Murali Swaminathan is the CTO of Freshworks — 15 years old, publicly traded, ~4,500 people — and he spent this hour (58 min) describing what it takes to rebuild that org for agents while the plane is still flying. Almost none of it is about model capability. It’s about the machinery around the model: review budgets, confidence thresholds, access control, and pricing that survives agents replacing the seats you used to bill for.

The number everyone quotes

  • The CEO said on an earnings call that more than half of Freshworks code is written by AI. Murali’s correction: not 100%, but “more than 50% in some cases, more than 70%” — and for newer products it’s above 70%.
  • For changes to existing code it’s a blend of human and AI, with at least 50% AI.
  • Throughput changed by an order of magnitude: engineers used to write ~1,000 lines a day; now it’s up to 10,000.
  • The consequence nobody plans for: review is the bottleneck now. Anything over 500 lines is hard for a human to read, so Freshworks pushes commits much sooner and caps them at 500 lines.
  • Governance went up, not down: AI reviewer tools plus human reviewers on a subset of changes — not every change gets human eyes.

He’s watched four tech waves

  • Pre-internet, client-server, SaaS, now AI — he migrated from India ~30 years ago and has worked through each redefining moment.
  • His one grand takeaway: curiosity. Every defining technology needs a finite amount of time to settle before people understand how it scales, how it secures, how it extends.
  • The SaaS parallel is the honest one: it took companies 10+ years to get scalable and secure, and it took customers 5-10 years to accept that on-premise was gone. Agent adoption will not be faster than that reality check.

Rebuild from scratch, or layer AI on top?

  • The source of record — the ticket schema, how tickets link to users, products, services — didn’t change and doesn’t need to. You still have to store it somewhere to report and act on it.
  • Freshworks was API-enabled from day one, and has since wrapped those interfaces in MCP so third-party AI tools can read and write without hand-coded integration.
  • The deterministic workflow engine stays. Resetting a password has a fixed sequence.
  • What AI adds is the front of the flow: figuring out whether a user with a login problem forgot the password, is on the wrong page, or is using the wrong ID — then calling the reset action.
  • His framing: the handyman’s toolbox. You don’t codify every scenario; the framework picks the tool at runtime, not design time.

MCP, in production terms

  • Two sets: inbound servers (built by Freshworks, how other tools talk to Freshservice) and outbound servers for third-party systems.
  • Why MCP over API: APIs need to be coded and wired; MCP is discoverable and introspective at runtime, so the model can map tool A to tool B and chain them without manual wiring.
  • He uses this himself: “I no longer go to a static dashboard. I no longer go to a database and run queries.” He operates through the MCP servers exposed by those data sources.

When classical ML still beats an LLM

  • Freddy dates to 2018 — traditional ML, classification, clustering, regression, before anyone said GenAI.
  • Most use cases have migrated to frontier models, but some genuinely can’t: ticket field suggestion (which team, which priority, which issue type) needs patterns across millions of records, and you can’t push millions of rows into a prompt.
  • His line: “You can’t use a Ferrari to go to the grocery shop.” Right tool, right job, measured on cost versus benefit.
  • Their heuristic is experimentation, not doctrine: a data-science group evaluates accuracy, the tipping point where accuracy drops as records scale, latency, and token consumption. Summarization goes to a frontier model; structured classification over historical volume goes to ML. Roughly 90% of use cases land on frontier models.

The hardest problem: 40% to 70% deflection

  • Their bar is 90% accuracy. They were at 40%; they’re now in the high 70s. The path was relentless evaluation — training data, data patterns, positive conversations, instruction tuning, feedback loops.
  • Guardrail for the irreversible stuff (payroll, provisioning): below 50% confidence, the agent does not act. It comes back with what it found and asks the human to pick the path — then learns from that choice.
  • “The confidence improves with actions.” More usage plus feedback raises trust, the same way driving experience works.

Bundling, protocols, and why agents can’t skip steps

  • The larger the company, the more the roles separate — at 10 people there’s no real separation; at 10,000 there is. Any agent you deploy has to respect the same protocols a human would.
  • “Just because you’re using an AI agent doesn’t mean I can just skip the mark.” If something needs an approval before processing, the agent routes the approval.

Pricing agents when agents replace seats

  • Per-seat breaks when the seat disappears, so Freshworks sells sessions: a web session covers a 24-hour conversation; email sessions run 72 hours, since a real support thread bounces back and forth before it closes.
  • It’s hybrid — seats for the human agents, sessions for the employees and customers talking to AI — sold in packs sized from historic volumes.
  • Customers want predictable spend, because nobody has an unlimited budget; that’s also why token-maxing behaviors became a problem. The industry is still testing resolution-based and credit pricing.

Code review, PRDs, and promotions

  • AI is verbose, and it isn’t only code: 20-page documents and 30-page slide decks. Freshworks built an “AI PDLC” — a one-page abstract expands into a multi-page BRD that pulls from the feature, the code, customer-reported issues, docs, and the UI.
  • The gate: the PM must review and correct the generated document before it feeds technical design or code. Nothing generated goes straight in.
  • Promotions stay outcome-based: shipping a product that generates revenue counts; a pile of prototypes that never saw light does not, even if it was a lot of activity.

How teams get structured now

  • Fewer specialists. One person often plays developer, designer, tester, and security reviewer on a small project; big projects still get real role separation.
  • Fluid teams of one, two, or five people that form around a problem; people pull tasks off a list instead of waiting on sprint ceremonies. For planned work, groups still form; for unplanned work, someone grabs it and does it end to end.
  • Managers have to become techno-functional investors — deciding what’s worth investing in, rolling up their sleeves, and experimenting themselves. Coding experience is no longer an excuse: “you can’t say I coded 10 years back and I can’t code now.”
  • They’ve flattened layers intentionally so each layer adds value instead of just managing people.

Code is the easy part

  • “Generating code is the easiest part.” The brains go into designing the solution: am I solving the right problem, will it work beyond this one case, what could go wrong.
  • That’s why planning and design agents matter more than the coding agent.
  • Asked how to become irreplaceable, he pushes back on the premise: you don’t want to be immortal, you want problems worth solving. “The more curious you are, the better you are as an engineer.”
  • On his own son, a biomedical engineer at a surgical robotics company: the hands-on, physical part of the work is the moat he sees — soldering circuit boards and working on the robots in a lab isn’t remote work.

Tools, bubbles, and non-human headcount

  • He turned the interview around: try the same task on Claude Code, Codex, and Cursor, and compare. Different tools use different models, and evaluating with one model makes you biased. “Don’t trust one system.”
  • On the bubble question: yes, in the Silicon Valley sense — the valley believes before things are real, and that’s how it evolved. His caveat is the useful half: what works for a tech-first Bay Area company may not work for a Midwest company that isn’t tech-first, so test there too.
  • Vertical vs horizontal AI is a race, not a verdict: “We’re all on a highway, driving at different speeds” — some get there faster, some crash.
  • Agent headcount: he can’t count the agents. Hundreds at least; each has its own thresholds; some he runs himself. His expectation: more agents than employees. SSO gives them visibility into who uses which AI where, and his target state is treating non-human agents like human ones — ids, classification by function (IT, HR, finance), access granted and revoked the same way. The rule for every agent is the same: authentication, authorization, role-based access, and no superpowers.

“Always invite AI to your table and learn to know its limits. Be the human in the loop. Don’t blindly trust its output — and assume this is the worst AI you’ll ever use.” The book he recommends is Co-Intelligence by Ethan Mollick.