---
id: "lib-2026-10-30-what-we-decided-so-we-can-keep-using-ai-panel"
type: "library"
title: "What we decided so we can keep using AI (panel)"
title_original: "AI를 계속 쓰기 위해, 우리가 정한 것들"
lang: "KO"
org: "Kakao, Kakao Pay Insurance, AWS (if(kakao) 2026)"
source_type: "Other"
date: "2026-10-30"
week: "2026-W44"
url: "https://nosov.ai/library/lib-2026-10-30-what-we-decided-so-we-can-keep-using-ai-panel/"
original_url: "https://www.youtube.com/watch?v=dfkKVbeBXew"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "if(kakao) 2026 panel with a Kakao developer, a Kakao Pay Insurance review-agent owner and an AWS specialist on AI cost and governance at three levels; practical, but the customer figures are relayed second-hand by AWS and nothing is measured on stage."
summary_basis: "transcript"
original_minutes: "42"
tags: ["ai technology", "enablement", "ai cost governance", "prompt caching"]
---

# What we decided so we can keep using AI (panel)

Original title: AI를 계속 쓰기 위해, 우리가 정한 것들

Tier B (Useful, limited): if(kakao) 2026 panel with a Kakao developer, a Kakao Pay Insurance review-agent owner and an AWS specialist on AI cost and governance at three levels; practical, but the customer figures are relayed second-hand by AWS and nothing is measured on stage.

Summary based on the full transcript.

## In short

A 42-minute Korean panel from if(kakao) 2026 (uploaded 30 October 2026) on what has to be decided for AI use to last once hundreds of people use it daily. A Kakao server developer covers individual habits: watching a weekly limit, adjusting model and effort first, never cutting planning or verification. A Kakao Pay Insurance engineer covers the cost of a team's AI code-review agent. An AWS specialist covers the company level: one official path through a gateway, with data, identity and logging rules set centrally, and budgets that degrade gracefully. The consensus: the company sets data, security and gateway defaults; the person doing the work picks model and effort.

## Key numbers

- 55%: people using a gateway-based rollout at a Korean chemicals and solar firm above its target, within a month (relayed by AWS)
- about 700: teams at a retail group, each with its own AI budget so one team's spend does not affect another
- 80%+: requests arriving within 5 minutes in one customer's gateway data, making a 1-hour prompt cache cost more than the 5-minute one
- $70+: cost of a single long session, traced to excessive reasoning
- about 2x: token price increase per model tier step, as the Kakao developer manages his weekly limit

## What it reports

- Individual: the developer shows his weekly Claude Code limit in the status line (it resets Monday 16:00) and checks it by Friday. Each model tier up costs about twice as much per token; higher effort helps hard problems but can overthink. He judges models by whether a reviewer would merge the result, and penalises unrequested work. Agents sometimes report done after building only a mock API.
- Team: the review agent comments on each new PR. Free rein reached goals but made cost unpredictable through long detours. Fixes: expensive models only for reasoning steps, cheap models for collection, summarised context instead of full data, and routing to the expensive model only when needed. Analyst and verifier agents and cross-checks were kept because cutting them hurt quality. Human reviewers now focus on business context.
- Company: name one official AI path and decide where data stays (training use, network, in-country), who uses which model and how (short-lived keys tied to SSO rather than long-lived personal keys), and what is logged.
- A chemicals and solar company with sensitive IP opened Claude tools including Claude Code through such a path; in a month, adoption across Korea and the US was over 55% above target, with onboarding in about five minutes.
- A retail group runs about 700 team budgets; new teams are added from a CSV file.
- Budget enforcement varies by workload: hard stop, a 10% grace band with an alert, or throttling at 80/90/100%; low-stakes work can be moved to a cheaper model or paused, critical work cannot.
- Cache and reasoning: gateway data showed the 1-hour cache default cost more than 5-minute caching, so the default was switched back. A session over $70 came from excessive reasoning and was handled by guidance, not blocking.
- A good standard is measured by how rarely people have to stop and ask whether something is allowed.

## How they got it

Panellists describe their own practice; AWS describes customer cases. The digest is based on cleaned auto-captions, which garble several names and product labels.

## Limits

- Customer figures are second-hand from AWS, a vendor with an interest in gateway adoption.
- No cost totals, quality metrics or before-and-after numbers for the review agent.
- Panel format; claims were not tested.

## Why it matters for companies

The panel gives a workable split of responsibility for enterprise AI spend: central rules for data, identity, logging and gateway defaults, with model and effort left to the user. The cache example shows why logging matters: a default that looked sensible cost more once real traffic was measured. Graduated budget controls tied to workload criticality are a better design than a single hard cap.

## Key passages in translation

> 모델은 한 단계 올릴 때마다 토큰당 비용이 약 두 배 정도 어 증가하게 됩니다.

Each step up in model raises the cost per token by about two times.

> 대부분의 이제 80%가 넘는 요청이 5분 안에 요청이 왔었고 결국 5분 캐시를 사용했을 때보다 비용이 오히려 더 많이 발생하는 경우가 어 생긴다는 것들을 확인을 했고

They confirmed that more than 80% of requests came within 5 minutes, so it actually cost more than using the 5-minute cache.

> 각각의 팀별로 예산 관리를함으로써 어떤 팀이 많이 사용했다고 하더라도 이제 다른 팀이 영향을 받지 않는 시스템을 구축을 했습니다.

By managing a budget per team, they built a system where heavy use by one team does not affect the others.

Original: https://www.youtube.com/watch?v=dfkKVbeBXew