The DiagnosisNo. 02

When to use an AI agent — and when not to

What automation means, what AI makes possible and how an agent differs from traditional automation.

Gonzalo GilgesHead of Practical AI

Sent: 15 October 2026
Reading time: 5 min

In a factory, a machine can move, sort or package products by following a sequence. In an office, we can do something similar with information: transfer order data to invoicing, update a record or send a notification. Products become data, but the work is still repetitive and follows defined steps.

Automation means having a system carry out those tasks without a person completing every step manually.

Its strength lies in connecting actions.

Doing those steps by hand makes poor use of a person’s knowledge and judgement. Automating them reduces the time spent sustaining routine operations and leaves more room for analysis, business knowledge and decision-making. AI is not always required.

Automated workflows

Consider collections follow-up. We can tell a system to check invoices every morning, identify those that are overdue, confirm which remain unpaid and prepare the corresponding reminders.

That sequence is an automated workflow. We define when it should act, what it should check and what it should do in each anticipated situation.

It can be programmed using languages such as Python or JavaScript, or configured with visual tools such as Zapier or n8n. This type of automation has been used for more than twenty years: it does not need AI to follow a defined procedure.

Fig. 1Workflow: rules determine what happens next

01 Input
Known trigger.
02 Check conditions
Workflow rules.
03 Condition met?
Workflow rule.
Yes04 Defined action
Workflow.
05 Result
Output.
No06 Stop or request review

In a workflow, people define the conditions and route in advance: the system does not choose how to continue.

What AI makes possible

The difficulty appears when information varies. One supplier sends an invoice as a PDF; another sends a scanned image. Amounts appear in different places and descriptions use different wording. Anticipating every combination through rules can be difficult.

A workflow can ask an LLM — an AI language model — to read those documents and return the supplier, amount and due date. The same workflow then continues with the planned checks.

There is AI, but not necessarily an agent: the model interprets information at one specific step without deciding how the process should continue.

Fig. 2Hybrid workflow: AI completes a specific task

01 Documents
Different formats.
02 LLM
Interprets the data.
03 Workflow
Applies rules.
04 Person
Reviews.
05 Result
Reviewed file.

A workflow can also call an LLM to interpret information. AI returns the result of that specific task; the rules still control the process.

An AI agent: give it an objective and the means to progress

An agent uses an AI model to choose and carry out actions. We give it an objective, instructions, context and access to tools. We also define limits: what it may do and when it must ask for help.

It acts, observes the result and decides what to do next. That is its autonomy. The difference between an agent and a workflow is who decides how to proceed.

Return to the RFP search from the previous edition. The objective might be to find opportunities that fit the company’s services.

The agent finds a tender and discovers that its terms are in an attachment. It consults the file. If it refers to a later clarification, the agent searches for that clarification before assessing the opportunity. Finally, it prepares a shortlist with sources, deadlines and open questions for the team.

The next step depends on what it finds. If it only reviewed fixed pages and always completed the same fields, a workflow with AI could be enough. The difference is not searching the internet; it is adapting the route according to the results.

Fig. 3Agent: the objective guides the route

01 Objective
Desired outcome.
02 Choose action
Agent.
03 Use tool
Permitted capability.
04 Check result
Agent.
05 Continue?
Agent decides. Yes Chooses the next action.
No06 Stop or ask for help

The agent chooses an action, uses a tool and checks the result. What it finds informs the next action, within the defined objective and limits.

The difference at a glance

	Automated workflow	AI agent
What do we define?	Steps and conditions.	Objective, instructions and limits.
How does it proceed?	It follows predefined routes.	It chooses actions based on what it finds.
Can it use AI?	Yes, at specific steps.	Yes, including to decide how to proceed.

Both require controls and an accountable person. A single process can combine them: rules to record information, AI to interpret it and an agent to investigate what is missing.

Autonomy must justify its cost

You would not hire an experienced lawyer simply to deliver documents to a court. Nor do you need an agent for a task whose route is already defined: that autonomy must justify its cost.

Each request to a model processes content measured in tokens: small units of text. With usage-based services, that consumption affects the price. An agent may make several requests to decide, act and review. It therefore tends to cost more than running simple code without AI.

Four questions to help choose the right approach

Can we define the steps and rules? Start with a workflow.
Is the format of the information the only thing that changes? AI within the workflow may be enough.
Does the next step depend on what is found, beyond what predefined routes can accommodate? Consider an agent with a clearly bounded scope.
Can we check the result, and do we know who is accountable if it fails? Establish that before delegating.

You do not need to know how to build a workflow or an agent. You do need to understand when each is appropriate. That judgement helps you evaluate your team’s proposals and allocate budget according to the problem they solve and the value they create.

Until the next issue.