Skip to content
caadoo
de

AI automation cost: what small businesses pay for

Ask what AI automation costs and you'll get a number nobody can justify or a shrug. This piece covers the line items that set the price, what you pay for with a project, a platform or a subscription, and the questions to ask before you sign.

By T. Krause8 min read

Ask what AI automation costs and you'll usually get one of two answers. One is a number, quoted before anyone has looked at the work. The other is "it depends", followed by nothing. Neither helps you set a budget.

A number worth trusting only exists once it's clear which workflow gets automated, which systems it runs through and how clean the data is. So this piece has no price ranges. It walks through the line items that make up the cost and compares the three common ways to buy the work: a one-off project, a software platform and an ongoing subscription.

I sell the third one myself, and you should know that before reading on. That's why each option below also gets its usual way of going wrong, the subscription included.

Photo: Unsplash

What drives the cost

The first build is rarely the biggest line item. The expensive parts are what comes after it and what nobody counted before it.

How many workflows, and how messy they are. A task everyone does the same way is quick to automate. If three people do it three different ways, they have to agree on one first, and that's process work. The software comes later. Exceptions cost the most: the supplier who sends invoices as phone photos, or the big customer with their own order form. Count for one week how often a task comes up and how often it strays from the normal case, and you'll know what you're actually ordering. The article on what to automate first shows how to run that count.

Systems and integrations. Every system an automation has to touch is its own cost item. Current accounting software with a documented API connects quickly. An ERP that only exports CSV files, or industry software with no API at all, needs workarounds, and workarounds break more often. For each system, ask whether there's an official API and whether your plan includes it. Some vendors charge extra for access.

Data quality. An AI model can read an invoice or an email. It won't fix a customer list where the same company appears four times under four spellings. Duplicate records, and free-text fields where a dropdown belongs, have to be cleaned up before the automation runs reliably. That cleanup is missing from a lot of quotes, and then it lands on your staff.

Running costs. Model providers bill language models by usage, usually per token, which roughly means per amount of text. Automation platforms bill per run, per task or per seat, and office software with built-in AI usually per user per month. For one run that's small. For a workflow that runs a thousand times a month and sends a long document to a model each time, it adds up. Ask what a single run costs and multiply by your volume.

Maintenance. Processes change, and so do tools. A supplier switches invoice formats, a SaaS vendor changes its API, a new hire saves files somewhere else. Any of these can quietly stop an automation. Who notices and who fixes it belongs in every quote. Without documentation, the answer is one person, which is the risk covered in automations that survive the builder leaving.

Your team's time. Your people have to explain how the work is done today, supply test cases, check results and sign off. Where the output reaches customers or moves money, a human review step stays in the process for good. Those hours show up on no invoice, but they come out of someone's week. The piece on human-in-the-loop approval covers how to keep that review step small.

Training. People who use AI tools need to understand them. In the EU, Article 4 of the AI Act, as amended by the Digital Omnibus, Regulation (EU) 2026/1744, requires companies to take measures that support the development of AI literacy. It no longer asks for a "sufficient level", and national authorities supervise it from 2 August 2026. What that means in practice is in the article on AI literacy under Article 4. For your budget, training and a record of it are a line item of their own, even if a small one.

Three ways to buy AI automation

Every line item above shows up whichever way you buy. What changes is who carries it and when it appears on an invoice.

A one-off project

An agency or freelancer builds a defined automation, at a fixed price or on time and materials. You pay for analysis, build, testing and handover. After that the solution is yours, and so is its upkeep, unless you sign a maintenance contract.

This fits when the workflow is well defined, rarely changes, and someone on your team can look after it once the agency has gone. A one-time data migration or a fixed monthly report are good candidates.

The usual failure comes a few months later. A connected system changes, the automation stops, and the agency is busy with its next client. The fix becomes a new job, often for someone who didn't build the original. Fixed-price work has a second catch. Anything that wasn't written into the spec is a change request.

A software platform

You license a tool that ships with automation or AI built in. That might be a no-code platform where you click workflows together, an AI assistant inside your office suite, or industry software that reads receipts. You pay for the licence, usually per user or by volume. Your own people build and maintain the workflows. The vendor keeps the platform running and leaves your workflows on it to you.

This fits when your problem matches what the tool was designed for and someone on staff has the time and the interest to learn it. For lots of small workflows in mainstream software, it's often the cheapest route.

The usual failure is the licence nobody uses. The tool is paid for, but nobody has time to set up the workflows, or the one person who could has left. You keep paying every month for something you meant to get round to. The other pattern is sprawl. Workflows pile up without a plan until nobody knows which one moves which data where.

An ongoing subscription

An outside person or small team takes on automation as a standing job, for a fixed monthly fee and an agreed scope. You pay for capacity, meaning analysis, building, upkeep and changes from the same people for as long as the subscription runs. Platform and model fees are either included or billed on top, depending on the contract, so ask which.

This fits when there's more than one workflow, your processes and tools change often, and nobody in-house has the time to own the topic. The maintenance that sits with you after a project is part of the service here.

The usual failure is a subscription without a task list. If nobody decides what gets built next and nobody checks whether the last build helped, you're paying a monthly fee for someone to be available. The second risk is lock-in. If logins, workflows and documentation live in the provider's accounts, leaving gets expensive. Both can be settled in the contract, which is exactly why you should ask. For the check itself, a number from before and one from after is enough, as described in measure before you build.

Which route fits which case

The three aren't exclusive. Plenty of companies buy a platform and bring in a project or a subscription to set it up. What matters is who ends up carrying the maintenance.

A clearly scoped workflow that rarely changes, with someone in-house to look after it, usually costs least as a project. A standard problem in software you already use, with a colleague who has the time, is often fine on the platform alone. Lots of workflows that keep changing, with nobody in-house who has the hours, point toward a subscription. And if you can't tell which case you're in, count first. A week of tally marks tells you more than three quotes.

Questions to ask before you buy

Whichever route you pick, get written answers to these:

  • Which workflow exactly gets automated, and how will we measure whether it paid off?
  • Which systems get connected, and do we need extra licences or API access for that?
  • Who cleans up the data, and is that in the price?
  • What does a single run cost in model and platform fees, and what does that come to at our volume?
  • Who notices when the automation breaks, and how fast does it get fixed?
  • What happens when a connected system or our process changes? Is that a new job?
  • How many hours will our team spend on explaining, testing and sign-off?
  • Who trains the people who use it, and who keeps the record?
  • Whose accounts hold the logins, workflows and documentation, and what do we get if we part ways?

A provider who can't or won't answer these has given you a number without knowing your costs.

caadoo sells the third route, as a subscription for companies of 20 to 250 people. What each plan covers is on the plans page. If you want to know whether a subscription fits your company at all, send me a short note about the workflow you have in mind. If a project or a platform fits better, I'll tell you.

Topics

  • AI automation
  • Costs
  • Small business
  • Automation
  • Vendor selection
All articles

Keep reading

  • 12 min read

    Fractional Chief AI Officer: what the role covers

    Large companies hire a Chief AI Officer. A company of 20 to 250 people rarely needs one full-time, but somebody still has to own the job. This piece covers what that job contains, the signs it's missing, and when a fractional CAIO is the wrong answer.

  • 8 min read

    Automations that survive the builder leaving

    Some automations run only because the one person who built them is still around. That works until they are on holiday, off sick or gone. This is how to take stock of what already runs and give every automation an owner, a one-page runbook, a team alert address, a shared vault and a quarterly check.

End of call

You can hang up now.

Or book 15 minutes. I'll ask you the questions I'd ask on the phone, and tell you at the end whether it fits. If it doesn't, we've both lost a quarter of an hour instead of a month.

Or call. Cold, if you like. My AI assistant picks up and I call you back. +49 5136 9414999

Analytics

With your consent I use Google Analytics to see which pages get read. Without it, no analytics code loads. Your answer is stored locally in your browser.

Privacy notice