The question sounds academic at first, but it has immediate consequences for project organisation: is an AI assistant a tool that someone operates – or a member of the team that you collaborate with?In the follow-up to a specialist programme on software architecture, Avraham Poupko and Ralf D.Müller examine precisely this distinction.In parallel, information from development practice at Google provides a sober view of how far the autonomy of such systems actually extends.
Tool or colleague: why the distinction is practical
A tool is used, controlled and put aside when needed.Responsibility for the result lies clearly with the person operating it.A team member, by contrast, contributes their own input, is involved in consultations and influences how tasks are distributed.Between these two images lies a considerable difference in expectations.
For development teams, this is not nitpicking over terminology.Anyone who describes AI as a colleague implicitly also shifts the attribution of errors, the design of code reviews – that is, the mutual review of source code before it is adopted – and the question of who ultimately takes responsibility for a decision.Anyone who sticks with the tool metaphor retains clear responsibilities, but must explain why interaction with such systems is increasingly dialogical.
What the figures from practice suggest
Statements by the Google engineer Haydn Belfield provide a reality check.According to these, more than 75 per cent of production code at Deepmind is now written by AI.At the same time, 88 per cent of changes still require human control.Both figures belong together: a high proportion of generation is not synonymous with autonomy.
Belfield also sees no measurable acceleration of AI research through AI itself.An internally used indicator, the General Capability Index, shows steady growth without a jump in pace.A self-improving AI remains a thing of the future for now.
The two sources examine different levels and do not contradict each other.One poses the conceptual question of the role within the team, the other provides metrics from a concrete development environment.Taken together, they produce a picture that supports neither the euphoria nor the defensive stance: AI's involvement in code production is high, leadership remains with humans.
Consequences for collaboration
If the majority of changes require human control, the focus of work shifts.Less time flows into writing individual lines, more time into specification, review and contextualisation.For teams, this means concretely:
- Review remains mandatory, not optional. Generated code goes through the same review steps as hand-written code – including tests, architectural alignment and security considerations.
- Responsibility remains tied to individuals. An assistance system cannot be held liable.Whoever approves something bears responsibility for the approval.
- Task descriptions gain in importance. The more precisely requirements and boundary conditions are formulated, the more useful the contributions will be.
- Building experience requires attention. If routine tasks fall away, teams must deliberately ensure that less experienced members continue to build an understanding of the overall system.
Choosing the metaphor with a sense of proportion
The colleague metaphor has a benefit: it describes that collaboration is iterative, that follow-up questions are asked and suggestions are discarded.But it also has a side effect.Anyone who treats a system as a colleague tends to place trust in it that is based on social experience – not on technical verification.
What is practicable, therefore, is an intermediate position: the interaction form of a counterpart, the accountability of a tool.Contributions are discussed, but not believed.Every adoption is justified.This attitude can be captured in team rules and definitions of done – that is, binding criteria for when a task is considered complete.
Managing expectations towards clients
For companies that commission software projects, the sober reading is the more robust one.A high proportion of AI in code generation does not automatically justify the assumption that projects will be proportionally faster or cheaper.The information from Deepmind's development environment suggests that the effort shifts rather than disappears.
It therefore makes sense to make transparent in proposals and project plans where AI support is used, which review steps are planned and who grants which approvals.This creates traceability and avoids discussions about whether an error is attributable to the system or to the service provider – the answer is in both cases: the service provider.
Assessment
For web and software projects, this changes less the question of responsibility than everyday work: quality assurance, architectural decisions and clear requirements become the actual bottleneck, while code generation loses weight.Teams should define which tasks may be worked on with AI support and which review steps remain mandatory.
Anyone who regulates this in writing retains control over quality and liability – regardless of whether the system ends up being called a tool or a colleague.
Sources
- software-architektur.tv: AI as a team colleague — heise developer News
- According to Google, self-improving AI is not in sight for a long time — The Decoder