We replaced Copilot Business about four months ago after 10 months of it not improving. 220 developers, mix of VS Code and IntelliJ. I wanted to share what changed and what didn't for anyone evaluating a similar switch.

The reason we moved was straightforward. After nearly a year, suggestions still didn't reflect our architecture. The tool had no memory of our codebase and token costs kept climbing with nothing we could do about either problem. We evaluated three alternatives and decided to implement tabnine which had a persistent context engine. After the context engine indexed our repos and Confluence docs the change in completions was noticeable within the first two weeks. Inline suggestions follow our patterns now. It suggests our custom decorator setup, uses our internal HTTP client instead of raw fetch, and follows our error handling conventions. Token costs dropped because each request isn't assembling a full context payload from scratch every time.

The VS Code extension itself is comparable to Copilot. Ghost text works well, keybindings are configurable, plays nicely with ESLint and Prettier. Initial setup is more involved than Copilot's install-and-sign-in flow because you're configuring the connection to the context engine endpoint. Not difficult but worth knowing going in. Where it's weaker is that the chat isn't as polished as Copilot Chat. Inline generation from comments is more conservative. No equivalent to Copilot Edit for multi-file editing, which is a real gap for refactoring work.