The Number You Already Know — and the Question It Doesn't Answer
Our previous article showed that 95% of enterprise generative-AI pilots produce no measurable financial impact (MIT, Project NANDA, 2025) — a number contested on methodology, but pointing to a real pattern. The obvious follow-up question is: how does an SME know whether its pilot sits among the 95% or the 5%?
The honest answer, in most cases, is that it doesn't know. Not because it doesn't want to know, but because the decision to continue or kill an AI project rarely rests on disciplined measurement. It rests on impression: "it seems to be helping," "the team likes it," "we haven't noticed problems." That's intuition — now validated by a new tool, not replaced by one.
It isn't an isolated number. RAND Corporation, in a separate investigation with 65 AI practitioners (2024), found that more than 80% of AI projects fail — twice the failure rate of non-AI IT projects. The causes it identifies converge with MIT's: insufficient data quality and poor integration into the real workflow, not the technology itself.
Two to Four Years Is a Long Time to Fly Blind
Part of the problem is the timeline. According to Deloitte, most organisations only report satisfactory ROI on a typical AI use case within 2 to 4 years — only 6% see returns in under a year. The World Economic Forum confirms the same pattern: cost and productivity gains arrive consistently, but rarely in year one.
That collides with the normal decision rhythm of an SME. Most technology investments pay back in 7 to 12 months; waiting 2 to 4 years for a reliable signal isn't realistic for anyone managing cash flow month to month. The result is predictable: without an interim measurement system — indicators that show progress well before financial ROI is visible — most companies decide to continue or abandon an AI project midway through, based on how the project "feels" like it's going.
The Bottleneck Isn't the Model — It's the Data
There's a second, more structural reason intuition keeps winning: even when a company wants to measure properly, it often can't. Data quality and availability is cited by 52% of companies as the number-one barrier to AI adoption (PEX Report 2025/26) — ahead of cost, skills, or resistance to change.
That has a direct consequence for how decisions get made. If the input data isn't reliable, any indicator built on top of it isn't either — and the team ends up, rationally, trusting what it sees with its own eyes more than a number it knows is contaminated. Intuition doesn't come back out of laziness: it comes back because, without clean data, it's literally the most reliable source of information available.
What Changes When You Actually Measure
Our previous article in this series showed that workflow redesign is the strongest differentiator for impact — done by just 21% of organisations (McKinsey). What's missing from that picture is that redesign and measurement aren't independent steps: you can only measure a process rigorously if it was deliberately designed to be measured.
Measuring properly means three things, in the right order. First, define the baseline — how the process worked before, with what numbers — because without a baseline there's no "better" or "worse," only opinion. Second, pick a single, simple indicator anyone on the team can verify, not a dashboard only its author understands. Third, fix the stopping criterion in advance — what would have to happen to conclude it isn't working — because decided midway through, under pressure to justify money already spent, that criterion stops being objective.
How THE iCON Approaches This
THE iCON's work with SMEs on AI adoption includes the measurement layer from day one — not as a final report, but as part of designing the process itself. Before any implementation, we define with the client the current baseline, the single indicator to be tracked, and the stopping criterion, in writing, before there's any pressure to justify it after the fact.
That also means confronting data quality early — if the underlying data isn't reliable, the work starts there, not with the AI tool. A pilot built on bad data doesn't become good by running faster.
Conclusion
Adopting AI is easy. Redesigning a process around it is harder, but more companies are doing it every year. What remains rare is measuring with enough discipline that the call to continue or stop stops being instinct dressed up as data. Without that, AI doesn't replace the judgement of whoever runs the SME — it just gives it a more convincing costume.