Skip to content
Gathena
da

Learn · Guide · DuffBeauty's platform

How the accounting agent is built, step by step

This page follows the data from the source systems to the bookkeeper. At each step it shows what runs, why we chose it, and what keeps it safe.

← The case: an accounting agent on Google Cloud

Runs
every night
Last run
about 10 minutes
Cost per run
about $0.11

Measured on 28 October 2026 on one nightly run; the cost is at Google Cloud list prices.

  1. Step 1

    What comes in

    Every night the platform reads the inventory and ERP system, the webshops, the accounting system, the card payments, the bank and the day's exchange rates. A pay-later source is built and waits to join the nightly run.

    Why this choice

    Each system has one shared client, so every pipeline talks to a source the same way, and a change in a source's API is fixed in one place.

    What keeps it safe

    Access to the sources is read-only, and the code enforces it: a pipeline can never write back into a system of record. When a source changes, the pipeline reports it instead of guessing.

    • Inventory · Rackbeat

      stock, orders, invoices

    • Webshops · Shopify

      orders, payouts

    • Accounting · e-conomic

      ledger, drafts, documents

    • Card payments · AltaPay

      card settlements

    • Bank

      statement export via a drive

    • Exchange rates · ECB

      daily rates

    • Pay later · Klarna

      built · not nightly yet

    pulled every night

  2. Step 2

    The nightly run

    A scheduler starts one workflow each night. Small jobs pull each source and load it. Then the transform step rebuilds the tables and runs its checks. The daily mail goes out only when every check has passed.

    Why this choice

    It is the smallest set of services that runs a nightly pipeline reliably, with no orchestrator to operate.

    What keeps it safe

    The checks stand between the data and anyone who reads it. A failed check stops the mail, and every run leaves a receipt.

    Scheduler · Cloud Scheduler

    starts the run at night

    starts

    Extract and load

    small jobs · Cloud Run · raw files kept as evidence

    then

    Transform · Dataform

    tables, views and checks

    only if the checks pass

    Daily mail job

    reads one prepared view

    loaded and copied

  3. Step 3

    Where the data lives

    Every raw response is copied to a file store exactly as it arrived. The warehouse keeps the data in layers: raw as received, staging cleaned and typed, intermediate with the business rules and matches, and core and marts with the tables finance reports and books from. A separate ops area holds the run receipts and the audit log.

    Why this choice

    Every finance-ready figure traces back through the layers to the raw file it came from.

    What keeps it safe

    The raw files are the evidence: when a figure is questioned, the file it came from is still there. All of it stays in the EU, in the company's own Google Cloud project.

    File store · Cloud Storage

    every raw response, kept

    Warehouse · BigQuery, EU

    raw

    as received

    staging

    cleaned and typed

    intermediate

    business rules, matches, checks

    core · marts

    tables for reporting and bookkeeping

    ops

    run receipts · audit log

    read through the ways in

  4. Step 4

    The ways in

    Nobody reads the warehouse directly. People and agents come in one of three ways: a data connection for questions, a drafts connection for the bookkeeping agent, and the daily mail. Both connections are MCP servers. An MCP server is a program that gives an AI assistant a defined way to read your data and, where allowed, act in a system. They run in the company's own cloud.

    Why this choice

    Each person signs in as themselves, so access is per person and can be withdrawn in one place.

    What keeps it safe

    Each role gets a tier with its own datasets and limits. Every call is logged, and if the log entry cannot be written, no answer goes out. The drafts connection can only write drafts. It can never book.

    • Data connection · MCP

      own cloud · Cloud Run

      • sign-in per person
      • a tier per role, with limits
      • query and export
      • every call logged

    • Drafts connection · MCP

      • writes drafts only
      • never books
      • every call logged

    • Daily mail

      webshop vs inventory gaps

    used by

  5. Step 5

    Who does what

    The operator runs and changes the platform. The finance team asks questions in their AI assistant. The bookkeeping agent prepares entries and writes them as drafts into the accounting system. The bookkeeper reviews the drafts and books them in the accounting system they already use.

    Why this choice

    Nothing is posted until a person approves it, in the place where they already work.

    What keeps it safe

    The only orange line on the map is the one that writes into a business system, and a person stands at the end of it. Each role's card shows what it can see and do, and its limits.

    • Operator · one account

      Through
      AI tools (Claude Code, Codex) → data connection
      Can see
      all datasets, all companies
      Can do
      queryexportrun
      Limits
      a daily scan limit; source extracts capped per day

    • Finance team

      Through
      AI assistant (Claude) → data connection
      Can see
      the finance datasets
      Can do
      query
      Limits
      read only; a cap per query and per day

    • Analytics · no members yet

      Through
      the same connector → data connection
      Can see
      the analytics datasets
      Can do
      query
      Limits
      a daily scan limit

    • Bookkeeping agent · one writer account

      Through
      AI agent (Claude) → drafts connection
      Can see
      journals and drafts; prepared webshop lines
      Can do
      write draftsimport
      Limits
      a cap per call; imports run as a dry run first; can never book
      → writes drafts intoAccounting · e-conomic · drafts; read back each night

    • Bookkeeper

      Through
      the accounting system (e-conomic)
      Can see
      the agent's drafts
      Can do
      review and book
      Limits
      outside the platform

    • Finance, by mail

      Through
      a mail each morning
      Can see
      yesterday's webshop–inventory differences
      Can do
      read and follow up
      Limits
      sent only if the checks pass

  6. Step 6

    How it was built, and how it changes

    A builder agent wrote the pipelines, the checks and the connections, inside the project's own rules: a written scope before any pipeline, checks per table, and a record for every decision. The operator reviews each plan and approves it before anything changes.

    Why this choice

    The operator approves a plan instead of writing pipelines by hand, and no change reaches the cloud without that approval.

    What keeps it safe

    Code ships from Git through Cloud Build. There are no service-account keys anywhere, and each pipeline and connection runs as its own identity with only the rights it needs.

Sources

Walk-through written 28 October 2026 from the platform as built; the figures were measured the same day.

DuffBeauty has agreed to be named and has accepted the detail and the figures shown here.

If you want to see how this would fit your own systems, we can go through it together, with your IT, in-house or your IT partner.

Book a conversation