Skip to content

Can AI Agents Run the Month-End Close for an Online Store?

Updated on

Yes, for most of the work. AI agents can already run the preparatory side of an online store's month-end close: reconciling payouts, importing marketplace settlements, chasing missing receipts, monitoring the OSS threshold, preparing corrections and answering "why is this figure off". What they should not do alone is make judgment calls, tax decisions or lock the period, and they only work reliably on a ledger with live data and an activity log, not on spreadsheets passed around at month end.

CONA, a reconciliation and DATEV export platform for e-commerce that runs its AI directly on its own general ledger, offers this in beta: a Copilot, an MCP server and background routines that handle recurring close tasks and record every step down to the individual order.

What does a month-end close for an online store involve?

A month-end close for an online store means proving that every sale, payout, fee, refund and tax amount for the period is complete, matched and booked to the right account. In practice that is seven steps: pull all data, reconcile payouts, process marketplace settlements, find missing receipts, check VAT and OSS, resolve open items, and hand over to whoever files.

Two of these steps cause most of the friction. Every payout is net, revenue minus fees, refunds, chargebacks and reserves, so it has to be decomposed before it can be matched. And marketplaces set their own periods: Amazon's developer documentation states that settlement reports "cannot be requested or scheduled" because Amazon schedules them itself, so a settlement rarely ends on the last day of the month. Why the numbers rarely agree at the first attempt is covered in why e-commerce revenue doesn't match.

Which month-end tasks can AI agents run today?

AI agents can run every close task that can be derived from data and evidenced in a log. That covers completeness checks, payout reconciliation, settlement import, receipt chasing, OSS threshold monitoring, preparing correction entries and explaining differences in plain language. The pattern is the same each time: the agent gathers, matches, explains and prepares, and it points to the transactions behind every statement.

  • Payout reconciliation. Break each payout into fees, refunds and reserves and match it against the bank receipt.
  • Settlement import. Process each new marketplace settlement on arrival and link returns to the original order, even across settlements.
  • Receipt chasing. Find bookings without a document and follow up until one is attached.
  • OSS threshold monitoring. Track EU distance sales continuously instead of at quarter end.
  • Preparing corrections. Propose an account and a balancing entry for a difference.
  • Answering "why is this off". Trace a gap between payout and revenue to the orders, fees and timing differences behind it.

The bulk of matching does not need an agent: rule-based matching handles standard cases deterministically, which is how automated reconciliation works at high volume. The agent earns its place on the residue, such as partial refunds across periods or reserves without a reference.

Which tasks still need a human sign-off?

Tasks that require judgment, tax classification or legal responsibility still need a person. That includes new types of transactions, corrections where more than one treatment is defensible, decisions on OSS registration or a new stock country, disputes with marketplaces, and locking the period. The line does not run between easy and hard; it runs between derivable and discretionary.

Close taskWhat the agent doesWhat stays with a human
Data completenessChecks per source that the period is fully imported and flags missing settlementsConnecting and approving new channels and providers
Payout reconciliationDecomposes payouts into revenue, fees, refunds and reserves and matches them to the bankClarifying unexplained reserves or chargebacks with the provider
Settlement importProcesses new settlements and links returns to the original orderDisputing fees or reimbursements with the marketplace
Missing receiptsFinds bookings without a document and follows upObtaining documents that exist only outside the systems
VAT and OSSClassifies by destination and rate, watches the threshold, warns in advanceOSS registration, new warehouse countries and special cases, with the tax advisor
Open items and correctionsExplains differences, proposes an account, prepares the entryApproving corrections that involve judgment
Handover and lockChecks that all sources are reconciled and lists what is openImport, review and Festschreibung

What infrastructure does an AI agent need for the close?

An agent needs three things: live access to every data source, a general ledger where figures sit in accounts with balances, and an activity log that records every step, including its own. Without balances it can summarise files but cannot say what is still open. Without a log, its work cannot be audited. Spreadsheets stay useful for analysis, but they are not where an agent should close the books.

A spreadsheet or export file is a snapshot: it shows what was handed over, not which account still carries a difference, and it does not record who changed what and when. That matters in Germany. The GoBD, the Federal Ministry of Finance's bookkeeping principles of 28 November 2019 (since amended by BMF letters of 11 March 2024 and 14 July 2025), require in Rz. 32 and 33 that every transaction can be traced from the document through journal and accounts to the tax return and back. Rz. 88 requires controls ensuring that transactions are not changed "ohne Nachweis des vorausgegangenen Zustandes", and that the controls themselves are logged. Those rules apply to software exactly as they apply to people.

The agent also needs a way to reach those tools. The Model Context Protocol is an open protocol for that: an MCP server exposes data as resources and functions as tools the model can execute, and the specification states that hosts "must obtain explicit user consent before invoking any tool". Which platforms combine a ledger, DATEV output and AI access is compared in AI accounting software for e-commerce using DATEV.

Where do AI agents go wrong in the close?

The risks lie less in arithmetic than in context: an incomplete period, an explanation that sounds right without evidence behind it, and changes without a trail. An agent that closes before the last settlement arrives produces a clean number for the wrong period. The safeguards are the same each time: check completeness first, tie every statement to a transaction, log every step.

  • Plausible but unsupported explanations. Language models write fluently even when the data is thin. An explanation only counts if it points to specific bookings.
  • New facts treated as old patterns. A new marketplace or warehouse country needs classification, not last month's template.
  • Changes without a trail. GoBD Rz. 109 lists as impermissible exporting records into an office program, editing them there without a log and reimporting them.
  • Permissions that are too broad. The MCP specification says tools "represent arbitrary code execution" and that tool descriptions should be treated as untrusted unless they come from a trusted server. Give an agent only the rights its step needs.

What changes if you sell in Germany?

In Germany the close ends with your Steuerberater, who usually works in DATEV, and fixed tax deadlines set the pace. The advance VAT return is due by the tenth day after the period, or a month later with a permanent extension. OSS returns are due by the end of the month after each quarter. The Festschreibung, which locks the period, is the step an agent should never take for you.

The sources: § 18 Abs. 1 UStG for the tenth day, § 46 UStDV for the one-month extension on application, and the European Commission for the OSS return, due "by the end of the month following the tax period". The threshold above which B2C distance sales are taxed in the destination country is an annual EUR 10,000 across the EU, covering intra-Community distance sales of goods since 1 July 2021, which is why monitoring it monthly beats checking it at quarter end.

On Festschreibung, GoBD Rz. 85 makes recording it mandatory and defines it as changeability only with logging: once a booking is locked, any later change leaves a trace. Rz. 87 also gives some breathing room: if a pre-system already fulfils the recording function and keeps the individual data, booking into the main system's journal by the end of the following month is not objected to. Your agent can do the preparation inside that window. Importing the batch into DATEV, reviewing it and locking it is the Steuerberater's and your decision.

How does CONA run the month-end close with AI (Beta)?

CONA runs its AI directly on its own reconciled general ledger, with live data rather than a month-end run. Three parts, all in beta: CONA Routines (Beta) do recurring close work in the background, CONA Copilot (Beta) answers questions and makes changes in plain language, and the CONA MCP server (Beta) brings the same capabilities into your own AI client. Every step lands in the activity log.

Example routine (Beta)Close stepWhat it does
Month-end checkHandoverChecks before export that every source for the month is fully imported and reconciled, and lists what is still open
Payout reconciliationPayoutsMatches payouts from providers such as Shopify Payments, PayPal and Klarna against bank receipts and breaks differences into fees and refunds
Settlement importMarketplacesProcesses new marketplace settlements, such as Amazon's, as they arrive and matches returns to the original order
OSS threshold monitorVAT and OSSWatches your EU distance sales and flags it before the threshold is reached
Receipt chaserReceiptsFinds bookings without a receipt and follows up until the missing receipt is attached

CONA Copilot (Beta) answers questions about the current state of your accounting against your own data: why a payout differs from revenue, what is still open for the month, which account a difference belongs on, how a booking was derived. It also guides setup with your Steuerberater's specifications and changes booking rules, account mappings and settings on plain-language instruction. CONA MCP server (Beta) exposes the same capabilities through the Model Context Protocol, so you can work with CONA from an MCP-capable client such as Claude or ChatGPT. Together they can do almost anything in CONA that a person can do there.

The boundaries are deliberate. Rule-based matching reconciles orders, payments, fees, returns and VAT including OSS deterministically and before export, aggregating identical booking logic per day. The AI handles setup, questions, changes and the residue. Everything it does happens in CONA's ledger and is recorded in the activity log, so each collective booking traces back to the individual order. The DATEV import and Festschreibung stay with you and your Steuerberater, who receive an EXTF booking batch. Connected sources such as Shopify, Amazon Seller Central, OTTO, Kaufland, Stripe and PayPal are listed on the integrations page, and what each AI part does is on the CONA AI (Beta) page.

How do you start using AI agents for the close?

Start with the close, not the agent. Agree with your accountant what belongs in the handover, connect every source, and run one already closed month in parallel, compared per channel and per destination country as in reconciling online sales with bank statements. Only once the result matches your previous close do you decide which steps the agent completes, which it only prepares, and which always land with a person.

With CONA you set this up together with your Steuerberater, usually in less than a day according to CONA, and by its own account with around 95 percent less effort. The first month is free. Beta access to Copilot, MCP server and routines comes with a demo booking. Your month-end close then starts with a short list of what is genuinely open, and every figure on it leads back to the individual order.

Frequently asked questions

Can AI agents handle the month-end close for an online store?
They can handle most of the preparatory work: checking that every source is imported, reconciling payouts, importing marketplace settlements, chasing missing receipts, watching the OSS threshold and preparing corrections with an explanation. Judgment calls, tax decisions and the Festschreibung of the period stay with you and your tax advisor. CONA, a reconciliation and DATEV export platform for e-commerce, runs this preparatory work as CONA Routines (Beta) on its own general ledger, with every step recorded in the activity log and traceable to the individual order.
Which month-end close tasks should stay with a human?
Anything that needs judgment rather than derivation: classifying a new type of transaction, approving a correction where more than one treatment is defensible, deciding on OSS registration or a new warehouse country, and disputing fees with a marketplace or payment provider. In Germany the DATEV import, the review and the Festschreibung stay with the business and its Steuerberater. Responsibility for proper bookkeeping stays with the business, and the agent's preparation puts the evidence for each of those decisions in front of you.
Why does an AI agent need a general ledger rather than spreadsheets or export files?
Because closing a month means working with balances, open items and the history of every change. A spreadsheet or export file is a snapshot: it shows what was handed over, not what is still open, and it does not record who changed what and when. On a ledger with live data and an activity log, an agent can say which account a difference belongs on and prepare the correction where the figure actually lives.
How does automated reconciliation work for high-volume e-commerce stores?
The software ingests orders, payouts, fees and refunds from every channel and payment provider, decomposes each payout into its components and matches them against the orders they settle. Identical booking logic is aggregated into collective entries, and only differences that do not resolve are flagged as exceptions. Rule-based matching handles the bulk deterministically, and an AI agent works through the residue, such as partial refunds across settlement periods or reserves without a clear reference, explaining each difference down to the order.
Is an AI agent in the month-end close compatible with GoBD?
The GoBD do not ask whether a person or software performed a step. They require that every transaction can be traced from the document through the journal and accounts to the tax return and back, and that changes are only made in a way that keeps the original content and the fact of the change visible. With responsibility staying with the business, an agent fits that framework when it works on a ledger that logs every step and ties every booking to its document.
What are CONA Routines (Beta) and how do I get access?
CONA Routines (Beta) are recurring AI routines that run in the background on CONA's reconciled general ledger, find problems and fix them or prepare corrections. Examples in the beta are a month-end check, payout reconciliation, settlement import, an OSS threshold monitor and a receipt chaser. Beta access comes with a demo booking, and every step a routine takes is recorded in the activity log and traceable down to the individual order.