Once a language model appears in a process, it is easy to describe the whole solution as an AI agent-even when the model is not making any meaningful decisions.
It might only classify an email, extract information from a document, or prepare a draft response. The rest of the process can still follow rules defined in advance.
An agent begins where the model is given control over how the task should be completed.
That distinction matters. The more decisions we delegate to the model, the more possible paths we need to test, secure, and monitor.
Good automation does not use the most advanced technology available. It uses the least complex solution that can handle the process effectively.
Workflow, AI-powered workflow, and AI agent
For this article, I will use the following definitions:
Workflow - code and rules determine which actions are performed and in what order.
AI-powered workflow - the path remains predefined, but a model handles selected steps that require interpretation.
AI agent - the model can decide which actions to take, in what order, and which tools to use.
A workflow might download an invoice from an email, store the attachment, extract selected fields, and send the data to an accounting system. The sequence is designed in advance.
If extracting the fields from different document formats requires an AI model, it becomes an AI-powered workflow. The model interprets the document, but it does not decide what happens next.
An agent could go further. It might identify the document type, decide what information needs to be checked, select an appropriate system, and choose the next action based on the results.
These approaches are not mutually exclusive. A single business process can combine rules, individual model calls, and a narrowly scoped agent. The important question is where the path is predefined and where the system has autonomy.
An agent is not simply a more advanced workflow. It is a system that has been given responsibility for some of the decisions about how the task should be completed.
Anthropic makes a similar distinction: workflows follow paths defined by code, while agents dynamically direct their own processes and tool usage. It also recommends adding complexity only when simpler approaches no longer produce sufficient results.
The key question: who chooses the next step?
Using AI does not automatically make a system an agent. Neither does the number of actions involved.
A process can contain twenty steps and still be a standard workflow if every step is defined in advance. An agent might perform only three actions but choose those actions dynamically based on the situation.
Several characteristics help determine which approach is appropriate:
- Predictability - can the path be described before the process starts?
- Ambiguity - can the input be handled with clear rules, or does it require interpretation?
- Exceptions - how often does the process move outside its standard path?
- Risk - what happens when the system makes the wrong decision?
- Reversibility - can an incorrect action be easily undone?
When the sequence is known, a workflow is usually enough. When only one step requires an understanding of the content, a model can be added to that step. An agent becomes relevant when it is impossible to determine all required actions in advance.
Google Cloud uses similar criteria when discussing agentic architectures, including task predictability, the need for dynamic planning, cost, latency, and the required level of control.
AI does not need to control the entire process
In many automations, AI is only required at a specific point-usually where the input varies or its meaning cannot be captured with a reliable set of rules.
A model might:
- classify an incoming message;
- extract information from a document;
- assess the priority of a request;
- summarize a case;
- prepare a draft response.
The rest of the process can remain deterministic. Code stores the information, applies a label, creates a task, sends a notification, or assigns the case to a specific person.
This is often a better option than handing the entire process to an agent. We know when the model is used, what output it should return, and what the system is allowed to do with that output.
AI should handle ambiguity, not replace rules that already work.
A single model call is also easier to test. We can prepare a representative set of emails or documents, measure classification quality, and determine the confidence threshold below which a case should be reviewed by a person.
A shared inbox does not need an AI agent
Consider a service company that receives approximately 2,400 messages per month through a shared inbox. An employee opens every email and routes it to sales, accounting, complaints, or administration.
It would be possible to build an agent that analyses each message, checks the customer in the CRM, chooses the next action, and replies without human involvement. But that level of autonomy might not solve an actual problem.
A closer look at the process could show that around 45% of messages can be handled with simple rules:
- messages sent to the
invoices@company.comalias go to accounting; - emails from known suppliers go to predefined folders;
- system notifications use a predictable subject format, such as
[ORDER #1234].
The remaining messages can be sent through a single model call. The model does not reply to customers or decide which business systems to search. It returns only: a category, a priority, a confidence score.
When the confidence score exceeds a defined threshold, the workflow applies a label or moves the email to the appropriate folder. The remaining cases are assigned to a person.
Under these assumptions:
- 1,080 messages are handled by rules;
- approximately 1,050 messages are classified by the model;
- approximately 270 messages require manual review.
If manually classifying one email takes 1.5 minutes, processing all messages takes about 60 hours per month. Reviewing 270 uncertain cases takes less than seven hours of classification work.
The actual time may be higher after accounting for rereading messages, correcting classifications, and handling exceptions-perhaps nine to eleven hours. That number is still an assumption and should be measured against the real process.
From a technical perspective, the solution could use Gmail filters and Apps Script, Microsoft Graph for Microsoft 365, or a small application connecting to a standard mail server through IMAP.
This is still an AI-powered workflow, not an agent. The model performs a constrained classification task. It does not plan the work, select tools, or decide how to achieve a broader objective.
When does an AI agent make sense?
An agent becomes valuable when the path to the result cannot reasonably be described before the work begins.
In-depth research is a good example. A system may begin with several sources, discover new issues, formulate follow-up questions, and decide where to search next. The full path cannot be known in advance because it depends on what the system finds along the way.
The same can be true for:
- diagnosing a failure by choosing successive tests;
- modifying code across an unknown number of files;
- investigating a customer case using several business systems;
- resolving an unusual request that does not follow one predefined procedure.
In each example, the value does not come from using a model alone. It comes from the system’s ability to adjust its plan based on intermediate results.
An agent makes sense when the system must decide not only what needs to be done, but also how to do it.
This does not require unlimited freedom. A production agent should have a clear objective, a limited set of tools, access only to the data it needs, and explicit stopping conditions.
OpenAI also recommends an incremental approach rather than starting with a fully autonomous, complex architecture. Agents are particularly relevant when a process involves ambiguous decisions, unstructured data, or a large number of rules and exceptions that would be difficult to maintain in code.
More autonomy means more control work
A workflow follows paths anticipated by its designers. An agent can create paths that were never written down explicitly.
That gives the system more flexibility, but it also expands what needs to be controlled.
An agentic solution requires attention to:
- Tool selection - did the agent use the correct system and operation?
- Sequence - did it perform an action too early?
- Stopping conditions - can it end the task instead of repeating unsuccessful attempts?
- Permissions - can it only read information, or can it change it?
- Failure handling - what happens when an external system returns incomplete or conflicting data?
- Escalation - when should the task be handed over to a person?
A model may complete nine steps correctly and fail on the tenth. Errors can also compound: an incorrect interpretation leads to the wrong tool, which produces information used in the next decision.
This is why checking the final output is not enough. The team also needs to understand how the result was produced: which information the agent received, which tools it called, and which actions it took.
The greater the impact of an error, the narrower the agent’s authority should be.
A person does not need to approve every action. Human review should focus on unusual exceptions, financial and legal decisions, external communication, and operations that are difficult to reverse.
Hybrid architecture is often the practical answer
The choice is not always workflow or agent.
Responsibility can be divided within a single process:
- Rules handle unambiguous cases.
- A workflow controls the process and integrations.
- AI interprets unstructured information.
- An agent handles a limited area that requires dynamic planning.
- A person makes high-risk decisions.
For example, a workflow can receive a complaint, check whether the required information is present, and retrieve the order history. A model classifies the issue and prepares a summary. An agent is triggered only when the case requires additional information from several systems. An employee makes the final decision about a refund.
The agent does not control the entire process. It receives a defined objective and operates within explicit boundaries.
A hybrid design is not always necessary. When rules or a single model call solve the problem, adding more layers only increases cost and operational effort.
How to choose the right approach
The decision can be reduced to five steps:
- Define the expected result - establish what should be true when the process is complete.
- Identify deterministic steps - mark the actions that can be performed using clear rules.
- Find interpretation points - locate the parts that require an understanding of text, documents, or context.
- Assess the need for dynamic planning - determine whether the system must choose its own next actions.
- Limit autonomy - delegate only the decisions that cannot reasonably be handled in a simpler way.
The number of steps does not determine whether you need an agent.
Start with a workflow. Add AI where interpretation is required. Use an agent only when the value of the process depends on the system choosing its own next actions.
