Every second enquiry we receive now mentions AI. That is a healthy sign — but the conversation usually starts in the wrong place. The question is rarely “which model should we use”. It is almost always “is our data in a state where any model could help”.
Start with the boring question
Before recommending anything, we ask a client to show us where a single number lives. Yesterday’s output figure for one line, say. In a surprising number of mid-sized manufacturers that number exists in three places — a machine log, a supervisor’s register and an ERP entry made the following morning — and the three do not agree.
No amount of machine learning fixes that. A model trained on inconsistent data produces confident, inconsistent answers. The unglamorous work of reconciling those three sources is what makes everything afterwards possible, and it is usually where the real return sits.
Where AI pays for itself first
In our experience the earliest wins are narrow and unromantic:
- Demand forecasting where seasonality is strong and historical sales data is already clean.
- Quality inspection using vision models, where the defect is visually obvious and you have photographs of past rejections.
- Document processing — invoices, purchase orders, dispatch notes — where staff currently retype information from a PDF into a system.
- Predictive maintenance, but only where sensors are already fitted and logging reliably.
What these share is a clear before-and-after measurement. You know what the forecast error was last quarter. You know how many hours go into invoice entry. That makes the result arguable in a board meeting, which matters more than technical elegance.
Where projects stall
The pattern is consistent. A pilot works well on a curated dataset, then meets production reality: a camera gets moved, a supplier changes their invoice layout, a sensor fails quietly for three weeks. The model keeps producing output, and nobody notices it has become wrong.
The fix is not more sophisticated modelling. It is monitoring, clear ownership, and an honest conversation at the outset about who maintains the system in month eighteen. We would rather scope a smaller project that survives than a larger one that quietly degrades.
A reasonable first step
Pick one process, measure it properly for a month, and only then decide whether AI is the right instrument. Often the measurement alone surfaces a process fix worth more than the technology. That is not a disappointing outcome — it is a cheap one.
