More Dev Interrupted episodes

You cant blame AI for the code you ship | Freshworks Murali Swaminathan thumbnail

You cant blame AI for the code you ship | Freshworks Murali Swaminathan

Published 17 Nov 2026

Duration: 00:36:39

Freshworks uses an AI Product Development Lifecycle to manage code quality and prioritize measurable customer impact over mere tool adoption.

Episode Description

AI can write the code, but a human still answers for it once it ships. Freshworks CTO Murali Swaminathan built his engineering org's approach around t...

Overview

Morali Swaminathan, CTO at Freshworks, describes how the company opened its gates to AI coding tools and quickly saw agentic workflows become the norm. The initial enthusiasm produced a flood of code, but more code did not translate into faster shipping. The cost shifted downstream to whoever had to review and test the output. In response, Freshworks developed an AI Product Development Lifecycle to reintroduce discipline into how code is built, tested, reviewed, and deployed.

He explains that engineers are now expected to do design thinking, architecture, and experimentation, and to think like product managers. The lines between backend, frontend, and architecture roles are blurring. Engineers are encouraged to own whatever they work on end-to-end, including testing and security checks, rather than relying on separate specialists.

On the customer side, Freshworks serves 75,000 customers across tech and non-tech industries. Swaminathan emphasizes that trust with non-engineer users is earned through outcomes, not granted. He cites Seagate, which migrated 30,000 employees to Freshworks in about three months and saw at least 30% of tickets deflected after deploying AI agents. He argues that adoption metrics matter less than whether the system helps users get their job done faster, with better quality, and reliably.

What If

  • What if you audit your AI-generated code for verbosity before it reaches human review?

    • Move: Set a complexity or line-count threshold in your CI pipeline that flags AI-generated functions exceeding a defined size, forcing a refactor before a reviewer ever sees the PR.
    • Why Now? Swaminathan describes how "anything can be done in 10 lines and not being done in 100 lines" once AI tools flood the codebase, and the review bottleneck is the new constraint.
    • Expected Upside: You could reduce review time per PR and prevent the accumulation of unmaintainable code, as a possibility rather than a guaranteed outcome.
  • What if you build a lightweight end-to-end ownership checklist that the same engineer completes from design through deployment?

    • Move: Create a short checklist covering design rationale, automated test coverage, security validation, and deployment verification that the writing engineer must complete before requesting review.
    • Why Now? Freshworks expects engineers to "do design thinking, architecture" and to not "rely on a tester to test your product" or "rely on a security person to make sure that your product code is secure."
    • Expected Upside: As a possibility, this could reduce handoff delays and make the engineer accountable for the full quality of their contribution, mirroring the "product builder mindset" Swaminathan describes.
  • What if you track one outcome metric per AI feature instead of adoption metrics?

    • Move: For each AI capability you ship, define a single customer-facing outcome (e.g., ticket deflection rate, time-to-resolution) and report on that rather than token usage or feature click-through.
    • Why Now? Swaminathan argues that "a customer doesn't care about any of those" adoption metrics and that "the impact is the most important one that people have to look at when you're rolling out software."
    • Expected Upside: You could align your roadmap with what customers actually value, and as a possibility, make it easier to justify continued investment in AI features that deliver measurable results.

Takeaway

  • Define a structured development lifecycle (your own "AI PDLC") that includes automated testing, review gates, and deployment checks so that AI-generated code is validated before it ships, not after.
  • Set a line-count or complexity threshold for AI-generated code in your CI pipeline to catch the verbosity Swaminathan describes, where "anything can be done in 10 lines and not being done in 100 lines."
  • Assign end-to-end ownership: the engineer who writes the code also handles its testing and security validation, rather than passing those to separate specialists.
  • Measure one customer-facing outcome per AI feature (resolution rate, time saved) rather than adoption metrics like token usage, because "the impact is the most important one that people have to look at."
  • Document reusable context and skills in your codebase so that knowledge lives in the system rather than in individual tribal knowledge, since "a lot of the knowledge is being transferred into the context that's being built by building new code."

Recent Episodes of Dev Interrupted

More Dev Interrupted episodes