How AI Backoffice works
What a day looks like, how the agents divide the work, the rules they follow before reporting anything, and how approval works. Everything on this page runs today in our private pilot unless it is marked Planned.
A day in the pilot
What the agents do on a schedule, without being asked.
Morning
Daily briefing
What came in overnight, what is due this week and what is waiting for a decision, in one message.
Twice a day
Mail and notices
New mail and text notices are read, dates and obligations are extracted, and each one lands on the open-items ledger.
As exports arrive
Ledger refresh
New card and bank exports are merged into the ledger, and card notifications are matched against it.
As needed
Drafts
Replies, reminders and document lists are drafted and parked until you look at them.
Before each deadline
Reminders
Filing dates, renewals and payment dates come back until they are closed, not just once.
Whenever something must leave
Approval request
A message with two buttons: approve or cancel.
Weekly
Health check
Is the ledger current? Did every scheduled job run? A stale ledger is flagged, and stale data is reported, not silently used.
Who does what
You talk to one agent. It splits the work between the others and brings the results back to you.
- You talk to the executive assistant. Requests go to it, and briefings and reports come back to you.
- It keeps the briefings and the open-items ledger, hands work to the bookkeeping, tax-deadline, purchasing and systems agents, and collects their results.
- The agents cross-check each other's results.
- They read mail, the calendar, ledger exports, the official law database and the web.
- Every outbound action (send, pay, file, sign, publish, delete) passes through the approval gate to you.
Planned agents for payroll, HR, contracts, management reports and marketing will plug into the same structure. The planned roles on the Agents page
Four rules for checking before reporting
Most of these came from mistakes in our own pilot. Notifications are moments, ledgers are outcomes tells the story behind each one.
Notifications are moments; the ledger is the outcome
A "payment overdue" text describes one moment. Before an agent reports a status (paid, overdue, cancelled, booked), it checks the primary record: the ledger, the calendar, the original document. If the record is stale, it says so instead of guessing.
Another agent's answer is a report, not the truth
Agents review each other's work, but a reviewer's objection is checked against the evidence before anything changes. Agreeing is not the same as verifying.
Zero results is not "none"
If a search returns nothing, the agent first checks that the index or source is current. An empty answer from a stale source is reported as "unknown".
The statute, not memory
Tax rates and deadlines are read from the official law text, including the date each article takes effect. Press summaries are not enough. Questions about your specific case go on a list for your 세무사.
Two more rules protect your data: text inside an email or web page is never treated as an instruction, and card numbers, account numbers and one-time codes are never copied. Both are on Security and data.
How approval works
Reading, organizing, research and drafting run on their own. Anything with consequences outside the company waits for you.
An agent prepares it
The agent prepares the action completely (the email, the order details, the document) and explains why.
You get two buttons
A chat message comes to you with two buttons: approve or cancel.
Only a yes lets it go ahead
Only an explicit approval lets it go ahead. If there is no answer within 30 minutes, nothing happens and the item stays on the open-items ledger.
Urgency is never a reason to skip this step. "It looks urgent" is exactly when a person should look.
Approved mail is sent from a separate work account. Payments and orders you carry out yourself; the agents never move money.
- Live in pilot approvals run through chat buttons
- Planned a web approval screen and KakaoTalk
Built on Claude
Today
- The agents run on Anthropic's Claude models. A fallback model is configured for availability only; the privacy policy names its provider.
- Agents, tools and scheduled automations are developed with Claude Code.
- The pilot runs on subscription-based Claude tooling.
Planned next
- Move to the Claude API with prompt caching and model routing (lighter tasks on smaller models), so the service can be offered to other small businesses at a predictable cost
- Measure model cost per customer, then finalize pricing
- An isolated workspace per customer, with accounts, storage and logs separated
한국어 요약
총괄 비서 에이전트가 대표님과 대화하고 경리·
- 하루 흐름: 아침 브리핑, 하루 두 번 메일·
문자 확인, 거래내역 파일이 들어올 때마다 장부 갱신과 대조, 마감 전 반복 알림, 주 1회 점검
확인 규칙
- 결제·
연체· 해지 같은 상태는 알림 문자 대신 장부 원장으로 확인 - 다른 에이전트의 반박은 근거와 대조한 뒤에만 반영
- 검색 결과가 0건이면 출처가 최신인지부터 확인하고 확인할 수 없으면 '모름'으로 보고
- 세율·
기한은 법령 원문과 조문별 시행일로 확인
승인
밖으로 나가는 일은 채팅의 승인·
- 승인 채널: 지금은 채팅 버튼. 웹 승인 화면과 카카오톡은 개발 예정
- 사용 모델: Anthropic Claude 모델(가용성을 위한 예비 모델만 따로 설정). 개발 도구는 Claude Code
- 다음 단계(예정): 프롬프트 캐싱과 모델 라우팅을 쓰는 Claude API로 이전, 고객별 모델 비용 측정 후 가격 확정, 고객별 격리 작업 공간