Source: In Context — Vijay Rayapati on Atomicwork
Source: In Context — Vijay Rayapati (Atomicwork), BoldCap, 1:24:32. YouTube: ytIphLE3j4A. Transcript: auto-generated captions — verify names/figures against a second source before external use.
What It Covers
BoldCap GP “Sathya” interviews Atomicwork co-founder/CEO Vijay Rayapati in BoldCap’s first podcast episode. The conversation spans Vijay’s serial-founder lineage (Trilogy/Jensen → Minjar → Nutanix), the “traditional SaaS is dead, AI is about workforce” thesis, why Atomicwork is rebuilding enterprise service management (the ServiceNow “back office” category) with an AI core, the shift from chat to voice/vision/action agents, technical hard problems (context, memory, verifiability, reinforcement fine-tuning), and the Khosla Ventures-led Series A. This is the highest-density founder-voice source for Atomicwork’s strategy and positioning.
Key Claims (with timestamps)
Founding timing & founder lineage (05:06)
Vijay traces an AI/ML thread across his career: early stint at Trilogy, which acquired “Jensen”^[ambiguous — company name garbled in auto-captions], an “‘80s real-time expert system company” of PhDs doing forward/backward-propagation anomaly detection (used even for International Space Station monitoring); a 2009–11 analytics company; then a cloud company “MAR” (almost certainly Minjar^[ambiguous — likely Minjar, garbled in captions]) “which I sold in N[u]t[a]n[i]x.” Atomicwork started “two months before ChatGPT” — GPT-3.5 was the “aha moment” that collapsed a planned 10-building-block NLP/transformer stack into something buildable (06:13).
”SaaS is dead; AI is about workforce” (09:01)
“I do believe that traditional SAS is dead from a workflow perspective … because AI is about workforce, you know, it’s not about workflows … SAS is all about workflows.”
He attributes the original “SaaS is dead” line to social media, not to Satya Nadella^[ambiguous — captions: “Satya Nadila”], who he says clarified at a Bangalore summit that he never said it.
SaaS caused fragmentation; Atomicwork’s convergence thesis (21:48)
SaaS drove an explosion of tools, with three side effects: fragmented data, fragmented workflows, fragmented experience. Atomicwork’s thesis is “it’s a time to converge … build a new digital backbone for business.” Traditionally called enterprise service management; they frame it as a “new digital service management” that uses AI to resolve that fragmentation — software should proactively detect when a user needs help rather than telling them to “go raise a ticket.”
Three category icons; “complex and boring” (26:44)
“Last 25 years … three great companies got built in enterprise software … front office it was Salesforce, for middle office it was Atlassian, for back office it was ServiceNow.”
He argues B2C is “complex and sexy” while B2B is “complex and boring” — and IT/ITSM is the canonical boring-but-essential job (“24/7 … everything is an interruption … an escalation … an abuse”) that AI will take over first, faster even than customer support.
”Believe in the model more than your mental model” (34:04)
Cites advice from “Bill Kuran … partner [at] … SQUA”^[ambiguous — name and firm garbled; possibly Bill Gurley or a Khosla partner — verify]: “if you think you’re an AI company then you have no company” — AI itself is an OS-layer business (OpenAI/Microsoft/Google/Anthropic). The opportunity is owning the work, jobs, workflows, and workforce before and after AI is applied. Echoes Sam Altman^[ambiguous — captions: “Samman”]: “don’t bet against the model.” He also credits seeing BabyAGI^[ambiguous — captions: “BB AJI”/“baby hi”] shortly after ChatGPT as the moment multi-agent systems felt “generational” (31:46).
Chat is “analog AI”; voice + vision + action (37:28)
“Chat is a low frequency interface for a human … I believe the chat AI is analog AI. The digital AI is voice and vision.”
Atomicwork launched an agent called “atom,” then upgraded it to a “universal agent” that is voice- and vision-driven (“it can see what’s happening on the screen”). Next is “action AI” (clicking, doing things) — voice + vision + action = “AI employees / digital workers.” Notes a behavioral split: in-office employees prefer chat, work-from-home users mostly talk to it.
”Decade of agency” (38:36)
“I think this is more than a decade of agents … this is the decade of agency … where software can have agency and such high autonomy” — and 24/7.
Context/memory/verifiability are the hard problems (43:35)
“Our work context is like a trillion-token context.”
Long-term enterprise memory (job roles, guardrails, metadata, allowed tools/workflows, approval chains) is unsolved; working context is improving but insufficient. Verifiability and auditability (“what you did, why you did it”) are critical in workflow automation, unlike codegen where the PR is the audit. Flags metadata as the gap AI models lack (“they have world’s data … what they don’t have is your metadata”) — citing Salesforce/Informatica and ServiceNow/“data world”^[ambiguous — likely Data.World] acquisitions.
99% per step compounds badly; reinforcement fine-tuning (47:57)
“If AI gets to 99% accuracy and you have a six-step process … 99 × 99 × 99 … it will probably drop to … 81, 72 … 63 … 45 … 30.”
Conclusion: AI will rewire six-step workflows into dynamic 2–3-step ones, and pushing accuracy from 90→95→99→99.9% for a domain requires reinforcement fine-tuning (now exposed via OpenAI’s RFT APIs) plus memory. He sequences AI adoption as CX → DX (developer) → enterprise/employee → partner experience.
The Khosla Series A and “I don’t care about ARR” (1:07:21)
Raised from Khosla Ventures (Vinod Khosla^[ambiguous — captions: “Venod/Vinod Kosla”] and “Kano”^[ambiguous — partner name garbled]), describing it as intentionally choosing investors who would “orbit-shift our thinking” over a higher valuation — “we want to get into the inner circle of AI.”
“I was at Khosla CEO summit … he said ‘I don’t care about ARR — build a great team, build a great product and amazing technology’ … the team you build is the company you build.”
ATOMIC culture acronym (1:04:35)
Atomicwork’s cultural framework spells ATOMIC: Taste, Agency, Ownership, Mastery, Impatience, Customer obsession. Taste framed as “doing tiny things that can shine” (cited example: keyboard shortcuts built by his co-founder).
”Every software will be rewritten”; Indian platform moment (1:18:06)
“Every software will be rewritten … can be rewritten and should be rewritten … from the core.”
Frames Atomicwork as rewriting enterprise service management with an AI core. Argues incumbents will adopt AI and survive but it’s “like upgrading the brain of an old person — you also need new hands.” Contrasts Zoho (~100+ apps), Freshworks^[ambiguous — captions: spelled correctly here] (~10 apps), and Atomicwork as “about one product,” and says Indian founders can now dare to build platform/product companies, not just app/use-case companies.
Biggest changed mind: autonomous agents (1:16:23)
“I thought autonomous agents will take much longer … I’ve been very wrong on that.”
As of “even 3–6 months ago” they were still planning AI workflows; now the goal is AI workforce. Bets for the next year: reinforcement fine-tuning + memory (1:21:29).
Limitations
- Auto-generated captions — proper nouns and figures are unreliable; every
^[ambiguous]above flags a likely transcription error (Minjar, Satya Nadella, Sam Altman, Vinod Khosla, BabyAGI, the Khosla/“SQUA” partner names, Data.World). Verify against a second source before external citation. - No hard numbers. The interview gives no Series A dollar amount, valuation, ARR, customer count, headcount, or named customer logos — all stated qualitatively.
- Co-founders not covered. Vijay references “my co-founder” once (~1:06:15) but the interview does not name or discuss the co-founding team (e.g., the ex-Freshworks founders) or their backgrounds — do not source co-founder claims from this page.
- No competitive detail vs. specific rivals. ServiceNow/Salesforce/Atlassian appear only as category archetypes; no head-to-head feature/pricing comparison.
- Heavy on macro/philosophical digressions (AI safety, bicycle history, humanoids) trimmed here as low-signal for competitive intel.