Why we sent engineers into the field: Kakao's FDE unit
우리는 왜 엔지니어를 현장으로 보냈나: 카카오 FDE 조직 이야기
B · Useful, limited if(kakao) 2026 talk by the Kakao lead of an internal forward-deployed engineering unit, with a clear bottleneck argument and one worked pilot; first-hand and specific, but the single outcome figure comes from a semi-automated pilot with no error or volume data.
Summary based on the full transcript·Original: about 27 minutes to watch
In short
A 27-minute Korean talk from if(kakao) 2026 (uploaded 30 October 2026) by the head of Kakao's forward-deployed engineering (FDE) unit. His argument: after a company "declares AI", individual productivity rises but company productivity does not, because most lead time sits in waits between people and teams, and the bottleneck moves from development to planning, design, QA and operations. Kakao's answer is to send engineers into the teams that own a bottleneck. The worked case automated part of seller contract termination at Kakao Gift, mostly with conventional statistics and RPA rather than LLMs.
Key numbers
- 34 / 15quick wins (effect in weeks to months) and strategic initiatives in the FDE portfolio
- about 80%automated judgements in the seller workflow handled by regression, ranking, ratio statistics, trend and threshold rules, not LLMs
- about 1 hour to 10 minutesa merchandiser's time per seller contract-termination case in the semi-automated pilot
- 4 of 7workflow steps automated: start, AI report, certified mail, 90-day legal-period watch
- about 2,700emoticon images reviewed by eye per day (about 90 sets of 30), the next use case
As reported by the source.
Korean·Video·opens youtube.com
What it reports
- Typical sequence after an AI announcement: general-purpose tools, then MCP and skills on internal data, then agent builders. Agents multiply unseen; even IT security cannot list them, and nobody can state the effect.
- A five-step line where one step handles 20 units an hour delivers 20 regardless of the other four.
- FDEs do not know the domain, so they embed with the owning team. Selection asks whether unblocking one step also speeds downstream steps and total lead time. Internally they are "omni-front managers", one person owning planning, design, development and communication for an initiative.
- Seller termination ran on spreadsheets and documents across 7 steps; data was scattered, decision logic lived in people's heads, and the certified-mail step ran on one shared Windows PC while staff used Macs.
- Design: a screen that decides rather than a dashboard; marts joining seller, product, order, behaviour-log, category and contract data; an ontology in business terms so a rule changes in one place; lineage for every number. Risk signals such as 12 months without a transaction trigger promotion limits, owner assignment or termination.
- Certified mail uses classic RPA, not an LLM computer-use agent, because the step is deterministic.
- Termination conditions, deadlines and final close stay with people. Even so, time per case fell from about an hour to 10 minutes. The demo used synthetic data.
- Emoticon review has five checks (plagiarism similarity, generated-image detection, real-photo detection, quality, ethics); staff review about 2,700 images a day by eye. No outcome figures were given.
- Direction: building each workflow from scratch is the systems-integrator pattern; the aim is a platform where earlier data, logic and actions are reused.
How they got it
The speaker describes his unit's own work and one pilot. The digest is based on cleaned auto-captions.
Limits
- One semi-automated pilot; no error rates, case volumes or effect on total lead time.
- The emoticon case has no results in the talk.
- Self-reported by the unit that built it.
Why it matters for companies
The talk gives a direct answer to why broad AI tool rollouts rarely show up in company results: they speed individual steps while hand-offs set the pace. Putting engineers inside the bottleneck team, and measuring candidates by downstream effect, is a concrete operating model. The 80% figure is the practical correction: most decision logic in an operational workflow needed rules and statistics that are cheap and checkable, with LLMs reserved for the remainder.
Key passages in translation
Korean original
부분 최적화를 아무리 해도 전체 최적화는 여전히 20입니다.
No matter how much you optimise the parts, the whole still delivers 20.
Korean original
전체 자동 판단의 80%가 여기서 처리됩니다. 충분히 강력하고 검증 가능하고 빨랐습니다.
80% of all automated judgements are handled here. It was strong enough, verifiable and fast.
Korean original
이것만으로도 MD 한 명이 판매자 한 건에 쓰던 한시간이 지금은 10분으로 줄었습니다.
This alone cut the hour a merchandiser spent on one seller case to 10 minutes.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com