Rules, workflows, AI or agents? Choose the simplest thing that works.
Not every automation problem needs AI, and not every AI problem needs an agent. Moving right buys capability and costs complexity, money, risk and maintenance.
Not every automation problem needs AI. And not every AI problem needs an agent. The further you move from simple rules toward autonomous agents, the more capability you gain — but usually also more complexity, cost, risk and maintenance.
“Could we use an AI agent for this?”
It is becoming a common starting point for software conversations.
Sometimes the answer is yes.
But often, the better question is:
What is the simplest approach that can solve this problem reliably?
A predictable pricing decision may need a rule. A sequence of repetitive actions may need a workflow. An email or PDF that needs interpreting may justify AI. A system that genuinely needs to decide what to do next may need an agent.
Those are different problems.
Treating them as the same one can make a simple solution unnecessarily expensive.
Start with rules
Rules are the simplest option when both the input and the decision are predictable.
For example:
- orders over a certain value receive a particular discount;
- invoices above a threshold require an additional approval;
- requests from an existing customer go to their account manager;
- a product category determines which price list applies.
There is no need for interpretation here.
If you can describe the decision clearly as “if this, then that”, conventional software is usually the right place to start.
Rules have another advantage: they are easy to understand.
When something goes wrong, you can usually see why.
That matters more than it sounds.
Use a workflow when several predictable steps need connecting
Sometimes no individual step is difficult. The problem is that somebody has to keep moving information between them.
A quoting process might be:
Request → pricing → document → email → follow-up
Each step is predictable.
The useful automation is not a model making decisions. It is a workflow making sure the next action happens without somebody having to remember it.
For example:
- capture the request;
- apply the pricing rules;
- prepare the quote;
- send it for review or approval;
- generate the document;
- send it;
- schedule the follow-up.
This is where traditional workflow automation is extremely effective.
And because the steps are known in advance, the system can remain largely deterministic: the same input should lead to the same expected action.
Use AI when interpretation is the difficult part
The picture changes when the input is not predictable.
A customer may explain the same requirement in ten different ways.
A supplier may send a specification in a PDF with a different structure every time.
A support request may need to be understood by meaning, not by looking for one particular keyword.
That is where AI starts to earn its place.
For example, it can:
- extract requirements from emails or documents;
- classify messages by intent;
- summarise a long conversation;
- identify relevant information in free text;
- turn unstructured input into structured data the rest of the workflow can use.
The important point is that AI does not need to run the whole process.
Often, it should handle only the ambiguous part.
Everything around it can still be ordinary, predictable software.
That usually makes the solution easier to test, monitor and maintain.
Use an agent when the system genuinely needs to choose what happens next
An agent is useful when the system cannot simply follow a fixed sequence.
It may need to:
- decide which tool to use;
- gather information from several systems;
- react differently depending on what it finds;
- retry or change approach;
- decide when it has enough information;
- escalate when it cannot resolve the case safely.
Imagine an exception in an order.
The system may need to check the CRM, inspect an invoice, look at stock, compare contractual terms and then decide which action is appropriate.
That is much closer to an agentic problem.
But with that flexibility comes a different engineering problem.
You now need to think about permissions, observability, failure modes, limits, human approval and what happens when the system makes the wrong choice.
An agent is not simply “better automation”.
It is a more capable system with a larger operational surface.
Moving right is a trade-off
Rules, workflows, AI and agents are not four generations of the same technology.
You do not “upgrade” from one to the next because the next one is more modern.
You move right only when the problem requires it.
As you do, several things generally increase:
Complexity. More components, integrations and possible behaviours.
Cost. More development, infrastructure, model usage and monitoring.
Risk. More ways for the system to behave differently from what you expected.
Maintenance. More things that can change: prompts, models, tools, permissions, integrations and business assumptions.
That does not make AI or agents a bad choice.
It makes them a choice that should earn their place.
A simple test before moving one level to the right
Before choosing the technology, ask:
Can the decision be expressed reliably as a rule?
If yes, start there.
Are the steps known in advance?
If yes, a workflow may be enough.
Is the difficult part understanding messy or variable information?
That is a strong candidate for AI-assisted processing.
Does the system genuinely need to choose its own next action?
Now an agent may make sense.
And one more question matters at every level:
What happens if it gets this wrong?
A system generating a draft description and a system approving a €100,000 transaction should not have the same level of autonomy.
The consequences of failure should influence the architecture.
The best system may contain all four
These approaches are not mutually exclusive.
A good system might use:
- rules for pricing;
- a workflow for moving the quote through the process;
- AI to interpret an unusual customer email;
- a human to approve an exception.
Or, in a more complex case, an agent might investigate the exception before handing it to the person who makes the final decision.
The goal is not technological purity.
It is to put each kind of problem in the simplest layer that can handle it reliably.
Start on the left
There is no prize for using the most sophisticated architecture.
If a rule solves the problem, an agent is not an upgrade.
If a workflow solves it, adding a language model may simply add another dependency.
If genuine ambiguity remains, AI becomes useful.
And when the system must reason across tools, context and changing circumstances, an agent may be exactly the right answer.
Start on the left. Move right only when the problem makes you.
If you are not sure which level a process actually needs, that is something we can look at in the 20-minute check. We take one real process, separate the predictable work from the ambiguous parts, and identify what is worth automating — and what is not.
Related: This process does not need AI. It needs fewer steps. · What we’d tell you not to automate. · Task calculator