According to a report, Microsoft and Meta plan to restrict the use of the AI coding assistant Claude Code in their development teams. Claude Code is a tool from the AI provider Anthropic that supports developers in writing, explaining and revising source code. The internal considerations behind this cannot be reliably inferred in detail from the available report – for now, only the direction can be noted: two companies with very large development organisations are scaling back a previously used AI tool.
That is precisely why it is worth looking at the questions behind such decisions. Because they also arise in mid-sized organisations, just on a different scale.
Why such course corrections are not surprising
AI assistants for development have quickly evolved from an experimental field into an everyday tool. In many teams, the introduction has been pragmatic: a tool convinces in a pilot project, licences are expanded, usage grows organically.
What is often missing is a second phase – the sober assessment of whether the actual benefit justifies the actual effort and whether the tool is being used in the right place. In this context, a rollback or restriction is not a failure, but a sign that an organisation is reviewing its toolbox.
Cost control: consumption instead of licence count
Classic software was licensed per workstation. With AI tools, however, costs often depend on consumption: the amount of text and code sections processed, the number of requests, the size of the context a model is to take into account. This makes expenditure harder to plan because it fluctuates with usage behaviour.
In practice, this means: anyone introducing AI tools needs transparency about consumption from the outset. What helps includes:
- a separate cost centre or at least a dedicated line in the project budget for AI tools
- regular analysis of which teams and which use cases account for the largest share
- upper limits or warning thresholds so that unwanted usage patterns are noticed early
- a counter-calculation: how much time does the tool demonstrably save, measured against concrete tasks?
Without this counter-calculation, any discussion about costs remains incomplete. A tool that costs noticeable amounts per month but reliably shortens routine work can pay off. A tool that runs mainly out of habit does not.
Tool strategy: not every tool for every task
Many organisations now use several AI tools in parallel, sometimes with overlapping functionality. This makes sense in a trial phase, but in the long run it becomes expensive and confusing – and makes it harder to build up experience.
A viable tool strategy answers three questions: which tasks should be handled with AI support at all? Which tool is the right one for that? And where does the work deliberately remain manual, for example with security-critical code or architecture decisions?
In practice, it has proven effective to name the fields of application instead of distributing licences across the board. Typical fields are explaining unfamiliar code during onboarding, writing tests, routine work in migrations or preparing documentation. For such tasks, the benefit can also be measured.
Think about dependencies early on
Anyone who couples workflows closely to a single AI product takes a risk. Providers change pricing models, adjust terms of use, reorganise features or limit quotas. There is also the strategic level: providers of AI models and their largest customers are sometimes in direct competition with each other – a tension that can play a role in tool selection at large corporations.
For mid-sized companies, practicable consequences can be drawn from this:
- Check replaceability: Can a model or provider be replaced at reasonable effort without having to rebuild processes?
- Keep knowledge in-house: Prompts, conventions and review steps should be documented – they are the actual operational knowledge, not the tool itself.
- Clarify data flows: Which content leaves the company, which does not? This question must be answered independently of the provider.
- Outline an exit scenario: What happens if the tool is no longer available tomorrow or is significantly more expensive?
Quality remains the team's responsibility
Regardless of the assistant used, AI-supported development shifts the focus of the work. Code is produced faster, but must be reviewed just as carefully as before. Code reviews, automated tests and clear quality criteria thereby become more important, not less.
Anyone who wants to assess the benefit of a tool should therefore not only look at the speed of writing, but at the entire path to a stable release – including rework, bug fixing and maintainability.
Assessment for web and software projects
The fact that even large technology companies are readjusting their use of AI tools shows that this market is not yet in a stable state, and that today's decisions are not commitments for the coming years. For web and software projects, an approach that treats AI tools as replaceable components rather than as a foundation is therefore advisable.
Anyone who clearly defines areas of application, regularly compares costs and benefits and builds processes in such a way that switching providers remains possible can use the productivity gains without committing themselves in the long term.