Kakao's AI Direction Is Wobbling: What Executives Should Learn (and Fix)
Kakao missed the cut for Korea's "independent AI foundation model" program, stumbled on a KakaoTalk overhaul that it later rolled back, and lost the top app spot in Korea to YouTube. Its stated vision of "people-centered AI" sounds good, but insiders and partners still ask what the company will actually build. The pattern points to strategic drift, slow decisions, and a product engine that isn't shipping clear wins.
This isn't just Kakao's problem. It's a case study in how a strong consumer platform can slip during an AI shift if leadership, product, and go-to-market aren't locked in step.
Recent moves, mixed outcomes
- Early momentum: KoGPT1.0 (2021) and minDALL-E (2021) showed initiative before the ChatGPT moment in late 2022.
- Distraction and delay: Founder-level scrutiny in 2023, then a merger of Kakao and Kakao Brain (May 2024) that signaled consolidation but also indecision.
- Platform reset: "Canana" (Oct 2024) unified models across language, multimodal, image, and voice-yet user-visible impact remains limited.
- Partnership whiplash: A 2025 tie-up with OpenAI added options but blurred the build-vs-partner line for core products.
Critics inside and outside the company point to unclear leadership on AI, a lack of signature features users actually love, and missed timing over the past two to three years. Several senior hires arrived with strong résumés but ran into organizational friction, unclear decision rights, and stalled delivery.
What actually broke
- Strategy whiplash: Internal model development proceeds in parallel with external partnerships, without a crisp boundary of where each wins.
- Weak product signal: Canana is used for summaries and templates, but there's no must-have capability the general public talks about.
- Governance gaps: A KakaoTalk overhaul pushed through over internal pushback and user testing eroded trust and wasted cycles.
- Leadership coverage: A non-technical top seat made it harder to challenge product decisions or set a hard technical bar.
- Talent friction: Big hires, unclear remit. Attrition risk grows when teams can't ship meaningful outcomes.
- Timing: The company missed the window when AI features could have been integrated steadily into the core messenger experience.
Strategic risk on the horizon
Messaging moats can vanish fast. NateOn looked unshakable on PC-and then it wasn't. In an AI shift, a challenger with a better assistant inside messaging could pull users away quickly. If KakaoTalk doesn't add obvious, everyday utility with AI, the brand strength won't save it.
A 180-day reset plan executives can run
1) Lock the AI thesis and win-conditions
- Define three user jobs to win inside KakaoTalk: faster replies, effortless recall, and safer conversations.
- Set clear non-goals. For anything outside messaging and its near adjacencies, partner instead of building.
2) Draw the build vs. partner line
- Build in-house: privacy-preserving message intelligence (on-device where feasible), safety filters, conversation memory with opt-in controls.
- Partner: general-purpose chat or image tools that do not require Kakao's proprietary context.
- Publish a one-page policy that states when the product routes to in-house models vs. a partner API, and why.
3) Ship three flagship features users can feel
- Smart replies that learn per-chat tone and context (opt-in, editable, transparent).
- Call and long-thread summaries with clickable citations back to the chat.
- Searchable, private "personal memory" for groups and 1:1 chats with strict retention limits and a clear off switch.
4) Make privacy and safety a feature, not a footer
- Default to on-device processing for sensitive signals when possible; sync insights only with explicit consent.
- Adopt federated approaches for model improvement; publish a plain-language explainer users can trust. Example: federated learning overview.
5) Fix governance and decision rights
- Stand up a Product and AI Council (CPO, CTO/CAIO, Head of Research, Head of Trust & Safety) with a single accountable owner for KakaoTalk AI.
- Require "stage gates" for major UI changes: user testing sign-off, a 1% canary, 24-hour rollback plan, and post-launch review.
- Cap any one leader's unchecked scope. Split experimentation (explore) from core product (exploit) with clear OKRs.
6) Talent and operating cadence
- Retain top ML and infra leads with milestone-based incentives tied to shipped features and unit economics, not vanity metrics.
- Adopt a six-week build cycle: plan (week 1), build (weeks 2-5), release and learn (week 6). Kill or double down based on data.
7) Capital and data
- Ring-fence budget for GPU capacity, evaluation tooling, red-teaming, and prompt safety. Treat this as core infra, not optional.
- Audit training and fine-tuning data rights. Document data lineage and user consent flows for every AI feature.
- Benchmark partner API costs per interaction and set a hard ceiling with graceful degradation paths.
8) Metrics that actually steer the ship
- Adoption: percent of DAUs using at least one AI feature weekly.
- Quality: response helpfulness score and edit rate for smart replies.
- Speed: median time-to-first-suggestion and end-to-end latency under load.
- Retention: 4-week feature-level retention lift vs. matched control.
- Economics: cost per AI interaction and margin impact per user cohort.
- Safety: incident rates, false positive/negative rates, and time-to-mitigate.
Leadership takeaways
AI strategy is product strategy. Pick where you win, publish the boundary between in-house and partner tech, and ship features users actually talk about. Guardrails on decision-making matter as much as model quality.
If your teams need a fast way to skill up for the above plan-product, data, and leadership included-consider structured paths here: AI courses by job.
Background on the market shift that raised the bar for everyone: ChatGPT introduction.
Your membership also unlocks:
