Dr Ovidiu Tierean is a Senior Advisor at PKF Malta
In boardrooms from Valletta to Vienna, the buzzwords arrive faster than the balance sheets can keep up. Generative AI was to be the locomotive of a new industrial revolution. Consultants painted visions of leaner operations, frictionless workflows and revenue curves bending upwards. But as the latest State of AI in Business 2025 from MIT makes abundantly clear, the train has left the station carrying more AI Models than paying passengers.
The report’s blunt arithmetic is difficult to ignore. 30-40 billion dollars have been poured into GenAI projects. Yet for all that expense, only 5% of these systems have crossed the finish line into production. The rest report no measurable shift in their profit statements. It’s a remarkable paradox: adoption rates are high, but genuine transformation remains as rare as rain in August.
Authors distil the results into two patterns. First, AI models have quietly become everyday companions for millions of workers. They shave minutes from tasks, summarise documents, draft courteous replies. But these micro‑efficiencies have not coagulated into enterprise‑level change. Second, the heavyweight, bespoke “enterprise‑grade” platforms that were meant to deliver structural gains remain largely stuck in pilot-project purgatory.
The sectoral picture is just as sobering. Only Technology and Media show signs of true disruption. Healthcare, energy, finance, and advanced industries are still circling the runway with proofs‑of‑concept trials. The problem is not hostile regulators or weak models. It is the more prosaic matter of fit: most tools simply fail to embed themselves in real workflows.
Against this backdrop, a second, unofficial economy has been thriving in plain sight. The study calls it the “shadow AI economy”. Here, over 90% of employees admit to using personal AI tools on the job, while only 40% of employers have sanctioned licences. It is grassroots, bottom‑up adoption where workers tap into consumer‑grade accounts for ChatGPT, Claude and others, tailoring them to their own rhythms. Freed from corporate integration cycles and rigid interfaces, these tools feel nimble, responsive, and above all useful.
By contrast, the official deployments – costly, over‑engineered and static – inspire little loyalty. They lack memory, cannot adapt over time, and so are swiftly abandoned for sensitive or complex tasks. The effect is self‑reinforcing: the more comfortable staff become with their clandestine companions, the less patience they have for corporate systems that lag behind.
Successful adopters share a recognisable playbook. They customise deeply for their own processes, measure outcomes rather than technical benchmarks, and put power users, so-called “prosumers”, in the driver’s seat of adoption. Significantly, projects built with external partners are twice as likely to make it to production as purely internal builds. This is not an argument against cultivating in‑house skills; it is a recognition that speed‑to‑value and trust can outweigh ideological purity.
Technically, the path forward is also coming into focus. The report makes a strong case for “agentic” systems — platforms that retain memory, learn from interaction, and improve over time.
There is also an emerging cultural battle, the so‑called “war for simple work”. Here, AI has already claimed decisive victories. Seven in ten employees prefer it for drafting emails; nearly two‑thirds for basic analysis. Yet when stakes rise, trust still tilts decisively toward human judgment, with 90 per cent favouring people for mission‑critical decisions. This reflects both the limits of current tools and the enduring premium placed on accountability.
Perhaps the report’s most bracing service is myth‑busting. Generative AI is not replacing vast numbers of jobs, nor has it rewritten the DNA of how business is done. Most firms have already invested heavily; their obstacles are less about model performance or regulatory drag, and more about the inability of current tools to adapt to, and improve within, the messy realities of enterprise workflows. Internal builds, the data shows, fail at double the rate of externally sourced solutions.
For leaders, the prescription is unapologetically pragmatic. Stop treating GenAI as a showroom exhibit. Weave it into operations where memory matters and where external spend is most burdensome. Be willing to buy before you build if that shortens the learning curve in your environment. Demand proof on your data, against your KPIs, that systems improve over time. Above all, enlist the power users in your own ranks, the ones already living in the shadow AI world, and make them champions of the formal rollout, supported by real change‑management muscle.
The “GenAI divide” is more than a catchy phrase. It captures the gulf between what is officially commissioned and what is actually delivering value on the ground. MIT’s NANDA report is a clear‑eyed reminder that this gap can be closed, but only if leaders abandon the theatre of adoption and focus instead on the economics of transformation. As with so many technological promises before it, the winners will be those who master not just the art of the possible, but the discipline of the profitable.
