There is a statistic doing the rounds in boardrooms this year, and it is quietly being used to justify standing still. The line is that around 95% of AI projects fail. A busy, discerning leader hears that and reasonably concludes the safest move is to wait and see.
It is worth knowing where the number comes from before it makes the decision for you. It traces back to MIT's NANDA initiative and its 2025 report on the state of AI in business. The finding is real, but it measures something quite specific: the profit-and-loss impact of mostly large, custom-built enterprise pilots, assessed within roughly six months of launch. It does not capture time saved, quality gained, or the considerable value already flowing through staff who quietly use tools like ChatGPT every day. A narrow measure became a broad headline on the way to the boardroom.
So is AI simply not ready, or not built for smaller firms? The evidence points somewhere more interesting.
The failures are operational, not technical
RAND's 2024 study looked closely at why AI projects fail and found a failure rate above 80%, roughly twice that of other IT projects. The single most common root cause was not weak models or insufficient computing power. It was leadership misunderstanding, or miscommunicating, the problem the project was meant to solve. The technology mostly worked. The framing around it did not.
The pattern among the minority who do see a return is consistent and, frankly, unglamorous. They start from one specific, painful workflow, often in the back office rather than the shop window. Rather than commissioning a bespoke system to be built from scratch, they take mature, well-supported tools and design how those tools are adopted around that single problem; the craft sits in the operational design, not in the engineering. They redesign the process around the tool instead of bolting it on. And they measure something concrete, such as hours saved or turnaround time, before and after.
The upside is well evidenced where it is framed this way. The UK government's SME Digital Adoption Taskforce reported in July 2025 firm-level productivity gains of 7 to 18 per cent per digital technology adopted, and named cost and switching barriers, not the capability of the tools, as the main brake. Honest counter-evidence exists too: some studies report strong returns, such as IDC's $3.70 back for every $1 invested, though those come from vendor-adjacent sources and deserve a degree of caution. Read together, the picture is dispersion rather than doom. Returns are real but unevenly captured, and the difference is mostly in how a business adopts, not which tool it picks.
Why this matters for Scotland's smaller firms
This is quietly good news for Scottish SMEs. The things that cause AI projects to fail, namely fuzzy problem definition, no redesign of the work, nobody owning the change, and measuring the wrong thing, are exactly the things a small, well-run business can control more tightly than a sprawling enterprise can. You do not need a data-science team. You need a clear problem, one owned workflow, and a way to tell whether it actually saved time or money.
This is why I look at AI adoption through operations rather than through the technology. The question that matters is not "which tool" but "where is your time and clarity being lost, and what would it take to win one of those back, properly."
I worked recently with a consultant whose business development had quietly become the bottleneck. It was eating a disproportionate share of his time, it was not being done as well as it could be, and it was also the single area most likely to move profit if it ran properly. Rather than reach for a new tool, we designed a system around the ones he already used: an AI-augmented system embedding his CRM, his call-recording and note-taking tools, his prospecting platform, and email. What came out of it was a largely automated business development process that kept him in the loop at the moments that mattered, overseeing the system rather than working every step by hand. The tools were not new. The operational design around them was.
The failure rate, then, is not a verdict on the technology. It is a verdict on how organisations adopt it, and that is a far more hopeful thing to be told, because it is within reach.
If we want Scotland's smaller firms to be among the most AI-confident in the country rather than the most hesitant, the conversation needs to move away from the scare statistic and the tool of the month, and toward something steadier: one real problem, adopted with discipline, and measured honestly. That is a move almost any well-run business can make.
Sources: MIT NANDA, The State of AI in Business 2025; RAND Corporation (2024) on AI project failure; UK SME Digital Adoption Taskforce (July 2025); IDC (vendor-adjacent ROI figure, treated with caution).