI tools deployed in organisations don’t retain context between sessions, don’t adapt to how a specific business actually operates, and don’t improve over time based on feedback. They’re static. The report summarises it bluntly: the tools that fail are those that don’t integrate into workflows and can’t evolve.
There’s a related finding that I think gets less attention than it deserves. The report noted that mid-market companies – not large enterprises – moved faster and more successfully from pilot to implementation. Large enterprises had more pilots, more budget and more staff assigned to AI. They also had the lowest rates of converting those pilots into something that actually worked. Bigger investment did not mean better outcomes.
Sound familiar?
I use ChatGPT, Copilot and Gemini and find them genuinely useful – for drafting, for working through a problem, for getting a first pass at something I’d otherwise spend an hour on. I’d be reluctant to give them up.
But I’m also talking to SME and NFP leaders who subscribed to Copilot because Microsoft included it, or signed up for a ChatGPT team plan because someone on staff pushed for it, and six months later aren’t sure what they’re actually getting from it. The tools are being used. Whether they’re delivering meaningful business value is a different question.
I’ve seen this pattern before, just with different technology. In the late 1990s it was ERP systems, and in the 2010s it was cloud migration. The specific technology changes but the dynamic doesn’t – organisations buy first and figure out the problem later. Sometimes they find a good fit. Often the subscription just quietly renews.
The question worth asking mid-year
Before you renew, expand or add AI tools before June 30, I’d suggest asking three questions.
What specific problem were we trying to solve when we signed up for this? If the answer is “we wanted to keep up with AI” or “the vendor demo was compelling” or “everyone else was signing up for it”, that’s worth noting. Tools adopted without a clear problem to solve tend to get used inconsistently and measured loosely. Not all of them are wasted – people find uses for things – but you won’t know whether you’re getting value without a clearer baseline than “it seemed like a good idea.”
Are people actually using it, and for what? The MIT report flagged a “shadow AI economy” – workers using personal AI tools more effectively than the official enterprise solutions their organisation deployed. That gap, when it exists, is diagnostic. It tells you something about whether the tools you’ve chosen actually fit how your people work, or whether they were selected for other reasons.
Is the tool integrated into how we actually work, or is it an add-on? An AI tool that lives alongside your workflows – requiring people to switch context, re-enter information, and remember to use it – delivers much less value than one embedded in how work actually gets done. If your team is using AI tools as a parallel step rather than as part of the process, you’ll get individual productivity gains at best. Business-wide impact requires integration.
If you’d like a structured look at what your current AI tools are actually delivering – or a clear-eyed assessment of where AI could genuinely help your organisation – get in touch. That conversation tends to be more useful than any demo.
