90% reduction in internal inquiry time with 3-second average responses. In-house AI chatbot handling government regulations with full security tracking.
In a strictly regulated B2B industry, regulation questions were eating into everyone's workday, and one person bore the worst of it
This client operates in a B2B industry that must follow regulations set by a national government agency. Finding specific documents, understanding procedures, knowing where guide documents are and how to fill them out: all of this requires internal domain knowledge that you can't find online. It only exists in the heads of people who've worked on-site for years. Dozens of these inquiries came in daily. Every employee lost some time to it, but one person in particular was buried, spending nearly the entire day fielding these questions instead of doing their actual job.
“Regulation inquiries affected every single employee. Everyone was losing bits of time to it, but one person in particular spent all day just answering questions and couldn't do anything else. You couldn't find answers by searching — you had to ask a specific person. We needed someone who could really solve this.”
— Client Company CEO
KakaoTalk chatbot as an in-house library, OCR-powered lookup, and security tracking
The client wanted one thing: "Instead of asking me, I want employees to ask the AI." Not a simple FAQ bot. It needed to actually find files in Google Drive and answer regulation questions. Plus a back office to track who asked what.
First, see the system in action. This video shows users searching for documents and uploading photos to automatically query the internal business system, all through a KakaoTalk channel.
KakaoTalk-based AI chatbot — from document search to OCR-powered auto lookup
We didn't try to build everything at once. The project was designed in phases from the start. In Phase 1, we built a document finder. Users add the KakaoTalk channel as a friend, send a message like 'Find the ○○ document', and the bot searches hundreds of files in Google Drive for keyword matches and returns the file URL. What previously required asking someone or digging through Drive folders now gets solved with a single chat message. Average response time: under 3 seconds from keyword input to file URL.
Once Phase 1 was running smoothly, Phase 2 added OCR-based automatic lookup. Take a photo of a document on-site, upload it to KakaoTalk, and the system extracts 5 key pieces of information via OCR, calls the internal business system's lookup API, and returns results in a vertical message format. We first tried a single AI model, but Korean character recognition was unreliable — it kept confusing similar-looking characters. Switching to a hybrid approach combining a specialized OCR engine with a lightweight AI model brought accuracy up to 98%. The cost difference was about 2 won (fraction of a cent) per request.
← Scroll horizontally to view the chart
Async callback flow that bypasses KakaoTalk's 5-second timeout for OCR processing
KakaoTalk chatbots have a strict 5-second response timeout. OCR processing alone takes about 4 seconds, and adding the system lookup pushes it past 10 seconds. To solve this, we implemented an asynchronous callback approach. When a photo is uploaded, a 'Processing your request' message is sent immediately, while OCR and lookup run in the background. Once complete, results are delivered via callback. From the user's perspective, they upload a photo, wait briefly, and receive cleanly formatted results.
Security was a real concern. With an external channel like KakaoTalk, you can't fully control user authentication and authorization. So we built a usage tracking system. Every interaction is recorded: KakaoTalk user ID, what they said, the full request and response payloads. Everything is searchable through the back office. If a security issue comes up, you can immediately trace who asked what and when.
KakaoTalk user ID-based chatbot message log management — tracking messages and timestamps
The back office also provides detailed views of individual conversations. Full request and response bodies are recorded, so when a security issue arises, you can immediately see exactly what question was asked and what answer was given.
Full request/response body logging — immediate traceability for security incidents
The whole thing runs on a single server. KakaoTalk Open Builder calls the skill server, which passes through a web server to the server's internal components: frontend, core backend, and chatbot system. The key logic analyzes user messages and routes them to either document lookup (Google Drive API) or OCR/AI engine processing. A web-based back office shows chatbot usage in real time. For a client with under 10 employees and no infrastructure team, we focused on keeping management overhead as low as possible.
Web Server + Frontend + Core Backend + Chatbot System integrated architecture — KakaoTalk Open Builder routes user messages to Google Drive or OCR/AI engine processing
Enter a keyword in KakaoTalk and the bot digs through internal documents in Google Drive to return the URL. Average response time under 3 seconds with 99% match success rate.
Upload a document photo and OCR automatically extracts 5 key fields, calls the internal business system API, and returns results in a vertical message format.
Combines a specialized OCR engine with a lightweight AI model to achieve 98% Korean character recognition accuracy. Processing cost is in the low tens of won per request.
Bypasses KakaoTalk's 5-second timeout with async callbacks. Immediate acknowledgment message → background processing → result delivery via callback.
Records every KakaoTalk user ID, message content, and full request/response payload. Enables immediate tracing of who asked what and when for security incidents.
Chatbot, API, and back office all run on a single server. Minimized management overhead so a team of under 10 can run it without dedicated infrastructure staff.
The time every employee was losing came back, and security tracking came with it
After deploying the chatbot, every employee who had been losing time to regulation inquiries could focus on their actual work. The one person who was almost entirely consumed by inquiry handling can now do their real job. Repetitive questions like 'Where is this document?' get resolved instantly via KakaoTalk. OCR-powered automatic lookup automated on-site verification tasks. The security tracking framework now covers legal risk management.
Regulation inquiries that had been eating into every employee's time moved to the chatbot. The one person who was buried in inquiry handling all day can now focus on their core work. The need for a dedicated inquiry-handling position simply disappeared.
Previously, staff had to manually enter values from documents to run lookups. Now they upload a single photo on KakaoTalk, OCR extracts key information, and the system lookup completes automatically. Results can be confirmed right in front of the customer, on the spot.
Using KakaoTalk for business information exchange felt risky because there was no way to know who asked what. Now every conversation is recorded, searchable by user ID in the back office. The system is ready for security audits.
Regulation knowledge lived only in specific people's heads. When they left, it was a crisis. Now the system holds that knowledge and gives the same answers no matter who's on the team. New hires can get up to speed without weeks of shadowing.
“Honestly, my first reaction was 'Yeah right, no way that works.' Find files through KakaoTalk? Upload a photo and get the internal system queried automatically? But it actually works. Employees don't ask me where documents are anymore — they all know to just ask the chatbot. The best part is the back office where you can see everything anyone asked. We were always anxious about security, but now that everything is tracked, we can relax.”
B2B Government-linked Business CEOCEO
Process AutomationFull direct sourcing pipeline automated with AI — sourcing time from 20h to 2h, list automation rate 90%+, and coffee chat conversion rate 5x increase. An AI headhunter case study.
PC Automation30 minutes to 1 minute, 50+ clicks to just 3. Monthly EBS exam paper downloads fully automated for an academy instructor, from login to organized folders.
Document Automation12 hours of document sorting cut to 7 minutes with 98% accuracy. AI-powered classification freed an evaluation consulting firm to focus on actual assessments.
Think "Our work requires too much domain knowledge for a chatbot"? This client thought the same at first. Let's talk — solutions become clear when we discuss together.