Artificial intelligence in mid-sized companies: from pilot to productivity

Few technologies have reached the agenda of mid-sized companies’ management teams as quickly as generative artificial intelligence. Somewhere between the two extremes – ‘we must automate everything immediately’ and ‘we’ll wait until it blows over’ – it is being decided right now who will have realised productivity gains five years from now and who will not.

Experience from technology projects shows that successful AI adoption in mid-sized companies rarely starts with the technology. It starts with a sober review of processes – focusing on where repetitive text work, classification, forecasting or document review create bottlenecks today.

Use cases with demonstrable value

Across industries, a set of applications is emerging in which AI already delivers reliable value:

• Knowledge work: summarising, structuring and drafting documents – from proposal templates to inspection reports.
• Customer service: pre-qualifying enquiries and drafting responses for staff to review and approve.
• Quality and maintenance: detecting anomalies in sensor and process data before failures occur.
• Administration: extracting structured data from invoices, delivery notes and contracts.

The real bottlenecks: data and people

Technically, these use cases are now well within reach. The bottlenecks lie elsewhere: in scattered, unstructured data, in unclear responsibilities – and, above all, in a lack of experience. AI expertise is scarce and expensive on the labour market, and for many mid-sized companies building an in-house data science team simply does not pay off.

This is exactly where the targeted use of external specialists has proved its worth: experienced AI developers and data architects who take an initiative from feasibility study to live operation and transfer their knowledge to the in-house team. The key is to treat these experts not as interchangeable suppliers but as partners you retain across several project phases – because every change of personnel means losing contextual knowledge of your data and processes.

Start pragmatically, stay measurable

A proven approach: select a single, commercially relevant process, build a near-production pilot within six to twelve weeks and measure from the outset – processing time, error rate, stakeholder satisfaction. If the results are positive, scale up; if they are negative, the lesson has been an inexpensive one.

In this way, AI turns from a buzzword into an ordinary investment decision with a business case – which is exactly where it belongs.