Skip to content
Back to Insights
ArticleAI & Data7 min read

Do not automate a bad process. Why AI should sometimes start with BPMN, not an LLM

In this article, you will learn why an AI initiative should not start with choosing a model, but with understanding the process. We show when an LLM makes sense, when a simple business rule is enough, and why the greatest value of an AI workshop may be a process diagram, not the model itself.

Łukasz WędzikManaging Director, QRQS.eu PSA15 September 2026

When an organisation starts thinking about using AI, the conversation very quickly turns to technology.

Which model should we choose? Should we use GPT, Claude, Gemini, or an open-source model? Do we need RAG? Agents? Our own infrastructure? How do we connect the model to our systems?

These are important questions. The problem is that they often appear too early.

Before we start thinking about which model to use, it is worth answering a much simpler question:

What exactly are we trying to improve?

That is why an AI project should sometimes start not with an LLM, but with a sheet of paper, a whiteboard, and a process map. Or, if we want to do it a little more formally, with BPMN.

AI does not fix a process. AI performs parts of it faster

Imagine a typical process in a large organisation.

A document arrives. An employee checks whether it is complete. Data is copied into one system. Then someone verifies it in another. If information is missing, an email is sent. After the response arrives, the case returns to the queue. Later, another person makes a decision, and someone else enters the result into the system.

At some point, someone says: „Let's add AI to this."

The model can read the document. It can extract data. It can prepare a response. It can classify the case. An AI agent can even execute several consecutive steps automatically. Technically, yes.

But have we solved the problem?

If the process has a dozen unnecessary steps, three manual re-entries of the same data, unclear decision rules, and five exceptions caused by historical system limitations, AI may simply make a bad process run faster. That is not always a success.

Process first, model second

BPMN may seem unexciting in the world of generative AI.

It does not generate impressive answers. It does not hold a conversation. It does not have an agent that „thinks" what to do next. But it has one major advantage: it forces us to show how the organisation really works.

Who starts the process? What data is needed? Who makes the decision? Which tasks are performed by a person? Which are performed by a system? Where do we wait? Where do we pass data between systems? What happens in an exception? Where is a business decision required, and where is it only a mechanical next step?

Only when we can see the process can we sensibly answer where AI really adds value. And often it turns out that we do not need AI everywhere.

Not everything that can be done with an LLM should be done with an LLM

This is one of the most important principles when designing systems that use AI.

If we have a clear rule: if the amount exceeds a defined threshold, start an additional approval path, we do not need a language model.

If the system has to fetch data from an API, calculate a value according to a known formula, and save the result in a database, we also do not need an LLM.

But if the system has to read an unstructured document, understand the content of a customer message, classify a case, or prepare a summary of several dozen pages of documentation, then AI starts to make much more sense.

This leads to a very useful split.

Keep deterministic parts of the process deterministic. Use AI where there is unstructured information, language, context, or a need for interpretation.

This approach may be less spectacular than the vision of an autonomous agent doing everything, but in production systems it has several important advantages: it is easier to test, monitor, and audit. Most importantly, it is clear what the model is responsible for and what the system is responsible for.

BPMN can be a map for AI

The most interesting point comes when we stop treating BPMN and AI as competing approaches. They solve different problems.

The process defines what should happen. AI can help decide what the information appearing in one of the process steps means.

Imagine an application handling process.

The system receives documents. AI classifies their content and extracts the required information. The process engine checks what the next step should be. The system fetches data from other sources. Business rules perform clear checks. AI analyses the part that requires interpretation. If confidence is too low or a defined exception appears, the case goes to a human. After the decision, the process continues.

In this model, the LLM is not the whole system. It is one component of the system. That is a very important difference.

Where do AI agents fit into this?

As AI agents develop, the boundary naturally starts to move. The model no longer has to only answer questions. It can use tools, call APIs, search for information, and execute subsequent actions. This opens completely new possibilities for automation.

But it also raises a new question: how much freedom do we want to give the agent?

Imagine two approaches.

In the first, the process precisely defines the sequence of actions, and AI is called only in specific places.

In the second, the agent receives a goal and a set of tools, then decides which steps to take.

The second approach is more flexible, while the first is more predictable. In many business uses, we will probably not choose only one of them. The most interesting architectures will be hybrid: the process will define boundaries, rules, and control points, while the agent will receive freedom to act within selected parts of the process.

Instead of asking „BPMN or an AI agent?", it is better to ask: „Which decisions do we want to control, and which can we delegate?"

Human-in-the-loop should not mean „a person checks everything"

This is another trap. In many AI projects, safety is attempted by adding a person at the end of the process. AI does the work, and the human approves everything. Formally, we have human-in-the-loop. In practice, we may create a very expensive system in which the human still does a large part of the work, only instead of creating the result from scratch, they check the machine.

The more interesting question is: in which situations is human intervention actually needed?

It may depend on the type of decision, transaction value, risk level, model confidence, data completeness, or the occurrence of a defined exception. A well-designed process allows these points to be defined clearly.

A human should not be an emergency mechanism added at the end of AI architecture. A human should be a deliberately designed element of the process.

The greatest value of an AI workshop may be a diagram, not a model

This sounds paradoxical, but it happens that while analysing a potential AI use case, the organisation discovers something more important than the possibility of using a model. It discovers that the process can be simplified.

Two steps turn out to do practically the same thing. One approval exists only because a previous system once required it. Data copied manually is already available through an API.

A report generated by an employee can be built without AI. And only in one specific place does a problem appear that is actually well solved by a language model.

In that case, the AI project has still succeeded. The greatest value simply appeared somewhere other than where we initially expected it.

AI should be part of the architecture, not the architecture itself

LLMs are an extremely useful new component of IT systems. But that is exactly how we should look at them: as a component.

We have databases, APIs, transaction systems, process engines, business rules, queues, document systems, authorisation mechanisms. And we have AI models.

The art is not to replace as much of the system as possible with a model. The art is to solve each problem with the right tool. That is why before starting the next AI project, it is worth doing one thing: do not immediately open the API documentation of the chosen model.

First, map the process. See where value is created, where time is lost, where exceptions occur, and where intelligence is actually needed. Only then decide where to place the LLM.

Because the best AI project is not always the one where AI does the most.

Sometimes it is the one where AI does exactly as much as it should.

Tags

  • AI
  • BPMN
  • LLM
  • automation
  • business processes
  • system architecture

Working on a similar topic?

Happy to compare notes on architecture, delivery and audit findings in your institution.

Get in touch