# How we run the month-end close Each month our Machine finance agent works the close checklist, ties Stripe revenue to the Google Sheets ledger, flags odd numbers and missing docs, and posts open items to Slack. A person closes the books. Canonical page: https://themachine.myapps.ai/use-cases/month-end-close Closing the month is the same routine every time, and it still eats several days. There's a checklist with a few dozen lines. Revenue has to match the ledger. Missing receipts and contracts need chasing. Then someone eyeballs the totals against recent months before anything gets signed. Not one of those steps is difficult. Together they crawl, and most of the effort goes into proving that the boring 90% is fine so the team can get to the 10% that isn't. So an agent on Machine now takes the first pass each month. It matches our Stripe revenue to the ledger, goes down the rest of the checklist looking for missing paperwork, points out numbers that look unusual next to previous months, and posts what's still open in our finance channel on Slack. Journal entries are off limits, and so is marking the period closed. That call stays with a person. - **Team:** Machine - **Runs on:** A schedule that fires once a month - **Connected systems:** Stripe · Google Sheets · Slack - **Mode:** Gathers and flags · a person closes the books ## The problem The list of tasks isn't what slows the close down. The drag is checking the same kinds of things over and over. Does booked revenue match what Stripe really took in? Does each line have the backup document it should? Has any number shifted far enough from recent months that someone will ask why? Any one of those takes minutes. All of them, done carefully, in a week already full of other close work, is where the time disappears. A formula in the spreadsheet will catch a perfect match, but a near miss slips past it. It also has no idea what "looks off" means unless a person brings that judgment, fresh, every month. So the whole checklist gets done by hand, at full cost, month after month, and it's only as thorough as the calendar allows. ## What we built Once the period has ended, a monthly **cron** wakes a close agent on Machine. Every run gets its own fresh cloud computer, or sandbox, that can read Stripe revenue plus the checklist and ledger tabs of a Google Sheet. The agent ties Stripe revenue to what the ledger booked, goes through the remaining checklist items for missing backup, compares this month with the trailing months to spot anything unusual, and posts a close-status summary to Slack listing every open item. Its job is to gather and flag. It never books a journal line and never signs off the close. ## How it works ### Fire once a month, after the period ends A **cron** trigger is a timer, and this one goes off monthly, after the previous period has wrapped up. Each time, Machine starts a new **session**: a clean sandbox used only for that run. The checklist and the ledger sheet hold the history, so the agent doesn't need to remember anything from last month. ### Write down what 'done right' means The agent's **skills** and **memory** are its playbook, and it reads them before looking at a single number. They spell out when a line counts as matched, how much slack a timing lag is allowed, which document each checklist item needs, and how large a jump against the trailing months earns a flag. The same yardstick applies every month. ### Connect the close, read and flag only Every tool is reached through a **connector**, a narrow link whose keys stay on our servers, so the model never sees a raw token. That lets the agent: - **Look at Stripe.** Revenue, refunds and payouts for the period just closed. - **Look at the ledger tab.** What got booked for that period, to compare against Stripe. - **Look at and mark the checklist tab.** It sees where every close item stands and adds its own notes and flags there. The ledger's booked entries are never touched. - **Message Slack.** One close-status summary with every open item. ### Set the limits Matching and flagging is the whole job. Closing is not. The agent will not write a journal entry, change an amount already in the ledger, or tick the period off as closed. That decision, and anything that follows from it, is a person's. Credentials are stored encrypted in the **Secrets Manager** and only handed to the agents you grant them to. ### Post what's still open Once that's all wired up, each month brings a fully worked checklist and a single Slack message. It lists what still doesn't reconcile, which items lack paperwork, and which numbers moved far enough from the trend to deserve a look. Each entry carries enough detail that nobody has to pull the figures again to act on it. Finance reads the list, then closes the books themselves. > **The pattern** > A **cron** once a month starts a clean session that can read Stripe and can > read and flag the checklist and ledger in Google Sheets. Matching tolerances > and anomaly rules sit in the agent's **skills** and **memory**. It gathers > and flags every open item, and a person closes the books. ## Guardrails Letting software near the company close is mostly about trust, so access is narrow and runs in one direction: - **Isolation.** Each run has a sandbox to itself and can only reach the systems it was given. The checklist flags and the Slack message are the only things it sends back out. - **Scoped secrets.** The Stripe and Sheets keys are stored encrypted in the **Secrets Manager** and only handed to the agents you grant them to, when the sandbox starts. - **No journal entries, no closing the books.** It can't add a journal line, change a booked ledger amount, or mark any checklist item or the period as closed. The agent flags. A person closes. - **Everything is code.** Checklist rules, tolerances and the permissions for each system are versioned files in the repo, changed only through a reviewed **change request**, never a dashboard toggle. ## The outcome - **Every month:** The full checklist, worked against this month's numbers - **Flag only:** The agent books no journal entries and closes no periods - **One post:** Unmatched lines, missing docs and odd numbers, all in one list The days we used to spend confirming the easy 90% now go to the short list the agent actually raised. The first pass happens on schedule whether or not anyone has found time for it, and when finance sits down to close, the open items are already waiting in Slack.