aceArt · AI Agency

AI automation for processes and workflows

In almost every company there is a routine that somebody has done by hand for years, because it never quite fitted into rules. Reading an email and filing it. Typing up a delivery note that arrived as a PDF. Checking whether two records mean the same company. This is exactly where AI is strong, and exactly where it is worth starting.

Get expert advice

Rules for the unambiguous, AI for the rest

Automation is older than AI and works for everything that can be described clearly. If an invoice comes in, file it and send a notification. Rules like that are cheap, fast and always do the same thing. What they could never handle is ambiguity: is this email a complaint or a question? Does the invoice number sit in the top right for this supplier, or somewhere in the middle of the text?

Language models are good at precisely that grey area. So we do not build workflows in which AI does everything. We build ones where it takes on the fuzzy step and the rest stays classic. That is cheaper to run, easier to check and fails less often. Where a rule is enough, we use a rule.

Routines that regularly pay off

Sorting the incoming post

Reading enquiries from the mailbox, sorting them by subject, assigning them to the right team, judging how urgent they are and preparing a draft reply. The gain is rarely the response time. It is that nothing sits in the wrong mailbox for days any more.

Reading documents

Invoices, delivery notes, orders, forms. What used to fail on differing layouts now works remarkably reliably, because the model understands the document instead of looking at fixed positions. Unclear cases are put to a person rather than guessed at.

Matching data between systems

CRM, ERP, shop and accounting know the same company under three spellings. Reconciling them is tedious, because it constantly calls for judgement. A routine that merges the clear cases itself and only puts up the doubtful ones saves a noticeable amount of time here.

Recurring compilations

Reports, overviews, minutes. Pulling numbers from several sources, putting them into shape, sending them out. This is the classic case where one person in the building does the same thing every Monday morning and nobody else knows how it works.

n8n, Make or custom code

For simple and mid-sized routines we use tools like n8n or Make. Both assemble a routine from building blocks, both talk to the usual systems, and both have the advantage that your team can change something afterwards without us. The difference is less in the feature list than in how they are run: n8n can be hosted on your own server, which is what decides it when sensitive data is involved; Make runs as a service and is quicker to get going.

As soon as a routine collects a lot of special cases, runs at high volume or reaches deep into your business logic, we build it as proper software. The point is usually easy to spot, namely when nobody can read the building-block diagram any more. A combination of the two is common: the frame in the tool, the complicated step as a separate service beside it.

What can go wrong, and how we catch it

An automated routine that quietly does the wrong thing is worse than no routine at all. So every workflow comes with the question, from the very start, of what happens when a step fails or when the model is not sure.

  • Uncertain cases land on a review list instead of slipping through. A routine that handles 80 per cent itself and puts up 20 per cent is a good routine.
  • Every run is logged, so it can be traced why something was decided the way it was.
  • Failures report themselves. A workflow that has not run for three weeks and that nobody missed was never important.
  • Cost limits, because every model request is billed and a mistake in the routine gets expensive otherwise.
  • Before going live, the routine runs alongside for a while without changing anything. That way you see the hit rate before it counts.

From the first routine to a system

We start with one routine, not with an automation strategy. The first one should be small enough to be running within weeks, and important enough that somebody misses it when it fails. What comes after that usually follows on its own, because other departments want it too.

When a routine needs several steps in a varying order, the move to an AI agent is a gradual one. And if it turns out that the automation belongs inside your own application rather than next to it, AI integration is where it continues. Which routine to do first is something we work out in the AI consulting.

Frequently asked questions about AI automation

Which processes can be automated with AI?
Best of all the ones where somebody reads, sorts and forwards today. Sorting the incoming post, reading documents, matching data between systems, producing recurring compilations. A rule of thumb: where the sequence is fixed, classic automation is enough. Where somebody has to think for a moment every time, AI comes into play.
n8n or Make, which one suits us?
Both can do most of it. Make runs as a service and is quicker to get going, n8n can be hosted on your own server, which is what decides it when sensitive data is involved. If your team wants to tinker with it afterwards, it comes down more to which one they feel comfortable in. We work with both.
What happens if the AI makes a mistake?
The routine is built so that uncertain cases are put to a person instead of slipping through. Every run is logged so decisions can be traced. Before going live we let the routine run alongside for a while without it changing anything, so the hit rate is known before it counts.
What do the running costs look like?
They come in two parts: the tool or the server the routine runs on, and the model requests, which are billed per run. The second part grows with volume and is easily overlooked. We work it out from your actual volumes before we build.
Can we change the routines ourselves?
With n8n and Make yes, that is what they are made for, and we set them up so the obvious adjustments work without us. With custom-built services it takes development work. Which of the two fits better depends on how often the routine is going to change.
How long does the implementation take?
A manageable first routine is often in production within a few weeks. When it takes longer, that is usually not down to the AI but to the access to the systems involved. If it needs to be quick, sort out in advance who is allowed to grant the interfaces.
Do we need clean data for this?
Less than you used to, because models cope well with differing wordings and layouts. What you do need to know is which system is the source for which piece of information. Without that answer you automate contradictions.
Does this replace jobs?
In the cases we build, manual work disappears that nobody enjoyed doing anyway, and there is more time for the cases that need judgement. To be honest, that depends on the project. If you discuss it openly with the team before the first routine goes live, everyone is spared an unpleasant surprise.

Related topics

AI agents

For when a routine needs several steps in a varying order. To the AI agents.

AI integration

For when the automation belongs inside your own application. To AI integration.

AI consulting

To find out which routine to start with. To AI consulting.

Ready for your KI project?

Tell us about your project. We'll get back to you promptly with an honest assessment.

Contact us
HOME › AI-AUTOMATION