Every integration is sold with the same three lines. An order lands in your shop, an invoice appears in your accounting system, and the stock count drops. Nobody sells you the fourth line, which is the one where your accountant spends the first week of every month correcting invoices by hand.
So this is a buyer's guide to the fourth line. What to buy, what to commission, where the sync gives way, and what your customers have been allowed to demand of you since July 2025. Most of the work is plumbing. Very little of it needs a language model, and the parts that do are not the parts vendors point at.
What a connector promises and what it delivers
A connector between a shop platform and an accounting system does one honest job well: it copies the clean orders. A customer pays in full, in one currency, for goods in stock, with no discount code and no later change of mind. That order becomes an invoice with the right lines and the right total, and nobody has to look at it again.
The trouble is that clean orders are the ones you were never going to spend time on. What costs you the first week of the month is the tail: the refund that came back partial, the exchange that went out as a new order, the business customer who needs their registry code on the document, the order that paid in two instalments. A connector either has a rule for each of those or it does not, and the demo never shows you which.
Ask any vendor for the list of cases their connector does not handle. A good one has the list ready and will read it to you. That single question sorts the market faster than any feature grid.
The four systems in the chain
Before anyone quotes you, write down what is in the chain. Most Estonian shops have four things, and the number of connections you need grows with the number of boxes, not with the number of orders.
- The shop platform. WooCommerce, Shopify, Magento or something bespoke. It owns the order and the customer.
- The accounting system. Merit Aktiva, Directo, SmartAccounts or the Business Register's own e-arveldaja. It owns the invoice and the ledger.
- The stock or warehouse system. Sometimes the accounting system, sometimes Erply or a warehouse tool, sometimes a spreadsheet that one person keeps.
- The e-invoice operator. The service that carries a machine-readable invoice from you to a buyer who asked for one.
Two boxes is a connector. Four boxes is a pipeline, and a pipeline needs someone to own it. That is the honest split, and it is the same split we describe on the automation service page.
Buy the connector, commission only the gap
If a ready connector exists for your two main boxes, buy it. The vendor maintains it against both APIs, absorbs the breaking changes, and charges less per month than one hour of anybody's development time. A commissioned copy of that same connector costs more, does less on day one, and then costs again every time either side ships an update.
What is worth commissioning is the gap the connector leaves. That is the system it has never heard of, the rule that only applies to your business, and the alert that tells you when the pipeline stopped. Nobody sells those, because they are different in every shop. We build them on n8n or Make depending on where the data has to live, and the choice is a hosting question before it is a features question.
The same logic applies to the tools around the shop, and we have written the general version of this argument in our guide to automation for ecommerce. Buy the solved thing. Commission the joins.
Where the sync gives way
Five places, in rough order of how often we find them.
Rounding and VAT
A shop usually computes an order total from prices that already include tax. An accounting system usually computes it from net lines with tax applied on top. On most orders the two agree. On a small share of orders they land a cent apart, and a cent is enough for an accountant to stop and look. Decide which system is the authority for the total before anyone writes code, and make the other one accept it.
Refunds, partial refunds and exchanges
A full refund is easy. A partial refund is a credit note with lines that no longer match the original invoice, and an exchange is often two documents that have to reference each other. Connectors vary wildly here, and this is where most of the manual correction lives.
Stock held in two places
If the shop and the warehouse system both believe they own the stock number, one of them will be wrong within a week. Pick the authority once. Everything else reads from it.
Customer identity
A business buyer needs a registry code and often a specific billing entity that is not the person who clicked buy. If your checkout does not collect it, the invoice cannot carry it, and no amount of clever mapping downstream will invent it.
Silence
The worst failure is the one that does not tell you. A token expires, a queue stalls, and orders keep arriving in the shop while nothing lands in accounting. Nobody notices until month end. Any pipeline we build has to say when it stopped, in a channel a human already reads.
The Estonian paperwork you cannot skip
Since 1 July 2025 the rules on e-invoices changed in a way that lands directly on shops. The Ministry of Finance describes the amendment plainly: the right to demand an e-invoice from a seller is given to every accounting-obliged entity that has publicly registered itself in the e-Business Register as an e-invoice recipient.
Read that carefully, because it is a right your buyer holds, not a blanket duty on every seller. What it means in practice is that a business customer can now ask, and you need a route that produces one. If your shop mails a PDF and calls it an e-invoice, you do not have that route. A PDF is a picture of an invoice. An e-invoice is a structured file another system can read without a human retyping it.
Tax filing sits downstream of the same pipeline. If orders reach the ledger late or wrong, the VAT return is late or wrong, and the deadlines published by the Estonian Tax and Customs Board do not move for anyone. That is the actual argument for fixing the sync, and it has nothing to do with technology being fashionable.
Stock is a different problem from invoices
People buy one integration and expect both. They are not the same job. Invoice sync runs once per order and can tolerate a few minutes of delay. Stock sync has to be close to live or you oversell, and overselling costs a refund plus a customer.
If you sell the same goods in a shop, a physical store and a marketplace, stock is the harder half by a wide margin, and it deserves its own decision. Get the authority right, decide what happens when a count goes negative, and accept that some of it will be a nightly reconciliation instead of a live feed. A nightly job that is always correct beats a live feed that is sometimes wrong.
What to ask before you commission anything
- Which cases does your connector not handle? Ask for the list in writing.
- Which system is the authority for the order total, and which for the stock count?
- What happens to a partial refund, an exchange and a two-instalment payment?
- How do I find out within an hour that the sync has stopped?
- Can this produce a structured e-invoice when a buyer asks for one?
- Who fixes it in month three, and what does that cost?
If the answers are vague, the correction work has not disappeared. It has moved onto your accountant's desk, where it is harder to see and easier to keep paying for. Our guide to what business automation costs works through the same question from the budget side.
What we would not automate
We would not let software decide a VAT treatment it is unsure about, because a wrong rate repeated across a year is a correction filing and not a bug report. We would not automate a credit note without a person approving it, because a credit note is money leaving. And we would not put a language model between an order and the ledger. Those records are deterministic, and a rule that fails loudly is worth more than a model that guesses quietly.
Language models do earn their place beside the pipeline. They handle the messages customers send about orders, and the calls asking where a parcel is, which is the same argument we make for accounting firms. That is service work, not bookkeeping, and the distinction is worth keeping.
Our own receipts here are integration and web receipts, not shop receipts. On a Tartu transport operator we rebuilt the site, took structured data from 4 items to 75, cut the hero image from 60KB to 14KB, published an llms.txt with a named AI crawler allowlist, and wired the back office with Make and n8n workflows. The four case studies we publish say what each build did not prove, which is the part most agencies leave out.
Questions people ask about ecommerce accounting integration
Should I build my own ecommerce accounting integration?
Only for the part nobody sells. If a ready connector exists between your shop platform and your accounting system, buy it, because the vendor maintains it against both APIs and you do not. Commission the wiring around it: the systems the connector has never heard of, the rules that are specific to your business, and the alerts that tell you when a sync stops.
Why do the invoices arrive wrong?
Almost always rounding, VAT treatment or partial refunds. A shop calculates the total from line prices including tax, an accounting system calculates it from net lines plus tax, and the two disagree by a cent on a fraction of orders. Partial refunds and exchanges then create documents the connector has no rule for. Ask for the failing orders as a list before anyone writes code.
Do I have to send e-invoices in Estonia?
You have to send one when the buyer asks and is entitled to. Since 1 July 2025 the Accounting Act gives every accounting-obliged entity that has publicly registered itself in the e-Business Register as an e-invoice recipient the right to demand an e-invoice from the seller. That is a right held by your buyer, not a blanket duty on every seller, so what it means in practice is that your shop needs a route to produce one.
Does any of this need AI?
The sync does not. Orders, invoices and stock counts are deterministic records, and a language model guessing at them is a downgrade from a rule that either holds or fails loudly. Language models earn their place next to the pipeline, on the messages people send about orders, and on the phone calls asking where a parcel is. Keep them out of the ledger. Anything that talks to a customer also carries a disclosure duty under the EU AI Act.
What does it cost to wire a shop to accounting software?
The connector itself is a subscription you pay the vendor directly. Our audit is free, takes about 30 minutes and ends with a written list of your leaks ranked by monthly cost. Build and run starts at 600 euros per month plus VAT, quoted after the audit, because the honest number depends on order volume, how many systems are in the chain and how odd your rules are. The wider market publishes 150 to 350 dollars per hour for senior independent talent and 20,000 to 150,000 dollars for a scoped build.
The shop is the part of this business that already works. The joins between it and everything downstream are where orders quietly turn into corrections. If you want that chain mapped for your own shop instead of a generic checklist, book the free audit. Thirty minutes, your order data, and a straight answer on whether the work is worth doing at all.