AI automation for business: a guide to getting started
By Jacob Slot · Published
AI automation is software that carries out the fixed steps of a workflow on its own, pulls and moves data between your systems, and hands the task to a person only when something falls outside the rule. The AI automation businesses actually run in production is therefore neither a chatbot on the front page nor an employee with ChatGPT in the browser, but a chain that runs by itself from trigger to finished result.
That is the difference most guides skip. 66 percent of small and mid-sized Danish businesses use at least one AI technology, but under 10 percent have AI integrated into their internal processes, according to CAISA's survey of Danish SMBs. Using AI is not the same as running on it.
Below you get the four layers an automation can sit in, where it pays off first, what a person still has to approve, and what it costs to get started.
What AI automation means for a business
AI automation in a business means that software performs a recurring task from start to finish, and that a model is used only for the step that varies: a receipt to read, an email to understand, a text to write. Ordinary integrations handle everything that is fixed.
Three words get used interchangeably, and the difference costs money. Digitalisation is moving the paper onto the screen. Automation is removing the clicks afterwards. AI is the layer that can handle text and documents that no rule can describe in advance.
A PDF invoice in the inbox is digitalisation. An invoice that gets read, coded and queued for approval on its own is AI automation.
The split across Danish business is lopsided. 96 percent of companies with more than 250 employees use artificial intelligence, against 57 percent of those with fewer than ten, according to Dansk Erhverv's analysis of AI use in 2025. The gap is rarely the technology. It is that nobody in the small company has had a week to write down what actually happens today.
The four layers of AI automation
Almost everything we build sits in one of four layers. They are worth knowing because price, risk and maintenance rise with each step, and because most tasks can be solved further down than a salesperson suggests.
| Layer | What it handles | What it requires | Example |
|---|---|---|---|
| 1. Rule-based integration | Fixed steps between two systems | An API at both ends | Webshop order becomes an invoice in the accounts |
| 2. Document understanding | Numbers and fields out of receipts and PDFs | A stable document flow | Purchase invoice read, coded, queued for approval |
| 3. Language and drafts | Reading, sorting and writing text | Your own examples of a good reply | Inbox sorted, reply written as a draft |
| 4. Multi-step agent | Tasks with decisions along the way | Clear limits and logging | Enquiry looked up in three systems, case created |
Layers 1 and 2 are where the money is for most small and mid-sized Danish businesses. They are cheapest to build, they fail visibly instead of silently, and they touch the time an employee spends every day.
Layer 4 is what gets sold hardest right now. It is also the only layer where a wrong decision can reach a customer before anyone sees it. We build it when the task genuinely requires it, and never as a first project.
Where AI automation pays off first in a business
Four places show up in almost every mapping we do. What they share is that they happen often, take time and follow the same rule every time.
Order to invoice. The classic. The order already exists in the webshop, the inbox or the CRM, and yet someone keys the lines into the accounting system again. This is layer 1, and it is typically the first automation we put into production.
The weekly report. The numbers sit in three systems, and an employee pulls them into a spreadsheet every Monday. The chain is described in detail in our guide to automatic reporting.
Purchase invoices and documents. Supplier invoices, timesheets and contracts that have to be read and coded. This is layer 2, and it is where the difference between AI and an ordinary integration becomes obvious.
The inbox. Enquiries that need sorting, answering or forwarding. Layer 3 writes the draft. A person presses send until the tone is right.
What is not on the list matters just as much: quoting that involves negotiation, HR cases and anything where the price depends on judgment. The parts around them can be automated. The decision cannot.
The systems, and where your data gets stuck
An automation is only as good as its access to the systems. In a small Danish business it is almost always the same names, and they fall into three groups.
Accounting systems have open APIs. e-conomic, Dinero and Billy can all be read and written from outside, which is why the invoice chain is the easiest place to start. What can be built in practice is covered in the guide to AI for e-conomic integration.
CRM and email are the second group. Pipedrive, HubSpot, Outlook and Gmail all have APIs good enough to read, create and update. The problem here is rarely access, but that the same details sit in three places in three versions.
The third group is the closed systems: industry-specific case systems, older ERP and the spreadsheet an employee keeps alongside everything else. There is no clean integration here, and the choice is between an export, a robot that clicks like a user, or leaving that step manual.
Be honest about the third group from the start. A chain that needs someone to download a file every morning is not automated. It has just moved the work.
What a person still has to approve
This question decides whether the project is safe to put into production. We work with three levels, and every link in the chain gets one of them.
- Runs by itself. Data moved between systems, record creation, reminders, internal reports. Nobody reviews them, and nobody needs to.
- Draft for approval. Customer replies, coding, quotes built on a price list. Software does 90 percent of the work, a person presses yes.
- Suggestion only. Decisions that involve judgment. The automation gathers the material and keeps quiet about the conclusion.
Four things should always sit in the two lower levels: money leaving the company, promises to a customer, personal data, and anything that cannot be rolled back. This is not a technical limitation. It is the line that lets management trust the rest to run.
And the exception has to be handled from day one. Every chain has a case that falls outside the rule. It must never stop the whole chain, and it must never disappear in silence. It should land with a named person and a message saying why.
What AI automation costs your business
A scoped build with us starts at 10.000 kr as a one-off, and an ongoing partnership from 5.000 kr./md. On top comes running the platforms and models the chain runs on, which is typically a three-digit amount per month for a single automation.
Key numbers
15 hours
freed up per week for a typical customer
20.000 kr
saved per month at Markeity
Grants change the sums noticeably. The SMV:Digital advisory grant covers 50 % of the advisory cost, up to 150.000 kr, for small and mid-sized businesses (Virksomhedsguiden). The scheme suits a scoped project with a defined outcome, and you apply yourself. The steps are in our guide to the SMV:Digital grant for AI.
If you want a number for your own task, the price ranges are broken down in the review of what AI automation costs, and you can work out the return with your own hours in the ROI calculator.
We write ROI into the agreement: if it is not measurable within 90 days, we keep working at no extra charge.
How your business gets started in 30 days
You do not need a strategy for the whole company. You need one chain in production, and a month is enough.
Week 1: map and measure. Sit down with the employee who does the routine and write down every step, with the system it happens in. Time the task at the same time: how many times it runs, how long each run takes, how often something goes wrong. Without that before-number you cannot show an after-number.
Week 2: pick one. Score each workflow on how often it happens, how long it takes, how fixed the rules are and what a mistake costs. The highest total wins. Not the most annoying one, and not the one the boss mentioned last. The order is described step by step in our method for automating workflows.
Week 3: build only that. Settle access and ownership before a line of code is written: who holds the account, who has the source code, what happens if the collaboration ends. Then build the one flow, test it on real data, and set the approval level on every link.
Week 4: measure and decide. Take the same three numbers as in week 1 and subtract. If it worked, you take number two on the list. If it did not, the mapping was too shallow, and the fix is cheap.
This is the route we have taken 53 times, and it is exactly the same whether you are five people or fifty. It is also the core of what we deliver as AI automation.
When AI automation does not pay off
Don't do it if the workflow changes every month. You will end up maintaining more than you save, and you lock in a process that is still evolving.
Don't do it if the task runs a few times a month and takes fifteen minutes. The sums do not add up, however annoying the task is.
Don't do it if the company cannot agree on how the workflow should look. Then you are automating a disagreement, and it only gets more expensive by running faster.
And don't do it if the data cannot be trusted. A chain that runs fast on wrong information costs more than the manual process it replaced. Clean up first.
FAQ about AI automation in business
What is the difference between using ChatGPT and having AI automation?
ChatGPT in the browser is a tool an employee opens and closes. AI automation is a chain that starts on its own, pulls data from your systems and puts the result back, without anyone having to remember it. The first makes an employee faster. The second takes the task off the calendar.
What is an AI assistant or an AI agent?
An AI assistant answers what you ask it in a chat. An AI agent carries out a scoped, multi-step task: it looks something up in a system, makes a decision within the limits it has been given, and writes the result somewhere. An agent is layer 4 in the table above and needs clear limits and logging before it may run without approval.
How long does the first automation take?
Expect about three weeks from mapping to production for a scoped workflow, with the first week spent writing down what happens today. Chains that touch four or five systems take longer, but they should still be split up so that something goes live along the way.
How big does the company need to be before it makes sense?
There is no lower limit on headcount, only on the amount of repetition. If the same task runs daily and takes more than fifteen minutes each time, there is something to gain, even in a company of three. If it happens once a month, there is not.
Can we get a grant for AI automation?
Yes, for the advisory part. The SMV:Digital advisory grant covers 50 % of the advisory cost, up to 150.000 kr, for small and mid-sized businesses (Virksomhedsguiden). The pools open in rounds, so check the dates before you plan the project, and expect to handle the application yourself.
AI automation in a business does not start with a tool. It starts with one workflow written down, and a decision about which link a person still has to approve.