An in-house AI platform succeeds on operating rules more than on model choice. That is the lesson from D-AISY, built by Daehong Communications — the Lotte Group’s full-service ad agency — with developer partner Ttokttokhan Developer. The agency was spending two days collecting broadcast and OTT schedule data, one to two hours daily rechecking pages for schedule changes, and another two days researching material for client meetings.
Time was not the only issue. Employees were already using outside AI tools for tasks like fact books and social copy, so work context and files were scattering outside the company. The team built individual AI tools and an internal AI portal to manage them at the same time.
Staff sign in with company accounts and use every internal AI service from one screen. The core structure covers login and permissions, a home screen, workflow and AI agent menus, and admin analytics. Five solutions include ad report automation, a program lineup AI, fact book support and social copy generation.
The first release centered on schedule data. It runs scheduled jobs to research new programs and detect schedule changes, and when a user uploads a programming grid, it extracts and stores the details by field.
Because the system depends on AI search and crawling, copyright and legal review came first. Instead of scraping pages directly from its servers, the system pulls URLs through official Naver and Daum APIs and passes them to Google Gemini’s URL context feature to read the content. Proposing options the client’s legal team could review, then agreeing on direction from test results, consumed the most time.
Early versions skipped items when a first search failed instead of retrying with new queries, and the team consolidated a multi-model setup into a single model. Inconsistent English spellings of OTT names were fixed by standardizing on Korean names. Search was limited to 20 major news domains, failed jobs retry up to three times, and schedule saves are wrapped so any mid-process error rolls everything back.
Phase two focuses on sharing prompts and workflows searchable by department, role and tag, plus admin views of participation by team. A mockup image generator and a competitive pitch assistant are also in the works.
The hardest part was not the model or the interface but deciding what data can be used, how it can be collected, and who catches wrong outputs. Before comparing tools, list which tasks must always be human-checked and which can run automatically. A similar lesson appears in Internal AI Agents Need Data Structure Before Models — How Buzzvil’s ‘Iris’ Took Hold in a Public Channel.
An internal AI portal at Daehong Communications where staff use task-specific AI tools — report automation, program lineup research, fact book support and social copy — from one screen.
It splits work into AI processing, automated validation and human review, so critical outputs such as parsed schedules are only saved after an owner confirms them.
Agreeing on legally acceptable data collection, the scope of usable data and who is responsible for catching wrong results.
To apply what you just read to your own site, start with a free audit of where things are now.
A strategist replies within 24 hours on business days.
Buzzvil's BD strategy team has zero developers but runs 32 AI workflow automations — from 7-minute revenue drop analysis to 90% fewer manual steps. How they scaled from personal hacks to team systems.
Meta’s Muse Wonder and Muse Wonder Full add Active Memory and Active Focus so its AI agent can handle tasks that take days. But a Toronto user’s Marketplace mishap shows the risks of agentic AI.
Pew cloned thousands of panelists as AI digital twins and found a 12.4-point average error. Five ways synthetic respondents fail — and where marketers can still use them in research.