GitHub brings together two developments that at first glance have little to do with each other: the licensing of AI assistance in the enterprise and the statistics on how work is actually done on the platform. Both point in the same direction. AI-assisted development is leaving the experimental phase and becoming part of the regular toolchain that must be procured, billed and monitored.
Copilot Access: simpler access for enterprise customers
Under the name "Copilot Access", GitHub simplifies access to the Large Language Model Copilot for enterprise customers. A Large Language Model, or LLM for short, is a language model trained on very large amounts of text that can complete or generate code and text.
For larger organisations, obtaining such functions was previously often a project in its own right: licensing models, user assignment and approvals had to be clarified before teams could even begin productive testing. This is exactly where the simplification comes in. The message to procurement and IT management is that the use of Copilot in the enterprise is becoming more plannable.
Model routing is intended to save tokens
The second point from the same announcement concerns ongoing costs. Better model routing is intended to help save tokens. Tokens are the smallest processing units of a language model – roughly speaking word or character fragments – and they usually form the basis for billing.
Model routing means that a request is not automatically sent to the largest available model, but to a model that suits the task. A short code completion requires different capacities than an extensive refactoring analysis. Routing smaller requests to smaller models lowers token consumption without sacrificing performance on complex tasks.
For companies, this is the really interesting lever. The cost of AI assistance does not scale linearly with the number of licences, but with the actual intensity of use. A routing mechanism that works in the background makes this cost curve more predictable – provided that usage is measured at all.
The activity report: agents in everyday work
The platform's latest activity report shows how strongly AI has now arrived in day-to-day business. According to the report, GitHub records 3.5 million merges – per day. A merge is the merging of a development branch into the mainline of a project, i.e. the moment when a change actually becomes part of the codebase.
This figure is more meaningful than pure commit or repository statistics because it counts completed, usually reviewed work steps. It describes throughput, not activity for its own sake.
At the same time, according to the report, TypeScript remains the most popular programming language on the platform – according to heise developer, also thanks to AI agents. TypeScript is an extension of JavaScript with a static type system that checks data types before execution.
Why TypeScript in particular
It is plausible that a typed language is ahead in a phase of stronger AI use. Type information gives both tools and generating models additional context: which fields an object has, which return values are permitted, where a call does not fit together. Errors that a model produces therefore come to light earlier – not only at runtime for the user.
The report provides a correlation here, not a fully formulated causal analysis. The connection between agent use and language distribution should therefore be read as an observation and not as proof. For technology selection in concrete projects, however, it is still an argument that one should know about.
What the two reports show together
The licensing announcement and the activity report describe two sides of the same development:
- Procurability: Access to AI assistance is becoming easier for companies and thus a normal licensing decision.
- Cost control: Model routing addresses consumption, not just the purchase price.
- Measurability: Merges per day are a metric for work actually completed, not for tool usage.
- Technological impact: The language distribution shifts noticeably when agents participate.
Classification for project practice
Neither report can be used to deduce how much code is produced by agents today or how high the quality of these contributions is. The sources make no statements on this. Anyone who wants to make statements in their own organisation needs their own metrics – such as pull request lead times, review effort or error rates after release.
Another important point remains: simplified access does not replace governance. Anyone rolling out AI assistance company-wide should first clarify which repositories are included, how sensitive content is handled and who bears review responsibility. Even with agents, approval remains a human decision.
What this means for web and software projects
For ongoing projects, the discussion is shifting from "Are we using AI tools?" to "How do we control consumption, quality and responsibilities?". Simpler licensing and model routing make AI assistance calculable, and the high merge throughput shows that the bottleneck in future lies more in review and quality assurance than in writing code.
Anyone who now establishes type-safe technologies, clean review processes and robust metrics will benefit from this development – instead of merely managing it.
Sources
- "Copilot Access": GitHub makes Copilot available to enterprise customers — heise developer News
- GitHub: TypeScript remains the most popular programming language thanks to AI agents — heise developer News