A manufacturing company that writes software runs into a predictable wall: the codebase grows faster than the headcount.

TCL's situation is typical. Software now accounts for more than half of the business, with over a hundred code branches accumulated internally — legacy debt on one side, cross-platform demands on the other. The goal was not to demo AI writing hello world. It was to put AI into a real engineering process.

How it was deployed

TCL rolled out an enterprise AI coding assistant to more than 90 percent of its core engineering teams.

Coverage is the operative word. The value of AI coding scales with adoption. If only a handful of engineers use it, coding style, review process and institutional knowledge all diverge, and the returns cancel out.

The number that matters

The clearest case is cross-platform migration: moving a project between Android and iOS dropped from two weeks to two days.

That class of work is repetitive, rule-bound and exhausting. Interface adaptation, API substitution, state management rewrites — humans make mistakes and hate doing it. Letting AI produce the first draft and humans review is where the gain is largest.

Why larger teams benefit more

Three reasons.

Big codebases have expensive search. A newcomer hunting for how a function is used loses hours; retrieval takes seconds.

Many branches mean heavy synchronization. A hundred branches implies constant merging and backporting.

Clear conventions make output verifiable. Large teams usually have style guides, test coverage and review gates, which gives AI output an acceptance standard.

The reverse also holds. If a team has no conventions and no tests, nobody can tell whether generated code is correct, and the benefit goes negative — rework costs more than writing by hand.

Risk controls that are not optional

Code review stays. AI-generated code looks plausible and can hide edge-case bugs.

Check dependencies and licenses. Models may pull in incompatible libraries or copyleft terms.

Isolate secrets. Do not paste keys or internal addresses into a chat window.

Keep branch strategy explicit. AI changes arrive in large chunks and conflict easily.

Preconditions for replication

1. Do you have a unified coding standard? If not, write one before adopting tools.
2. Is test coverage adequate? It is the only automated way to judge AI output.
3. Do you have a clear list of migration or refactoring tasks? Best starting point.
4. Will the team change how it works? Tools without process change accomplish little.

The returns from AI coding are not evenly distributed. They are highest on repetitive, rule-bound tasks with clear acceptance criteria, and lowest on ambiguous requirements and architectural design. Finding the former beats pursuing "AI everywhere."