AI automation costs anywhere from a few hundred euros a month for a single off-the-shelf workflow to well into six figures for a custom multi-system agent. The price is driven by how many steps you automate, how many systems you connect, the quality of your data, and how messy the exceptions are. A simple, high-volume task like invoice data entry is inexpensive to automate; a process that touches several systems and needs human-level judgement is expensive. At Crux Digits we work in fixed-scope projects — an AI Audit at EUR 2,500, a Proof of Concept at EUR 20,000, and a Production Launch from EUR 50,000 — so you know the number before the work starts.
Why "what does AI automation cost" has no single answer
When people ask what AI automation costs, they are usually asking about two very different things at once. One is the cost of using a ready-made tool to automate a narrow task — say, an email classifier or a transcription service you pay for monthly. The other is the cost of building an automation around your own process, your own systems, and your own rules. Those two sit a long way apart on the price scale, which is why a blunt "it depends" is honest but unhelpful.
The useful question is not "what does AI cost" but "what does it cost to automate *this* process". A single, well-defined task with clean inputs is often low-cost and quick. A process that spans several departments, pulls from systems that were never designed to talk to each other, and needs a person to step in when something looks odd — that is a real engineering project, and it carries a real engineering budget.
The rest of this article walks through the things that actually move the number, gives you concrete low-cost-versus-expensive examples, explains when a subscription tool beats a custom build, and shows how to think about payback without leaning on invented statistics. If you want the wider context first, our AI automation overview explains where automation fits alongside strategy and AI agents.
The five things that actually drive the cost
Almost every estimate for automating a process comes down to the same handful of factors. None of them are exotic — but together they explain why two automations that sound similar can differ tenfold in price.
- Number and complexity of steps. Automating a three-step task is inexpensive. Automating a fifteen-step process with branches, approvals, and conditional logic is not. Each decision point is something that has to be specified, built, and tested.
- Systems to integrate. Reading one inbox is easy. Pulling from your ERP, writing to your CRM, posting to a finance system, and checking a contract repository means four integrations — each with its own API, authentication, and quirks. Integration is usually the single biggest hidden cost.
- Data quality. Clean, consistent, well-labelled data makes automation low-cost. Messy data — inconsistent formats, missing fields, scanned PDFs of varying quality — means you spend most of the budget on cleaning and validation before any AI runs. This is where solid data engineering earns its keep.
- Exception handling. The happy path is the easy 80%. The cost lives in the other 20% — the unusual invoice, the ambiguous request, the edge case nobody documented. Deciding what the system does when it is unsure (and how a human steps in) is often half the work.
- Volume and reliability needs. Automating something that runs ten times a day is forgiving. Automating something that runs ten thousand times a day, where an error is expensive or regulated, demands monitoring, logging, fallbacks, and testing that all add cost.
If you only remember one thing: integration and exceptions cost more than the AI itself. The model is rarely the expensive part. Connecting it safely to your real systems, and handling everything that does not fit the pattern, is where the budget goes.
Low-cost vs expensive automations — concrete examples
It helps to anchor the abstractions in real scenarios. Here is roughly where common automations land and why.
On the less expensive end: extracting structured data from invoices or receipts, transcribing and summarising calls, classifying and routing incoming email, drafting first-pass replies, or tagging documents. These share a pattern — one clear input, one clear output, high volume, and tolerance for a human to glance at the result. The process is contained, the integrations are few, and the AI is doing one well-bounded job.
In the middle: a customer-intake assistant that answers questions from your own documents, checks a couple of systems, and creates a ticket. Or a reporting pipeline that gathers data from several sources, generates a draft, and flags anomalies for review. More steps, more systems, more to get right — but still a recognisable, repeatable shape.
On the expensive end: a multi-system agent that takes an instruction, reasons about it, reads and writes across several business systems, makes decisions, and acts with limited supervision. Think of an agent that handles an order end-to-end — checking stock, pricing, credit, and logistics across four systems and deciding what to do at each turn. The cost here is not the model; it is the orchestration, the guardrails, the testing, and the trust you have to build before you let it act. Our piece on AI agents vs RPA explains why this kind of judgement-driven work is a different animal from rule-based scripting.
The lesson is simple: the more a process resembles "one input, one output, do it many times", the less expensive it is. The more it resembles "understand a situation and decide across several systems", the more it costs.
Build vs subscription: when each makes sense
A real cost lever is whether you buy a subscription tool off the shelf or build something around your own process. Both are legitimate; they just suit different situations.
Subscription tools are right when your need is common and your process can bend to fit the tool. Transcription, generic chatbots, document OCR, and many marketing tasks have mature products you can pay for monthly. The cost is predictable and low to start, you carry no maintenance burden, and you are live in days. The trade-off: you adapt to the tool, your data may leave your environment, and per-seat or per-usage fees can climb as you scale.
A custom build is right when the automation is specific to how *you* work, touches systems no off-the-shelf product knows about, or handles data you cannot send to a third party. The upfront cost is higher, but you own it, it fits your process exactly, and you control where the data sits — which matters under the EU AI Act and GDPR. Over time, a build can be less expensive than rising subscription fees, and far less expensive than the cost of a process that never quite fit the tool.
Often the honest answer is a mix: subscribe for the commodity parts, build the slice that is genuinely yours. A good consultant will tell you when *not* to build — that is part of the value of an honest AI consulting engagement.
How Crux Digits prices it: fixed-scope, no surprises
Open-ended day-rate projects are where automation budgets quietly run away. We work differently. Crux Digits prices AI automation in three fixed-scope steps so you always know the number before the work begins.
- AI Audit & Strategy — EUR 2,500 (fixed). We map your process, check whether AI is genuinely the right tool, score the data, and tell you what is feasible, what it would cost, and what the payback looks like. If automation is not worth it, we say so. This is the least expensive way to get a grounded estimate for your specific case.
- Proof of Concept — EUR 20,000 (fixed). We build a working version on your real data and your real process, so you can see it run and judge the results before committing to production. It de-risks the big decision. See what an AI proof of concept costs for the detail on what is — and is not — included.
- Production Launch — from EUR 50,000. We harden the system, build the integrations, add monitoring and exception handling, and put it into daily use. "From" reflects the real range — a single-system automation sits near the bottom; a multi-system agent sits higher.
The point of fixing the scope is that you are not signing a blank cheque. You decide at each gate whether to continue. Full detail lives on our pricing page and the AI consultancy pricing breakdown.
Estimating payback from hours saved
The cost only means something next to the return. The most reliable way to estimate payback for a process automation is not exotic — it is hours saved, valued honestly.
Start by measuring how long the task takes today, per item, and how many times it runs. Multiply by a realistic loaded hourly cost for the people doing it. That gives you the recurring cost of doing it by hand. Then estimate, conservatively, how much of that the automation removes — rarely 100%, because someone still reviews exceptions. The gap between the two is your annual saving, and the project cost divided by that saving tells you roughly how many months until it pays for itself.
A few honest cautions. Do not count time "saved" that simply gets absorbed into other low-value work — count it only if it frees capacity you will actually redeploy or avoid hiring for. Factor in the soft gains too — fewer errors, faster turnaround, less weekend firefighting — but treat them as supporting reasons, not the headline number. And include the run cost (below) on the cost side of the ledger.
We deliberately avoid quoting an "average payback period" — it would be a made-up number, because it depends entirely on your wage costs, your volume, and your process. The Audit exists precisely to put real figures into this calculation for your situation.
Maintenance and run costs — the part people forget
The build is a one-off; running the automation is forever. Leaving run costs out of the estimate is the most common budgeting mistake, so plan for them from the start.
- Model and infrastructure usage. Every AI call has a cost, whether it is a per-token API fee or your own compute. For high-volume automations this is a genuine line item, though usually modest next to the labour it replaces.
- Monitoring and support. Production systems need someone watching that they still work, that accuracy has not drifted, and that an upstream system change has not quietly broken an integration.
- Change and drift. Your process evolves, your source systems get updated, and AI models themselves change. Periodic tuning and re-testing keep the automation accurate and compliant — and keep it aligned with the EU AI Act as the rules phase in. Our note on keeping AI agents reliable in production goes deeper here.
As a rough mental model, treat ongoing run and maintenance as a recurring fraction of the build — higher for complex, multi-system automations, lower for a single contained task. The good news is that this cost is predictable, and it is exactly the kind of thing a fixed-scope Audit will lay out for you up front.
Bringing it together
So, what does AI automation cost? Less than you fear for a contained, high-volume task; more than a subscription for a true multi-system agent — and the honest figure for *your* process depends on its steps, its integrations, your data, and its exceptions. The most expensive mistake is not the build; it is automating the wrong process, or signing up for open-ended work without a clear scope.
That is why we start small and fixed: a EUR 2,500 Audit gives you a real number, a real payback estimate, and a straight answer on whether to proceed — before you spend more. If you would like that number for a specific process you have in mind, book a free consultation and we will talk it through. No hard sell — just an honest read on whether automation is worth it for you. You can also explore how this plays out by sector on our industries pages.