The discussion about artificial intelligence in companies has shifted. The question in the foreground is no longer whether individual teams may use a language model, but what place AI actually occupies in the organisation at all. A current assessment puts it in a nutshell: many companies have so far used AI only selectively as a tool. Anyone who wants to work AI-native, by contrast, cannot avoid structural changes – including new roles.

This difference sounds academic, but it has very concrete consequences for web and software projects. A tool is brought out when needed. Infrastructure, by contrast, is always there; it shapes how work is organised in the first place – just like version control, deployment pipelines or a central content management system.

The pattern: many tools, little impact

The typical entry point follows a similar path. A development team uses a code assistant, the editorial team has texts pre-formulated, the sales team tests a summarisation function. Each of these steps achieves something in its own right. Together, however, they rarely produce a noticeable change at process level.

The reason lies less in the technology than in connectivity. Results are produced in isolated interfaces, not in the system in which the work actually takes place. They are copied, reworked, passed on manually. Time gained in one place is lost again at the next interface.

On top of that: selective use remains invisible. There is no shared understanding of which work steps are even supposed to be AI-supported, what quality is expected and who remains responsible when a result is substantively wrong.

What the infrastructure mindset changes

Thinking in terms of infrastructure means moving from the individual application to the level of processes, data and responsibilities. Three shifts are particularly relevant here.

  • From tool to process step: AI is not regarded as a separate application, but as part of a defined workflow – for example in the editorial workflow or in code review.
  • From interface to API: What matters is whether a model is connected to the existing systems via a programming interface, instead of existing only in the browser window.
  • From individual initiative to role: It takes people who are responsible for quality, data access and the further development of AI-supported workflows – not just committed individuals in the departments.

The last point in particular is underestimated. New roles do not arise automatically just because a technology is available. They must be described, filled and equipped with decision-making authority.

Data and systems as a prerequisite

AI results are only as good as the context they can access. In many companies this context is distributed: in the CMS, in the CRM, in project management tools, in file storage, in the heads of individual employees. As long as these sources are not structured and accessible, any use of AI remains an exercise with incomplete information.

For web projects this means, very practically: cleanly modelled content types, consistent metadata, documented interfaces. A well-structured TYPO3 setup is valuable not only for editors and the frontend, but also the foundation for AI-supported functions to be able to dock on meaningfully at all – for example for suggestions on tagging, translations or internal linking.

Governance: no scaling without rules

As use becomes broader, the need for clear guardrails grows. Which data may be passed to which systems? Who checks results before publication? How is it documented that a piece of content or a section of code was created with AI support?

These questions seem to slow things down, but they are the prerequisite for using AI beyond pilot projects at all. Without binding answers, any expansion remains a risk decision made case by case – and that is precisely what prevents scaling.

A pragmatic path

The infrastructure mindset does not require a complete overhaul. A step-by-step approach makes sense, starting with a concrete workflow and growing from there.

  1. Select a process that is run through frequently and delivers measurable results – for example the creation and approval of content.
  2. Document the workflow and mark the points at which information is transferred manually.
  3. Integrate AI not as an additional interface, but at precisely these handover points.
  4. Define responsibilities: who decides, who checks, who maintains the connection?
  5. Observe the results and only then transfer them to further processes.

The difference from the tool approach lies less in the model used than in the question of what happens afterwards: does the result remain in a chat window, or does it flow into a system from which others can continue working?

Assessment

For web and software projects, the focus thus shifts: it is not the choice of assistant that determines the benefit, but the quality of the underlying structure of data, interfaces and responsibilities. Anyone who is overhauling their CMS, their content models and their approval processes anyway should factor in connectivity for AI-supported workflows from the start, instead of retrofitting it later.

Point solutions deliver quick wins – lasting relief only comes when AI is planned and governed as part of the infrastructure.

Sources