Almost every business wants to automate its processes with AI, and almost every project starts in the wrong place: with the tool instead of the process. This article explains what automating with AI and agents means in practice, which processes to tackle first, how to grow from fixed rules to an agent without losing control, and which pitfalls we run into most often.
Automating a business process means that recurring steps are no longer carried out by a person, but by software. That is nothing new: integrations between systems, automatic reminders and if-then rules have existed for decades. What AI adds is that software can now also handle information that has no fixed shape. An incoming email, a PDF from a supplier, a conversation with a customer, a half-completed form: exactly the kind of input classic automation stalled on, because a person first had to look at it to determine what it was.
An AI agent takes that one step further. Where an AI model answers one question or assesses one document, an agent carries out several steps in a row: looking up information, updating a system, involving a colleague, creating a task. It uses tools and integrations to do so, and stays within boundaries you have defined in advance. That last part is not a detail. An agent without boundaries is not automation but a risk.
Once that distinction is clear, you immediately see where most projects go wrong: they start with the agent, while the process underneath has never been mapped out.
It helps to see process automation as three levels. Not because the highest level is the goal, but because each level builds on the previous one.
Several steps in a row, independently, with tools and logging. Only once level 2 is demonstrably stable.
Classifying, extracting, drafting. An employee approves before it moves on.
If this, then that. No AI model needed, yet the biggest and fastest win in most businesses.
Level 1: rules and integrations. Everything that works with a fixed rule. An order that comes in appears in the planning. An invoice that is thirty days overdue gets a reminder. A new contact in the CRM is also in the newsletter tool. No AI model is involved here, and this is still where the biggest win sits in most businesses. Duplicate entry, retyping and lists that are updated by hand: they are all rules nobody has automated yet.
Level 2: AI as a step in the process. On the points where an assessment is needed, an AI model takes over that step. An email is classified as a complaint, question or order. The relevant data is extracted from a PDF. A draft reply is ready before the employee opens the email. The human stays in the process and approves, but no longer has to read, search and type everything personally.
Level 3: agents that take over several steps. Once the AI step from level 2 demonstrably works well, an agent can run several steps in a row without intermediate approval. The complaint is recognised, the file is looked up, the customer receives a confirmation, the right colleague gets a task with all the context attached. The employee sees the result, not every intermediate step.
A complaint, question or order, in free text, sometimes with an attachment.
Classifies the email, extracts the key data and finds the customer file.
Sees the proposal with all the context and confirms, corrects or takes over.
Sends the confirmation, updates the system, creates a task and logs every action.
The order is not theory. Whoever attempts level 3 without levels 1 and 2 builds an agent that has to work with data that is not right, in a process nobody has mapped out. That is why so many AI pilots remain an impressive demo.
Not every process is a good candidate. The processes where automation with AI pays off fastest share a few characteristics: they occur often, they run largely predictably, the required information is available, and if something goes wrong once, it is recoverable. A process that happens three times a year and where a mistake costs you a customer is not the one to automate first, however annoying it is.
In practice we keep running into the same candidates at SMEs:
If you recognise several of these, still pick one. The first process is not meant to deliver the most value, but to learn how automating with AI works in your organisation. To base that choice on numbers rather than on a hunch, read spotting repetitive work: which task do you automate first?
There is one condition that matters more than which tool you choose: the data the process runs on must be available and reliable. An agent that has to handle a complaint needs access to the customer data, the order history, earlier correspondence. If that is spread across a CRM, an accounting package, a mailbox and a pile of spreadsheets, the agent spends most of its time searching and guessing.
This is exactly the pattern we described in five signs your business has outgrown its CRM: AI plans rarely fail on the AI, but on fragmented data. The answer is not a bigger licence or a smarter tool, but a foundation where all business data sits together and agents can plug in. How such an open, in-house data layer works, and why it replaces the monolithic CRM rather than supplementing it, is explained on the AI Native Data Layer page.
If your data is still stuck in a package that will not cooperate, a move comes first. How to do that in a controlled way, without losing anything, is in the data migration plan.
This is the order we follow ourselves. Every step delivers something you can use, even if you then decide not to go further.
Who does what, with which information, and where does it get stuck? Usually the process runs differently than everyone thought.
How much time does it take now, how often does it go wrong, how long does a customer wait? Without a baseline you cannot prove anything later.
Integrations, status updates, reminders, calculations. Often half the manual work disappears here, without an AI model.
Classifying, extracting, drafting. An employee approves; track how often that approval becomes a correction.
Only once step 4 is stable. Define what the agent may and may not do, what always goes to a human, and log every action.
Compare with the baseline, adjust the boundaries. If it works, the next process is up.
The pattern is always the same: start small, prove that it works, and only then expand. Whoever keeps that up has, after a few months, not one impressive demo but a series of processes that simply run.
As soon as AI looks into a process, the question arises which information may be used for that. The honest answer is that it differs per step. Classifying an email requires the content of that email, but not the complete customer file. A draft quote requires the product data, but not the HR records. Whoever records per step which data the model gets to see keeps control, and can explain it to a customer or regulator too.
Beyond that, it matters which provider does the processing. The major AI providers differ considerably in how they handle business data, whether they train on it and which agreements they offer business customers. We have a clear preference here, and explain on the AI Consultancy page why it lies with Claude and what the alternatives are.
Finally: make sure you can get out again. A process that is fully dependent on one provider is a new kind of lock-in. From 2027 the European Data Act gives you the right to leave a software provider without switching costs; what that means concretely is in Data Act: from 2027, switching costs you nothing. But rights are only worth something if your data and your process logic are practically yours as well.
Starting with the tool. A licence is signed, and then a process is sought to apply it to. Turn it around: first the process, then the question of which level of automation it needs, and only then the tool.
Wanting everything at once. Automating five processes in parallel means five half results and nobody who knows what works. Taking one process all the way delivers more, including in knowledge.
Automating a bad process. If the process itself is wrong, automation only makes it wrong faster. Mapping it out in step 1 is exactly there to discover that before anything gets built.
No owner. An automated process needs someone who keeps an eye on it, handles the exceptions and adjusts the boundaries. Without an owner it slowly dies, or worse: it keeps running while nobody looks anymore.
No fallback. What happens if the model is briefly unavailable, or if the agent encounters something it does not recognise? A good process has a path back to a human, and that path is tested before it is needed.
Classic automation follows fixed rules: if this, then that. AI adds the ability for software to handle unstructured information such as emails, documents and conversations, and to make an assessment. An AI agent goes one step further: it carries out several steps in a row independently, uses tools and systems to do so, and stays within boundaries defined in advance.
Processes that occur often, run predictably, where the required data is available and where a mistake is recoverable. In practice these are usually handling incoming email, preparing quotes, transferring data from documents, checks in the administration and recurring reports.
No. Most of an automated process usually consists of fixed rules and integrations that involve no AI model at all. A model only looks at the steps where judgement is needed, and there you decide in advance which data may be used for that and which provider processes it.
Automating a first process generally takes days to a few weeks, not months, provided you start small and the data is available. The biggest delay is not the technology but fragmented data: if the required information is spread across systems and mailboxes, it has to come together first.
You do not need an AI strategy to begin. You need one process everyone recognises as a waste of time, an afternoon to map it out, and the discipline to automate the rules first before you put a model to work. Everything after that follows from what you learn in that first round.
If you would rather not do that alone, that is exactly what our AI Consultancy is for: we come to you, look at your processes together, set up the environment and get the first automation and dashboards working. Not as a six-month project, but as a first step that has to prove itself before the next one follows.
In a free advisory call we look at your business processes, your systems and your data together, and determine where automating with AI pays off fastest for your organisation. On site or online.
Book a free advisory call