For the past few months, AI agents have been constantly discussed online, both in tech communities and across almost every social media platform.

These agents are often presented as the solution to everything.
“Stop writing prompts, build an agent” is a sentence many of you will probably have seen on Instagram in recent weeks. They are increasingly presented as a cure-all for any problem we want to delegate to AI.

But do we really need to build an agent for everything we want AI to do?

As I mentioned in the article AI for Architects, where to start, architects should first identify the repetitive or cumbersome tasks they want to automate, and only then ask which system is best suited to speed them up or automate them completely.

The main risk is becoming too attached to a technology, or to the idea of one, before understanding whether it is actually the right one for the problem. If you don’t yet have a clear idea of what you want to achieve, there is little point in deciding how to build it.

AI Automation, workflows and agents: what’s the difference?

There is an important difference between traditional automation, AI-powered automation, AI workflows and AI agents: it comes down mostly to how much autonomy the system has over the process.

Traditional automation

Traditional automation existed long before AI. It follows fixed, deterministic rules that produce predictable results through a series of basic and repeatable actions.

It doesn’t need to understand the situation or make complex decisions beyond:

If X happens → do Y

Some examples include:

• Email auto-responders: Sending a preset “out of office” reply immediately when an email arrives.
• Invoice generation: Automatically creating and emailing a PDF invoice when a customer completes an online purchase.
• Data transfer: Moving structured data from one database to another using fixed connections.
• Automated file sorting: Moving email attachments into a specific folder based on the sender or file type.
• Form validation: Preventing a form from being submitted when required information is missing.

The main advantages of such automations are their simplicity and predictability. They are usually easy to understand, test and maintain.

The main limitation, however, is that they struggle when the situation changes or when the input doesn’t fit the rules they were designed around.

Traditional automation doesn’t need to understand the situation. It needs to recognize the condition.

AI-Powered automation

AI-powered automation is an automated process where AI is introduced into one or more steps.

The overall process remains predefined, but AI can make individual steps more flexible because it can interpret information rather than simply follow fixed rules.

For example:

Incoming email → AI reads → classifies → extracts information → sends to the correct destination

The process itself is still predetermined. The AI isn’t deciding how to redesign the workflow or what the overall objective should be.

Instead, it is making individual steps more capable of dealing with unstructured information and context.

The workflow decides what happens next. AI helps with individual steps

AI Workflows

With AI workflows, we start moving towards more AI-native processes.

The overall path is still defined in advance, but AI can have more freedom and reasoning capabilities within individual steps.

For example:

Project data

↓

AI Researcher

↓

AI Strategist

↓

AI Writer

↓

AI Critic

↓

Human approval

↓

Publish

The order of the workflow is still determined by the person who built it. AI decides how to operate within the individual steps, but it doesn’t necessarily decide to completely change the process.

You design the path. AI operates within it.

AI Agents

This is where the system is given a goal rather than a completely predefined sequence of steps.

An agent can interpret the goal, decide which tools to use, determine what to do next, observe the results and change its approach when necessary.

For example:

«GOAL

↓

AGENT

↓

Search / CRM / Web / Documents

↓

Observe results

↓

Decide next step

↓

Act

↓

Evaluate

↓

Repeat if needed»

The important difference is that the agent has some autonomy over how the goal is achieved.

It doesn’t mean that an agent should have unlimited freedom. Its tools, permissions, boundaries and human approval points can still be clearly defined.

A workflow follows a path. An agent can choose the path.

The difference is autonomy, not AI

This is perhaps the easiest way to understand the difference.

Ask yourself:

Who decides what happens next?

With traditional automation, the human defines the rules and the actions.

With AI-powered automation, the human defines the workflow while AI helps perform individual steps.

With an AI workflow, the human defines the overall process, while AI has more flexibility within individual stages.

With an AI agent, the human defines the goal, tools and boundaries, while the agent can determine some of the steps required to reach that goal.

The important difference is how much control the AI has over the process.

Not everything called an AI agent is an agent

This is becoming an important distinction as more AI products and services are marketed as “agents”.

For example, you might have a system that looks like this:

«Trigger

↓

GPT

↓

Gmail

↓

GPT

↓

Notion»

This can be a very useful system.

But the fact that it uses GPT and several different tools doesn’t automatically make it an AI agent.

If the process is rigidly: A → B → C → D, and the system cannot decide whether to skip B, return to A, use another tool, search for additional information or change its strategy, it would often be more accurate to describe it as an AI-powered workflow or automation.

Using multiple AI tools doesn’t automatically make a system an agent.

An agent is not defined by how many tools it uses.

An agent is defined by how it decides to use them.

So what does your architecture practice actually need?

The answer depends on the type of process you are trying to improve.

Traditional automation

Traditional automation works well for simple and predictable tasks such as:

• organizing files;
• recurring calendar actions;
• standard notifications;
• fixed document routing;
• repetitive administrative tasks.

If the process is stable and predictable, there is usually no reason to make it more complicated than necessary.

AI-powered automation

AI-powered automation becomes useful when the process is predictable but the information isn’t.

For example:

• classifying emails;
• extracting information from PDFs;
• summarizing documents;
• processing project information;
• turning unstructured information into structured data.

The workflow can remain predefined while AI handles the parts that require interpretation.

AI Workflows

AI workflows are useful when several AI-assisted steps need to work together.

For example:

• client onboarding;
• proposals;
• meeting follow-ups;
• FF&E information;
• knowledge management;
• project administration.

The process has several stages, but the overall structure can still be designed in advance.

AI Agents

Agents become more interesting when the path to the result isn’t completely predictable.

For example:

• a client request could require different actions depending on what the system finds;
• the system may need to choose between different tools;
• information may need to be researched and evaluated before deciding what to do next;
• the system may need to perform several actions and reassess the situation along the way.

The more variable the path, the more useful agentic behavior can become.

Don’t build an agent just because you can

More autonomy also means more complexity.

An agent can introduce:

• more things to test;
• more possible failure points;
• more monitoring;
• more maintenance;
• potentially higher costs and latency.

That’s why it’s important to start with the goal and workflow in mind: sometimes a simple automation that works reliably every day can be much more valuable than an advanced agent that needs constant supervision. 

The goal is to reach the most useful degree of autonomy.

What should stay human?

No matter which system you choose, some parts of the process should remain under human control.

Professional judgment

AI can prepare information for a decision, but the architect should make the professional decision.

Sensitive communication

Difficult clients, negotiations, conflicts and important project decisions often require context and judgment that should not simply be delegated to an autonomous system.

Creative decisions

AI can help explore possibilities, but architecture still requires creativity, experience and an understanding of people and context.

Responsibility

A system can support a decision or even perform actions on behalf of a practice, but responsibility still needs clearly defined human ownership.

The more autonomy you give a system, the more important its boundaries become.

How do you choose?

Before deciding what technology to use, look at the workflow itself.

	Traditional automation	AI automation	AI workflow	AI agent
Process	Fixed	Predefined	Predefined	Dynamic
AI role	None	Individual steps	Multiple steps	Decision-making
Path	Fixed	Fixed	Mostly fixed	Can change
Tool use	Predetermined	Predetermined	Predetermined	Can be selected dynamically
Best for	Repetitive tasks	Unstructured inputs	Multi-step processes|	Variable, goal-driven tasks
Human control	High	High	High	Defined by boundaries

A simple rule can help:

Start with the simplest system that solves the problem.

If a traditional automation can solve it, use one.

If the workflow is predictable but requires AI to interpret information, use AI-powered automation.

If several AI-assisted steps need to work together, build a workflow.

If the system needs to determine its own path towards a goal, an agent may be appropriate.

Final takeaway

Do you actually need an AI agent?

Maybe.

But you should only know that after understanding the workflow.

Automation isn’t outdated.

Workflows aren’t less intelligent.

Agents aren’t automatically better.

The right system depends on the problem you are trying to solve and how much autonomy that problem actually requires.

Don’t build an agent because you can. Build one because the workflow needs one.

And, as with everything else we discussed in the first article:

Start with the workflow. Choose the technology second.