Generative AI is changing how software is created. Language models formulate classes, configurations and test cases in seconds, and the pure writing work on code is losing weight. A current article in the development environment sums it up simply: Large language models will fundamentally change software architecture – however, the fundamental problems of software development remain untouched by this.

Precisely this dual movement is interesting for project managers. It becomes faster to build something. It does not automatically become easier to build the right thing.

The bottleneck moves forward

In classic projects, a large part of the effort lay in implementation: connecting interfaces, mapping data models, elaborating templates. When this part becomes cheaper and faster, what was previously hidden by it comes to the fore – the question of what behaviour the system should actually exhibit.

Unclear requirements used to produce delays. Today they produce functioning but wrong code, and often in considerable quantities. A language model does not ask whether a discount rule in this form actually corresponds to the business logic. It generates a plausible implementation of what was in the task description.

Thus the value of collaboration shifts to where it always was, but rarely became visible: into the joint clarification of terms, responsibilities and processes between the business department and development.

Collaborative modelling as a way of working

Modelling here does not mean drawing diagrams that are as complete as possible. What is meant is joint work on a shared understanding: Which business objects are there? When does their state change? Who may trigger what? Which rules apply, and which exceptions are permitted?

Such clarifications arise in conversation – at a wall with cards, at a whiteboard, in a workshop with the people who execute the process daily. The result is a model that all participants can read, not just the development department.

The practical benefit is twofold. First, a clean model prevents a misunderstanding from being cast into code. Second, it provides exactly the precision with which AI-supported tools can be meaningfully controlled: anyone who has a domain model with clear terms formulates significantly better specifications than someone who works from a vague requirements text.

What architecture must still deliver

Architecture decisions are decisions about consequences that can later only be corrected at great cost. They concern, among other things:

  • Domain cut: Which areas belong together, which are deliberately separated?
  • Data sovereignty: Which system is the leading source for which information?
  • Integration paths: Synchronous or event-driven, and which error cases must be expected?
  • Operations and security: How is deployment done, how is monitoring done, how are permissions mapped?
  • Changeability: Which parts should be allowed to change frequently, which must remain stable?

For these questions, a language model provides suggestions, but no responsibility. The weighing of effort, risk and organisational reality remains a human task – and it requires knowledge of the specific company.

Consequences for TYPO3 and web projects

Especially on the web, the effect becomes apparent quickly. A content management system maps editorial processes, permission structures and content types. If this structure is modelled uncleanly, workarounds arise: fields that are misappropriated, content elements that no one can assign any more, approval paths that bypass the organisation.

Such legacies cannot be resolved with generated implementation speed – they tend to multiply. Conversely, every hour that flows in advance into a clear model of content types, roles and publication processes pays back twice when implementation becomes cheaper.

In practice, this means a shift in project scoping:

  1. More time for requirements and domain workshops with the business department, less for pure specification documents.
  2. Document decisions, not just results – so that later it remains traceable why a structure looks the way it does.
  3. Upgrade review and tests: anyone who generates more code in less time needs robust verification mechanisms, otherwise the effort merely shifts into troubleshooting.
  4. Use business terms consistently in code and in configuration, so that model and system do not drift apart.

No automatism

It would be premature to derive an end to programming work from this development. The more cautious reading is sensible: the weights are shifting. Typing speed becomes less of a distinguishing feature, judgement more. How strong the effect turns out to be in a concrete project depends on the domain, system landscape and maturity of the organisation.

For web and software projects, this specifically means: the critical path increasingly lies in joint clarification with the business department, not in implementation. Anyone who describes their domain model, their roles and their processes cleanly actually gets speed out of AI-supported development. Anyone who does not gets, above all, faster results that nobody ordered that way.

Sources