Podcast Wiki

AI R&D Rollout (adopting AI across an engineering org)

5 min read

Summary: Rolling AI out to an R&D org is mostly change management. Start with education and licenses, let people try the main tools (Copilot, Cursor, Claude Code), converge on what works without locking in, and drive the mindset shift from “AI as a chat helper” to delegating whole features. Mandates and licenses come from the top, but the spark usually comes from below: an enthusiast with a working POC who goes team to team, and a guild of early adopters that spreads success stories and workshops. Deliberate challenges (a sprint with no hand-written code) force the shift. Keep humans accountable throughout.

Key ideas

  • Path at Salt Security: Copilot licenses → everyone tries all tools → three-month accelerator → mainly Claude Code (codebase understanding, plan mode), with Copilot kept for tab completion (langtalks—62-ai-rd-rollout, iko-azoulay).
  • Avoid lock-in: tools leapfrog each other; don’t build anything that ties you to one forever (langtalks—62-ai-rd-rollout).
  • Tab completion is a “safety belt” for people not yet through the mindset shift. Choosing Claude Code (no tab completion) can push the shift; compliance reasons include Bedrock models and Cursor’s server-side indexing (langtalks—62-ai-rd-rollout).
  • Prediction: model labs will build the best coding agents because they fine-tune on their own tools (langtalks—62-ai-rd-rollout).
  • GenAI guild: early adopters share success stories in wider forums, run internal workshops, and generate ideas, pulling laggards forward (langtalks—62-ai-rd-rollout).
  • The “no-code sprint”: he suggested a sprint of writing no code by hand. Harder at first, and people sometimes take pushback hard, but those who push through make the shift (langtalks—62-ai-rd-rollout).
  • Give product people access too (Cursor plus repos) so they can ask the code instead of asking engineers (langtalks—62-ai-rd-rollout).
  • A dedicated AI-enablement person can find internal “customers” and build Slack-channel agents for them (e.g. an analyst agent answering a GM’s data questions, with a human analyst verifying SQL; a marketing copywriter agent) (ignore-instructions—24-wonderful-road-to-100m).
  • Measured acceleration is modest overall: the host estimates 10–20% at Mixtiles (a few × in places, none in others). Gurevich sees ~3× on new products, where a founder can clear old processes. A flatter org with far more seniors than juniors; dashboards give way to asking Claude to investigate (ignore-instructions—26-ai-cyber-pavel-gurevich, pavel-gurevich).
  • Little acceleration in sales and G&A. Gurevich argues automated outbound is a step backward and pushes enterprise sales back to in-person meetings, since buyers stake their careers on vendors (ignore-instructions—26-ai-cyber-pavel-gurevich, pavel-gurevich).
  • More than 20% of monday’s code is written by agents (startup-for-startup—362-agent-architecture-fit, netanel-abergel).
  • The main blocker is mindset: people micro-manage agents with 100 small tasks instead of stating the goal and reviewing the agent’s plan; when something breaks, connect tools (e.g. a browser) and tell it to check itself rather than dictating fixes (startup-for-startup—362-agent-architecture-fit, netanel-abergel).
  • Separate fundamentals (model and infra progress, which is steady and logical) from product hype (“the movie app doesn’t actually work”); try things hands-on, since month-old verdicts are already history (startup-for-startup—362-agent-architecture-fit, roy-mann).
  • Three org responses to coding AI: cautious minimum, shallow enthusiasm (licenses plus one training), and a small group that rebuilds practices. Passive “take it and see” doesn’t reach depth (osim-tochna—gong-ai-for-developers, ohad-parush).
  • Someone must own it: at Gong the Developer Experience team owns AI adoption (Claude Code in Docker per Anthropic’s security guidance, manifests, practices), plus a training team running hands-on sessions on developers’ own code (osim-tochna—gong-ai-for-developers, ohad-parush).
  • Top-down mandate works: on one new product, an all-senior team had to work with Claude Code, and seniors became the most hooked; it shone on unfamiliar tech and less on integrating with existing React code (osim-tochna—gong-ai-for-developers, ohad-parush).
  • No caps on sessions or tokens; the cost stays reasonable (osim-tochna—gong-ai-for-developers, ohad-parush).
  • Bottom-up: managers and team leads build weekend side projects (search features, prototypes), and the fast gratification is addictive and spreads (osim-tochna—gong-ai-for-developers, ohad-parush).
  • The spark comes from below: at Cato, licenses came from the top, but one curious engineer got poor results from Cursor at first, learned to work with rules and skills, realized they could be shared org-wide, and became the “missionary” going team to team with a POC in hand (hidden-layers—cato-networks-avishai-zvinsky, avishai-zvinsky).

Disagreements & open questions

Takeaways

ai-sdlc · ai-engineering-metrics · coding-agent-workflow

Graph View

Backlinks