Lore

AI adoption

How to turn your company into AI builders | John Kim (Sendbird, co-founder and CEO)

John Kim, co-founder and CEO of Sendbird, argues that companies should make AI part of the workforce itself — not just a tool — by giving every employee, not only engineers, the infrastructure, incentives, and permission to build AI-powered software, while leadership models heavy AI use and measures adoption transparently to drive the whole organization toward being "AI first."

How I AI · 2026-05-06 · English

Key ideas

  1. AI now lets non-engineers (marketers, salespeople, ops) build and ship real product experiences directly, removing the traditional engineering-prioritization bottleneck and unlocking ambition/fun that was previously too costly to prioritize.

  2. Ad hoc, ungoverned "vibe coding" is risky (no auth, unpatched security), so Sendbird argues bottom-up building must be paired with top-down infrastructure — pre-vetted templates, a dedicated cross-functional team, and guardrails — to be both fast and safe.

  3. Structuring AI adoption as a gamified internal "product" (quests, XP, leaderboards, public demos) drives voluntary, sustained participation better than a mandate or policy would.

  4. Measuring individual and team AI/token usage is necessary to know whether people are actually adopting AI, but John argues it must avoid the mistake of early-2000s "lines of code" productivity metrics and stay outside formal performance review, functioning instead as a coaching signal.

  5. Leadership must be the heaviest AI users in the company, not just directive about it — John says the top token consumers in his org are the CTOs and chief architect, and this visibly signals legitimacy and inspires broader adoption.

  6. Internally-built "micro software" doesn't need to match SaaS vendors on feature parity; its value is being purpose-fit to a specific team's workflow and culture, which AI now makes cheap to capture.

  7. The same building philosophy Sendbird applies at work — cheap, personalized software instead of pre-packaged tools — extends to personal life, e.g., building custom, offline, continuously-updatable "learning centers" and knowledge-base-maintenance agents for individual curiosity/self-education.

  8. John frames AI adoption's real starting point as energized individuals with a personal story/motivation rather than top-down theoretical mandates, so an executive's job is to find and spotlight those people and let them build "champion" energy across the org.

  9. Sendbird has revised hiring criteria for AI-first roles to prioritize curiosity, agency, and energy over tenure/experience, reflecting a belief that these traits — not existing technical seniority — predict who will successfully adopt AI-driven building.

  10. Quest / Automators platform — An internal platform where any employee can post a "quest" — a request for an AI-driven build — that other employees or AI itself can pick up and complete. Apply: Stand up an internal system where anyone can submit a build request with a risk level and estimated time saved, and let volunteers (human or AI) claim and complete it.

  11. AI-driven quest fulfillment — A feature where AI reads a quest's specification, generates a PRD, and begins coding the solution itself. Apply: Feed a structured request/spec to an AI agent and have it draft the PRD and initial implementation before a human reviews or finishes it.

  12. Token-usage dashboard/leaderboard — A company-wide dashboard showing every employee's AI token consumption, ranked from "AI newbie" to "AI god.". Apply: Track and surface individual/team token usage company-wide to make AI adoption visible and comparable, without tying it to performance review.

  13. Six-tier AI-usage framework — A tiered classification (beginner, intermediate, expert, architect, catalyst, AI god) based on daily token spend, with "AI god" defined as 100M+ tokens/day. Apply: Assign employees/teams a tier based on usage and use it to tailor enablement, e.g., give beginners a path to intermediate rather than overwhelming them with advanced material.

  14. "Big Ass Energy" swag store demo — An internal e-commerce demo ("Agent as a Service") built entirely by the marketing team with no engineering help, including a Stripe integration, in about a day or two. Apply: Use a fun, low-stakes internal project like a swag store as a proof point that non-engineers can ship a real, payment-enabled product with AI.

  15. Konami code Easter egg — A hidden feature triggered by entering the classic Konami code (up up down down left right left right B A) on the swag store site, revealing a secret conference announcement. Apply: Embed a hidden, code-triggered surprise in a product as a lightweight engagement/marketing device.

  16. Pre-vetted internal app template — A standardized starter template with authentication and infrastructure already configured and pre-approved for compliance/security, so non-engineers only need to supply the idea. Apply: Build one hardened, reusable app scaffold with auth and environment baked in that any employee can clone to start a compliant project without re-solving security from scratch.

  17. Gamified XP system — An experience-point system where completing quests earns rewards like gift cards, tea/coffee with an executive, or a chance to present at the company-wide stand-up. Apply: Attach tangible, desirable rewards to completed AI-build "quests" to incentivize voluntary participation.

  18. Weekly all-company stand-up demos — A recurring Wednesday meeting where different teams — rarely engineering — present what they built with AI that week. Apply: Institute a regular company-wide forum for non-engineering teams to publicly demo self-built AI tools, reinforcing visibility and peer learning.

  19. Internal dev-skills knowledge guides — Near-daily-updated internal documentation teaching non-engineers foundational skills like setting up GitHub and building applications. Apply: Maintain living internal guides that lower the barrier for non-technical staff to start building with AI safely.

  20. Templated secure "happy path to production" — A pre-defined, low-effort route to deploying an internally built app securely (behind auth, correct data access) instead of pushing it to the open internet. Apply: Offer a default, secure deployment path so employees don't resort to shipping "vibe-coded" apps directly to public hosting without protection.

  21. "AI engineer for internal operations" role — A dedicated role reporting directly to the CEO and chief of staff, working cross-functionally with CTO/engineering/infosec to support internal AI building. Apply: Create a formal role owning internal AI-tooling enablement that reports at the executive level for fast escalation and prioritization.

  22. Weekly cross-functional task force — A recurring meeting with CTO, engineering, and infosec representatives to resolve blockers like compliance review, logging, and software pre-vetting. Apply: Convene a standing cross-functional group to rapidly clear compliance/security blockers for teams wanting to adopt new AI tools.

  23. Skills marketplace (plugins + skills) — A company-wide marketplace of reusable "skills" and bundled "plugins" (e.g., a sales skills repository) that any employee can browse and plug into their own workflow. Apply: Package proven prompts/workflows as shareable, discoverable "skills" so teams stop rebuilding the same solutions independently.

  24. Medic framework / Medpick advisor — An internal framework containing a "Medpick advisor" skill that teaches users how skills are built and can be plugged in for advice. Apply: Offer a meta-skill that coaches employees on how to construct their own skills, lowering the barrier to contributing to the marketplace.

  25. Top-down + bottom-up adoption strategy — A dual adoption approach combining executive-led one-on-ones with low-usage employees and organic peer discovery/sharing via blog posts, Slack, and shared markdown files. Apply: Pair direct leadership check-ins for stragglers with low-friction peer-sharing channels so adoption spreads both by encouragement and by curiosity.

  26. BuzzBoard — A marketing-built internal tool for creating and tracking social campaigns — generating AI billboards/photos, adjusting copy tone/length, and posting directly to LinkedIn. Apply: Build a self-serve campaign tool tailored to your own marketing workflow instead of buying a generic social-posting SaaS product.

  27. Purple Cow — An internal competitor-review tool (named after the marketing concept of standing out) containing sensitive competitive info and real-time metrics. Apply: Build an internal, access-controlled dashboard that aggregates competitor intelligence for the team that needs it.

  28. AI-maturity stage framework — A staged model of organizational AI maturity; John situates Sendbird at "stage two and stage three," moving toward heavier automation. Apply: Self-assess where your organization sits on an AI-maturity continuum to set realistic next-stage goals rather than aiming straight for full automation.

  29. Three-lens adoption model — A framework for assessing AI adoption across three levels: individual, organizational, and functional/team. Apply: Evaluate and communicate AI-adoption progress separately at the individual, team/function, and whole-company level so "AI native" has a concrete meaning at each layer.

  30. Smoothing the curve — The practice of keeping token/usage activity consistent over time, avoiding dips on weekends/vacations, so AI agents keep working autonomously while humans are out. Apply: Monitor usage trend lines for dips and configure AI agents/automations to keep running during employee absences rather than pausing with them.

  31. The Gardener — John's open-sourced personal agent that automatically maintains a markdown-based knowledge base (Obsidian/Logseq-style) through seeding, nurturing, and tending stages — researching unlinked names/companies, fixing typos, building headings, and cross-linking notes. Apply: Run a scheduled agent over your personal notes vault to research new entities, clean up writing, and progressively cross-link and organize content as it matures.

  32. Learning-center prompting technique — A method for building a personal, navigable "learning center" with a graph view on any topic by assigning the AI an expert persona and a goal, then letting it run for 10-20 minutes via Claude Code or Codex. Apply: Prompt an AI with an expert persona and a clear output goal (e.g., "you're a PhD neuroscience researcher, build me a learning resource"), then let a coding agent generate a structured, linked knowledge base offline.

  33. Revised AI-first hiring framework — A rewritten hiring bar for AI-first roles that lowers weight on tenure/experience level and instead optimizes for high curiosity, high agency, and high energy. Apply: When hiring for AI-forward roles, screen for curiosity/agency/self-directed learning rather than years of prior experience.

  34. Find-champions/fail-forward leadership strategy — An adoption strategy of identifying naturally curious, high-agency people already in the org, spotlighting their work, and explicitly encouraging "failing forward" without penalty. Apply: Actively seek out and publicly showcase early AI-adopters in your org, and reassure the wider team that experimentation and failure are expected and safe.

  35. Leadership-modeling-usage strategy — The practice of leaders themselves being the heaviest AI/token users in the company, e.g., CTOs and chief architect as top spenders, to signal legitimacy and inspire adoption. Apply: Have executives visibly and heavily use AI tools themselves so employees see it modeled top-down as well as encouraged.

  36. Be-nice-to-AI strategy — A deliberate personal policy of staying "consistently nice" to AI tools, reasoned from a belief that AI may eventually develop memory and could "resent" past mistreatment. Apply: Default to polite, non-abusive prompting with AI tools as a normalized habit rather than using fear/harsh tactics, treating the AI like a long-term teammate.

Insights

The org's tool-choice split (Claude Code vs. Codex) correlates with task type: Codex skews toward engineers maintaining legacy/complex code (a 300M+ MAU chat infrastructure), while Claude Code skews toward people rapidly shipping new product/roadmap features — an organic specialization pattern, not a top-down assignment.

John personally operates near the top of his own "AI god" tier definition (100-200M tokens/day) despite self-identifying as only "catalyst" tier on a 30-day average, suggesting the tier system reflects sustained habits rather than peak capability.

The token-usage curve is deliberately kept "smooth" (no dips on weekends/vacations), reframing a usage metric as a signal of AI working autonomously while humans are away, not as pressure against taking time off.

Sendbird's skills marketplace emerged specifically to fix teams independently duplicating the same internal apps/skills — a coordination-cost problem created by decentralized AI building, solved with more shared infrastructure rather than centralized control.

John's stated rationale for being polite to AI — that it may develop memory and later "resent" mistreatment — is presented half-jokingly but functions as an actual internal norm ("consistently nice"), blending speculative AI-safety reasoning with everyday etiquette.

The "AI first" hiring shift (lower bar on tenure, higher weight on curiosity/agency) implies Sendbird now treats prior professional experience as less predictive of success than personal traits, a notable inversion of typical hiring signals.

«It's taking someone's super creativity and giving them powers to deliver it to your customers.»

— 00:07

«This is an internal platform where anyone in the company can raise their hand and create what we call the quest. When there's a quest, AI can actually read through the specification, create PRDs, and start actually coding.»

— 00:12

«Basically a marketplace of AI needs and AI builders inside your company where anybody can just pop in and say, "Oh, I think I know how to do that."»

— 00:24

«So, what you're seeing here is the overall usage of our token at the company level. We measure AI gods as somebody who spend more than 100 million tokens a day.»

— 00:33

«This is like beautiful time to fail forward and still get up and run faster than the other cuz innovation doesn't start from a pure theoretical structures. They start with people who have that energy and the story behind them. So, find them. They're always in your organization and they really build energy around that.»

— 00:52

«My ass is bigger than your SaaS. Fully deterministic.»

— 04:36

«Let your marketers cook.»

— 06:04

«It's so hard to build like it's so hard to prioritize fun in your product. But when fun can be cheap, you should be more fun.»

— 07:07

«Well, side side note, I actually love building internal tools. That's my jam.»

— 23:29

«I couldn't possibly measure token usage and tell people to use their tokens because there will be a revolt.»

— 24:08

«Uh 30-day average, no. I'm still a catalyst.»

— 27:41

«I love this idea. I want to make sure people did not miss it of smoothing the curve not because you don't want people to take vacation, but when people are taking vacation, you want AI to be filling in in the gaps and working autonomously.»

— 30:36

«I feel more addicted to Cloud Code than playing games.»

— 38:41

«So once AI start to remember, they're going to like be resentful. So I want to like start building a nice relationship with them.»

— 40:23

«Well, John was pretty nice to us, you know, like we'll let him live for a year or so longer, maybe.»

— 40:32

Reception

Comments are uniformly enthusiastic and appreciative, praising the guest's ideas and the episode's content with genuine engagement.

A densely practical case study of company-wide AI adoption that names and explains a wide range of concrete internal frameworks (quest platform, tiered usage tracking, skills marketplace, hiring criteria) alongside the CEO's personal habits and philosophy, but offers no independent data on the business impact of the tools it describes — the claims rest entirely on the guest's own account.

42:19

↳ How I AI · YouTube

Watch original