AI automation for back-office work

You get one process that runs to the end on its own, and one clearly marked place where a person steps in. Back-office work repeats in the same few places, and these are the ones we build on:

• Supplier invoice PDFs extracted and posted as purchase invoices in the ERP.
• Bank payment notification emails read and reconciled against open invoices.
• Sales orders synced from the ERP into Excel or OneDrive.
• Delivery planning generated from sales-order delivery dates.
• Monthly file creation, maintenance and archiving.
• An AI phone agent that answers calls and passes on what needs a person.
• Automatic quote pricing from the inputs a customer supplies.
• An MCP interface so AI agents can talk to your systems directly.

Every workflow ships with the same three parts: the connection to the system, a review queue for exceptions, and a log showing what the machine did and who approved it. Without those three you have a demo rather than something a finance team will sign off on. A full example, awkward documents included, is written up in the supplier invoice workflow.

Source code, infrastructure and rights stay with you, and the workflow is not tied to a single AI vendor, so you can move to a cheaper or more accurate model when one appears.

The review queue is the most important part of this service. The honest answer to what happens when the model reads a document wrong is not that it will not happen. It is that confident cases post automatically, and uncertain ones land in a queue where a person sees the original document and the model's best guess side by side, then approves or corrects it.

Before anything posts, deterministic checks run. They use no model, so their verdict is reproducible. On an invoice there are three: line items plus VAT must equal the stated total to the cent, the supplier and invoice number pair must not already exist, and the currency and tax rate must be ones the company actually uses. Anything that fails a check never reaches the ledger.

Four decisions get made during the build: where the confidence threshold sits, who works the queue, what happens to a document nobody reviews, and what the audit log has to contain. Designing that queue is a build item in its own right, and a quote that never mentions it is pricing only the path where nothing goes wrong.

One buyer put the requirement in the same sentence as the goal: cut manual data entry while keeping exception handling and manual review wherever the information cannot be reliably identified. That second half is the engineering problem, and it is a sharper specification than most briefs contain.

You almost never have to replace anything. Exact Online, Merit and e-conomic all expose APIs, as do most mainstream cloud accounting and ERP platforms, and the automation is built on top of what you already run. A migration is a far larger project than the automation it would supposedly enable, and it is rarely the cheaper route.

Where there is no API, three routes remain: scheduled file exports, browser automation, or reading data out of screens. All three depend on an interface the vendor can change without notice, so that fragility stays on the books as maintenance for as long as the workflow runs. We say which route applies before quoting, not after.

Two things usually get built alongside. An MCP interface makes your systems callable by AI agents without the agent having to guess the format. We run one ourselves at /mcp, and our lead endpoint answers a GET request with its own schema, so an agent that finds the URL learns the fields instead of guessing them.

When data cannot leave the building, the model runs on your own server and documents never leave the internal network. The same build pattern runs our own business, aiasemu.ee, where a customer types an address, AI measures the lawn from a satellite photo and prices the job in seconds instead of a site visit, and an AI phone agent takes the calls.

Automation is quoted per project, and one well-defined workflow is normally a one to three week build. Four things set the number: how many systems have to talk to each other, whether those systems expose an API, how deep the exception handling has to go, and how many documents pass through each month. What drives an automation price works through each of the four in detail.

We publish fixed prices for the work that is genuinely repeatable: 290€ for an AI-readiness audit, 990€ to implement the fixes on an existing site, from 3 999€ per project for a new website, and 590€ a month for a visibility retainer. Automation stays in the quoted column because the price depends on software we have not opened yet. A firm number offered before that review is either padded to cover the risk, or it becomes a change request later.

Quotes come in stages, each priced separately, with the first stage small enough that you can stop after it. Running cost is given as a formula rather than a monthly lump sum: cost per processed document times expected monthly volume, plus hosting, plus an agreed maintenance arrangement. That form survives a change in volume and lets you check the bill against reality.

Start with one process. A five-workflow quote prices five unknowns at once, and every unknown carries a risk margin.

Three things to start, and nothing beyond them.

• The list of processes in your own words, with who performs each step today.
• System names at product level: which ERP, which accounting platform, and where the files live.
• Sample documents, the awkward ones included: two-page invoices, a credit note, a supplier whose name does not match the ledger, an unexpected currency.

The awkward samples matter most. Pulling a total off a clean invoice is an afternoon of work. The budget goes on the exceptions, and without real examples an estimate is a wish rather than a number.

You also need one person who is allowed to decide where the confidence threshold sits and who works the review queue. That is a business decision rather than a technical one. If the accountant does not trust the system, they will re-check every automatic entry anyway and the saving disappears.

Access comes after the quote is accepted, and only to the systems the workflow actually touches. Scoping runs on documents and system names first, which is also why we can tell you early when a process is not worth automating at all.

Three cases where we say do not start, instead of writing a quote.

The process runs too rarely. Payback comes from hours saved, so a half-hour task performed twice a month will not repay a build, however well the build is done. Write the steps down and keep doing it by hand.

The process is too variable. If every case is decided on its own and nobody can state the rule in words, there is nothing to hand the machine. A model can read a document. It cannot invent a rule the company does not already have.

Nobody owns the process. Automation makes decisions on someone's behalf, and if that someone does not exist, the review queue sits unread and the exceptions pile up quietly.

Two things we will not promise in either direction: that the model will never be wrong, and a specific saving before the first workflow has run long enough to show its real exception rate. If you are not sure which of your processes fall on which side of that line, send the list with the system names and we will tell you which parts we would leave alone.