AI projects in operations keep multiplying, and so do the pilots — but few reach production at scale. The causes are almost always the same: insufficient data, processes that are not structured enough, and expectations that do not match what the technology delivers in a real setting. The organisations that get value out of it are not the ones that invested most — they are the ones that started from a precise problem.
Three kinds of use create value in production
The use cases that work in a real operating environment share three characteristics: a clearly defined problem, data that is available and of sufficient quality, and an organisation ready to fold the results into its processes.
- Demand forecasting: significant gains on stock levels are documented in the sector literature — provided the input data is dependable, which usually means a data quality exercise first.
- Document classification and handling: automatic extraction of information from invoices, contracts or forms. This is one of the use cases with the fastest return, because the cost of manual handling is known and measurable. High recognition rates are achievable on structured documents.
- Automation of support and repetitive tasks: request classification, suggested answers, automatic resolution of simple cases. These uses work well when the underlying processes are structured and the historical data is sound.
What AI does not replace: situational judgement, handling of complex exceptions, decisions in undocumented contexts. It increases capacity to process — it does not stand in for business expertise.
Data quality accounts for most of the real effort
The first obstacle identified on AI projects is invariably data quality. A model fed with incomplete or biased data will not produce dependable results, however sophisticated the technology. This preparation work is consistently underestimated in the initial estimates.
The second obstacle is choosing a technology before defining the problem. "Putting AI into our operations" is not a project, it is an intention. An AI project that starts well starts with a precise question: which costly, frequent, measurable problem are we trying to solve? That is the purpose of qualifying the use cases.
A use in production degrades if it is not maintained
Going live is not the end of the project — it is the start of a phase of continuous maintenance. Quality degrades over time if the data shifts without the system being reassessed. Planning the resources, the skills and the update rhythm at scoping avoids discovering this problem eighteen months after launch.
Selecting the right problem matters more than choosing the technology. AI in operations creates value — but only when it is applied in the right place, with the right data, in processes structured enough to benefit from it.
