Six data science problems that stall companies before any model is built

AI adoption is rising fast: 20% of EU enterprises used AI in 2025 ^[1], and in McKinsey's 2025 survey 88% of respondents said their organizations regularly use AI in at least one business function ^[2]. Yet nearly two-thirds have not begun scaling AI across the enterprise ^[2]. In my experience the reasons are rarely about algorithms. They are the six problems below.

1. There is no decision to improve

A project that starts with "let's do something with our data" has nothing to measure itself against. A useful project starts with a decision that is made again and again, such as how much to order, which customers to call or which report management reads on Monday, and asks how data can make that decision faster or better.

2. The data isn't ready

Gartner predicts that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data ^[3]. "Ready" means more than "exists": the data has to be complete, consistent and recorded the same way over time. A quick check of missing values, duplicates and mixed formats tells you more than any model will. You can run one on your own file with the free Data Quality Check; the file never leaves your browser.

3. Every department has its own numbers

When sales, finance and operations each pull figures from their own exports, the same KPI comes out differently in every meeting. Before any model, the company needs a single source of truth: one agreed definition per metric and one place it is calculated ^[4]. This is unglamorous work, and it is where reporting automation usually begins.

4. The success metric is wrong

Forecast accuracy is often reported as MAPE, the mean absolute percentage error ^[5]:

The point is broader than forecasting: pick the metric with the business, check how it behaves on your data, and write it down before the work starts.

5. Averages hide the real story

Simpson's paradox is a trend that appears in several groups of data but disappears or reverses when the groups are combined ^[6]. The classic example is UC Berkeley's 1973 admissions: overall, men were admitted at a higher rate, but department by department the picture changed, because women applied more often to the most competitive departments ^[6].

Store A converts better in both segments, yet looks far worse in the combined figure, because most of its visitors come on weekdays, when everyone converts less. Dashboards that only show totals invite exactly this mistake; good ones let you break every number down.

6. Nobody owns the system afterwards

Among EU enterprises that considered AI but don't use it, lack of relevant expertise is the most cited barrier ^[7]. The same gap shows up after a project: a model nobody understands, retrains or monitors quietly decays. Plan the handover from day one, whether that means training your team, documenting the system, or having a fractional data science lead keep it on course.

A quick self-check

1. Can you name the decision this project improves?
2. Have you checked missing values and duplicates in the data it needs?
3. Does everyone agree on how the key numbers are defined?
4. Is the success metric written down, and does it suit your data?
5. Can every total on your dashboard be broken down by segment?
6. Is it clear who runs the system once it's live?

Want a structured version of this check? The free Data Maturity Test scores your company in a few minutes, entirely in your browser.