Topic · Generative AI

The process a model is built into determines its value.

Generative AI produces text, code and images. What counts for a business is which process becomes faster or better as a result, and whether that shows up in the numbers.

A chat assistant for all staff is an obvious place to start. The bigger gains come in specific processes such as proposals, contracts, customer enquiries, documentation and software development, because that's where time and quality can be measured.

The model needs four things alongside it: reliable data, criteria for when an output is good enough, control over running costs and a person who answers for the result.

Context

What’s changing and how we see it

What studies show

The effect depends heavily on the task and on experience. In a study of 5,172 customer-support agents, those with an AI assistant resolved 15 per cent more issues per hour on average. Less experienced agents gained most, while quality dipped slightly for the most experienced (Brynjolfsson, Li and Raymond, Quarterly Journal of Economics, 2025). In an experiment with 758 consultants at Boston Consulting Group, participants using AI worked 25 per cent faster on tasks within the model's capabilities. On a task outside that frontier, they were 19 percentage points less likely to reach the correct answer (Dell’Acqua et al., Harvard Business School, 2023).

What follows from this

How much generative AI delivers has to be measured in your own process. Even users' own impressions can mislead. In a randomised study, experienced open-source developers took 19 per cent longer with AI tools, yet believed they had been around 20 per cent faster (METR, July 2025). So we agree before the start what will be measured, and compare against the process without AI.

What matters

Ground rules

  1. Home with the process

    The starting question is which process should get measurably better. The tool follows from that.

  2. Define quality

    Decide up front when an output is good enough, and check it continuously in operation.

  3. Plan costs

    Design model choice, usage and operation so that the benefit exceeds the running costs.

Key questions

What you should clarify

  1. In which processes does generative AI pay off?

    Where a process runs often, involves a lot of text or code and the output can be checked. Whether it pays off shows in a comparison with the existing process.

  2. Who answers for a generated output?

    The person who approves it or passes it on. That should be set out explicitly for every application, along with when an output may go out unchecked.

  3. One model for everything, or several?

    A small selection, chosen by task, data protection and cost per transaction. What matters is being able to swap a model without rebuilding the process.

  4. What do we have to label?

    Since 2 August 2026 the transparency obligations under Art. 50 of the EU AI Act apply, for example when people interact with an AI system or when deepfakes are published. For machine-readable marking by providers (Art. 50(2)), the July 2026 amendment grants a transition period until 2 December 2026.