Q: What is the Google Claude access policy and how does it affect engineers at Google? A: The Google Claude access policy defaults to using Google’s internal tools like Gemini, but select DeepMind teams have been granted exceptions to use Anthropic’s Claude for coding. This split has created tension because most Googlers remain restricted to internal models while some DeepMind engineers can use Claude, and managers are adding AI goals to performance reviews. Q: Why does Google prefer engineers to stick with Gemini and internal tools? A: Google argues that using its stack protects code and data, reduces vendor risk, and enables dogfooding to improve Gemini. The company also points to custom-built internal infrastructure and IP protection as reasons for restricting external models. Q: Why has allowing some DeepMind teams to use Claude caused internal friction? A: Engineers outside DeepMind see uneven access as unfair and some believe Gemini performs worse for certain coding tasks. That perceived gap matters because teams are being asked to adopt AI and some AI-related goals are now factored into performance reviews. Q: How should a team approach getting approval to use Claude under the Google Claude access policy? A: Teams should present a narrow, testable plan that protects IP, includes measurable KPIs, and can be turned off if goals aren’t met. The article advises defining scope and success criteria, building security and compliance guardrails, running a small pilot, and showing credible ROI and an exit strategy. Q: What security and compliance guardrails does the article recommend for external model use? A: It recommends prompt filters to block secrets, automatic redaction of keys and tokens, data classification to limit sharing to low-risk code, and routing usage through a proxy with logging and rate limits. The guidance also calls for human-in-the-loop code reviews and license checks on AI-generated outputs before merge. Q: What does the suggested pilot program look like in practice? A: The pilot should be small and time-boxed, typically 8–12 engineers across two repositories for about 6–8 weeks, with 2–4 weeks of pre-pilot baseline metrics and a matched Gemini-only control group. The article outlines a 30/60/90 cadence: set up guardrails and baselines, expand to refactors and test generation at midpoint, then run limited feature scaffolding and a final go/no-go review. Q: If a team is denied an exception, how can it still make progress with Gemini? A: Teams should optimize Gemini for tasks it handles well, such as boilerplate, test generation, and doc strings, and use concise prompts that include file context and coding standards. Maintain the same review, scanning, and logging guardrails and provide feedback and examples to internal product teams to improve Gemini. Q: Which metrics are most persuasive when seeking approval to use an external model like Claude? A: Decision-makers look for throughput metrics (PRs per engineer, lead time), quality measures (pre- and post-merge defects, test coverage delta), reviewer load, and reliability indicators like rollback rate. Security metrics such as secret-leak blocks and policy violations are also essential to demonstrate safe use under the Google Claude access policy.