How Much Does AI Business Automation Cost?

One well-defined back-office workflow is normally a one to three week build, quoted per project. That single process is the unit worth pricing. It has a start, an end, and a defined path for what happens when something goes wrong. Anything larger is the same unit repeated.

We publish fixed prices for the parts of our work that are genuinely repeatable and quote the rest per project. Most agencies in our market answer the cost question with a contact form instead. That habit also costs them mentions when someone asks an assistant to compare suppliers, a problem we covered in why ChatGPT does not recommend your business.

Automation sits in the quoted column for a reason. A website build has a knowable shape before anyone starts work. An automation project depends on systems we have not opened: which of them hold the data, whether they will hand it over through an API, and how often the incoming documents are messy enough to need a human eye.

Four factors set the shape of an automation quote. Everything else is detail.

The number of systems. Every system in the chain needs its own connection, its own credentials, its own rate limits and its own error handling. Moving data between two systems is one bridge, and a single bridge with no judgement in it is often cheaper on an off-the-shelf tool than in a custom build, a trade-off we work through in choosing between Zapier, Make and a custom AI build. Adding a third bridge rarely adds a third of the work, because the new system also has to agree with the first two about what a customer, an invoice and a date are.

Whether those systems expose an API. This is the single largest swing factor and it is covered in the next section.

How deep the exception handling goes. A workflow that stops and asks a person whenever it is unsure is cheaper and safer than one that tries to be right on every document. How far you push that dial is a budget decision, and it should be an explicit one.

Volume. Volume sets the running cost rather than the build cost. Ten supplier invoices a month and ten thousand need the same logic, but different infrastructure, different batching and a very different model bill. Volume also decides whether the project is worth starting. The payback comes from hours saved, so a process that runs rarely will not repay the build, however well it is built.

A documented API turns an integration into a known quantity, and its absence turns the same integration into an open-ended one. With an API, the work is reading the documentation, handling authentication, mapping fields and testing the error cases. Without one, the options are scheduled file exports, browser automation or reading data out of screens. Each of those three depends on an interface the vendor is free to change without notice, so the fragility stays on the books as a maintenance line item for as long as the workflow runs.

You almost never have to replace the system to automate around it. Merit and e-conomic expose APIs, as do most mainstream cloud accounting and ERP platforms, so 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. We go through the common shapes of this work in AI back-office automation for SMEs.

You can run the same test on your own systems. A system that is cheap to automate describes itself. Our lead endpoint at /api/lead answers a GET request with a schema, so an agent that finds the URL learns the field format without guessing, then posts the enquiry. We publish an MCP server at /mcp for the same reason. When a system tells you how to talk to it, the integration estimate drops.

The happy path is the cheap part. Extracting the total from a clean supplier invoice and posting it to the ledger is a demo you can build in an afternoon, and it is the smallest piece of a supplier invoice workflow that holds up in production. The budget goes into the invoices with two pages, a credit note stapled to the front, a supplier name that does not match the one in the ledger, or a currency the template did not expect.

One buyer who came to us about back-office automation put the condition in the same sentence as the goal: cut the manual data entry, but keep 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.

That review queue is a build item in its own right: designing the review queue means setting a confidence threshold, building the queue and the interface, keeping an audit log and deciding what happens to a document nobody reviews. It is worth checking whether a low quote covers any of that, because the happy path on its own is far cheaper to build. The gap shows up at go-live, when finance declines to sign off on a process that can post a wrong number with nobody watching.

A single automated process gives you a measured result before you commit to the next one. Buyers usually arrive with a list rather than a process. A typical one, from a real enquiry: supplier invoice PDFs read and posted as purchase invoices in the ERP, bank payment notification emails reconciled against open invoices, sales orders synced from the ERP into a spreadsheet, delivery planning generated from sales-order delivery dates, and monthly file creation and archiving.

Quoted as one project, that list prices five unknowns at once, and every unknown carries a risk margin. Built one at a time, the first workflow pays for the discovery: afterwards you know how that ERP behaves under load, how clean the documents really are and what share of them lands in the review queue. The remaining four get estimated against evidence instead of padding.

We hold ourselves to the same test. On aiasemu.ee, our own lawn-care business run end to end by AI, a customer types an address and the system measures the lawn from a satellite photo and prices the job in seconds instead of a site visit. One week of PostHog data: 57 address searches, 39 refined estimates, 9 bookings started, 2 completed, roughly 22 leads a week. That funnel produces enquiries and loses people at the booking step, so our internal conclusion was to treat the site as a lead source rather than a booking system. Measuring one workflow is what tells you where the following build hour belongs.

A quote you can hold a vendor to breaks the work into stages and prices each one. Use this as a buyer checklist, and ask for whatever is missing before you sign:

• Stage-based fixed pricing, plus a total range for the whole build.
• A timeline, and the team with their roles.
• An hourly rate for work outside the agreed scope.
• Infrastructure and maintenance cost estimates.
• Third-party API cost estimates, stated separately from the build fee.
• The technical risks, written down rather than discovered later.

The same buyer required that source code, infrastructure and rights stay with them, with no lock-in to a single AI vendor. Both conditions are worth copying. Ownership decides whether you can change supplier without rebuilding. Model independence decides whether you can move to a cheaper or better model when one appears, and a workflow welded to a single provider cannot follow the price down.

Buyers with a procurement function ask for all of this as a matter of course. Smaller companies often do not, and the gaps surface later as change requests.

Running cost has three components, and only one of them scales with your volume. Hosting is usually the smallest and the most predictable. Third-party model and API usage scales directly with how many documents, emails or calls pass through, so a quote should give you a per-unit figure rather than a monthly lump sum you cannot check. Maintenance depends on how the integration was built. API-based connections mostly keep working. File exports and browser automation need attention whenever a vendor changes something.

Ask for the running cost as a formula rather than a single number: cost per processed document multiplied by 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. It also keeps the payback calculation honest, since you can set hours saved per month against running cost per month with the build amortised over a period you choose.

For comparison, the one recurring price we publish is the 590€ a month visibility retainer, which covers ongoing work plus a monthly report on how often AI assistants recommend you. Automation maintenance is quoted per project, because a workflow that touches one API and a workflow that touches four are not the same commitment.

Our scoping runs in that order. Send the list of processes, name the systems by product name, and share a handful of representative documents including the awkward ones. We check API availability, look at how variable the documents actually are, then quote in stages with the first stage small enough that you can stop after it.

One pattern is worth naming. The operations buyer who contacted us about back-office automation opened by asking whether integrating their accounting platform through its API was inside our expertise at all. Price came later. Scope anxiety arrives before price anxiety, and a vendor who leads with a number is answering the second question while the buyer is still asking the first.

If your question is about your website rather than your back office, start with the free readiness check. It runs eight deterministic checks with no model scoring, so the result is reproducible. If your question is about a process, send us the list and we will tell you which parts are worth automating.