What is an AI agent and where is it used

One thing separates an agent from a chatbot and from a fixed workflow: who chooses the next step. Start the trial with narrow permissions, a log, a way to undo and a person’s approval.

· 14 min readAI

Short answer

An AI agent is a system that picks its own next step by looking at a goal, rather than following a path drawn in advance, and uses tools to take it. In a fixed workflow you decide the steps. Try an agent with narrow permissions, a log, a way to undo and a person’s approval on risky steps.

Picture this: a small online shop (the firm in this article is invented and describes no real business). In the inbox this morning, a handful of customer emails say “where is my order?”. One vendor offers you an “AI agent”: it will read the email, look up the order record, write the reply and, if needed, start a refund. Another vendor uses the same word to sell the chat box on your site. So the question of what an AI agent is comes first as a question of definition, and then as a question of permissions: which job will it do, with which tools, and with whose approval? This article takes the definition from the model providers’ own writing and explains how to run a small first trial that is logged and reversible.

Agent, chatbot and workflow: three different things

There is no single definition of the term. Anthropic’s engineering article says that some customers call fully autonomous systems that run independently for long periods agents, and others use the word for more prescriptive implementations that follow predefined workflows. Anthropic calls all of these “agentic systems” and draws an architectural line between them: in workflows, language models (the AI models that produce text) and tools are run through code paths written in advance. In agents, the language model directs its own process and its use of tools as it goes. (The article is dated December 2024, and the page itself notes that much of the tooling it describes has changed since. We took only the definition and the principles from it.)

OpenAI’s guide also defines an agent, as a system that independently accomplishes tasks on your behalf. By the guide’s account, an application in which the language model does not control how the workflow is carried out is not an agent. That covers a simple chatbot, a single one-off question to a model or a classifier that sorts comments into positive and negative. The practical test that follows from the two definitions is one question: who chooses the next step? The table below is our own summary of that test; the examples belong to the invented shop.

Chatbot Fixed workflow Agent
Who chooses the next step? There is no step; it writes an answer to a question You, drawn in advance The model, by looking at the goal
Example Tells the customer the return terms in the chat box When a new order arrives, writes the record into the accounting program Reads a tangled complaint, checks the order record and suggests what to do

We have a separate guide to a chat assistant that builds its answers from your own documents: how to set up an AI assistant (chatbot). What a fixed flow is can be read under workflow in the glossary and in the article on your first automation workflow. A short definition is in the glossary entry for AI agent.

The four parts of an agent

OpenAI’s guide lists three components at the base of an agent: a model, tools that act on outside systems, and instructions that define its behaviour. Anthropic says agents are, most of the time, language models using tools in a loop on feedback from their environment, and it mentions adding stopping conditions, such as a maximum number of rounds, to keep control. We gather what a business owner should ask into four headings; this set of four is our own framework:

  1. Goal. The job in one sentence, and the measure for “done”: “identify the topic of each incoming complaint email and prepare a draft reply.”
  2. Tools. The operations the agent can touch: reading email, looking up an order record, writing a draft. These connections between programs are usually made through an API, a doorway through which one program talks to another; we explained what that is in what an API and a webhook are.
  3. Data it can see. Which records and files it can read. For systems that answer from documents, our RAG article explains how this permission is limited at index level. In OWASP’s example, an agent that makes recommendations from a product database needs read access to the products table only, with no right to insert, update or delete.
  4. The limit on what it does alone. Which operations depend on a person’s approval and how many steps the agent may take. OWASP says the root cause of an agent being able to take a damaging action is typically one or more of three things: more functionality than needed, more permissions than needed and more autonomy than needed.

Which jobs suit an agent and which suit a plain workflow

Anthropic’s advice is plain: find the simplest solution possible and add complexity only when needed, which may mean not building an agentic system at all. Anthropic says workflows offer predictability and consistency for well-defined tasks and agents offer flexibility, and it adds that agentic systems often accept slower responses and higher cost in exchange for better task performance. It considers agents suited to open-ended problems where it is hard to predict the number of steps and a fixed path cannot be set in advance.

OpenAI’s guide makes this concrete with three signs: work involving nuanced judgement and exceptions (its example is refund approval in customer service), rule sets that have become hard to maintain, and work that means interpreting written text or documents. Without these, the guide says, a rule-based solution may suffice. The table below sorts five jobs of the invented shop by those tests; the “our suggestion” column is our own reading.

Job Are the steps known in advance? Our suggestion
Writing a new order into the accounting program Yes, the same every time Plain workflow
Sending a fixed-text notice when an order ships Yes Plain workflow
Preparing a draft reply to “where is my order?” from the shipping record Partly; the question varies The agent writes the draft, a person sends it
Reading a tangled complaint email and suggesting its topic and next step No, every email is different Agent candidate, with a person’s approval
Deciding a refund amount and making the payment Needs judgement and money leaves At most the agent prepares a suggestion; a person approves the payment

To sort your own jobs you have to write them down by hand; we described how in where to start with automation, and you can find fixed-flow examples in business process automation examples.

Where a person must approve

OpenAI’s guide recommends human intervention in two cases: when the agent exceeds a set limit on retries or actions, and on risky steps. It describes a risky step as sensitive, irreversible or high-stakes, says such steps should depend on human oversight until confidence in the agent’s reliability grows, and gives examples such as cancelling orders, authorising large refunds and making payments. The guide also suggests rating each tool as low, medium or high risk by read or write access, reversibility, account permissions and financial impact. OWASP likewise asks for a person’s approval before high-impact actions.

We bring these tests together in the table below; the rows and approval rules are our own suggestion:

What the agent does Example Approval
Reads Looks up an order record Not needed; what it read goes in the log
Prepares a draft A draft reply to a customer A person reads it before it is sent
Writes a reversible record Adds a tag to a customer record Pick samples and check them in the first weeks
Does something that cannot be undone or involves money Sends an email, refunds, pays, deletes A person approves every time

Our service page says the same thing: we measure which tasks can be safely automated, and keep human approval where it is needed. The figure below shows how a single step passes through these rules; we will set up the log and rollback boxes in the next section.

One step of an agent: goal, the agent choosing a step, risk and limit check, a person approving a risky step, an allowed tool doing it, log and rollback Goal One-sentence job Agent chooses its next step by itself Check High-risk step? Limit reached? A person decides Approves or stops it Uses a tool Only the tools it is allowed Log What it saw, what it did Rollback If something goes wrong, check the log and undo it Yes No OK One step of an agent: goal, the agent choosing a step, risk and limit check, a person approving a risky step, an allowed tool doing it, log and rollback Goal One-sentence job Agent chooses its next step by itself Check High-risk step? Limit reached? A person decides Approves or stops it Uses a tool Only the tools it is allowed Log What it saw, what it did Rollback If something goes wrong, check the log and undo it Yes OK No
Figure: At every step the agent first passes a risk and limit check; a person approves a risky step, only allowed tools carry it out, everything is written to a log and can be rolled back if something goes wrong.

A first trial in six steps

The steps below are our own suggestion and rest on the principles in the sources above. The aim is not to put the agent to work at once but to see, through a log, how it behaves in a small area.

  1. Pick one job. Choose one where a wrong result does little harm and where you know how long it takes by hand; in the invented shop, “sort a complaint email and write a draft reply”.
  2. Write down the limit. Put the four parts above on one page. Let the first trial be read and draft only, and write how many steps it may take on one job.
  3. Open a log. Record at each step what the agent saw, which tool it called, what it produced and who approved it. Anthropic counts transparency among its three principles and asks agents to show their planning steps; OWASP says logging and monitoring help limit the damage.
  4. Build the way back first. For everything the agent writes, there should be an answer to “how do we return it to how it was?”. If there is none, that operation stays with a person.
  5. Let a person see every output for two weeks. OpenAI’s guide says human intervention helps find failures and rare cases, especially early in deployment. On our service page, an automation pilot also lasts two weeks, and every flow goes live with an error alert and a log.
  6. Make a written decision to continue. Note before the trial how long the job takes by hand and how many errors come up, and take the same measures at the end. If the result is not good, going back to a fixed workflow or to doing it by hand is a valid outcome too.

Three things that can go wrong

A wrong action. Anthropic writes that the autonomy of agents brings higher costs and the potential for compounding errors, and it recommends testing extensively in an environment that does not touch real systems, with guardrails. By OWASP’s account, a damaging action can occur in response to an unexpected, ambiguous or manipulated output from the model.

Made-up facts. OWASP lists the model making things up (hallucination) among the common triggers of excessive agency. The KVKK guide on generative AI also describes hallucination as content cut off from context or untrue. It says that wrong information in outputs containing personal data can harm the people concerned, so outputs should be monitored regularly and supported by human oversight where needed. The term is defined under hallucination.

Instructions hidden in what it reads. OWASP calls this “indirect prompt injection”: when a model takes input from outside sources such as websites or files, data in that content can change the model’s behaviour in unexpected ways. The effect does not have to be visible to a human; it is enough that the model processes it. In one OWASP example, an assistant that summarises emails has a send function next to its permission to read, and a maliciously crafted email steers the assistant into searching the mailbox for sensitive information and forwarding it to an address. As prevention it names a tool that only reads mail, a read-only permission and a person’s approval of every drafted message before it is sent. OWASP’s prompt injection page, for its part, says it is unclear whether any fool-proof way of preventing this exists; it lists only measures that reduce the impact. Our own conclusion: do not give an agent that reads customer emails, attachments or web pages the right to send, delete or pay on day one.

On personal data we can only repeat what the KVKK guide says. It advises users to take care not to share identifying data such as name, address, phone number and identity details, or other people’s personal data, with these systems, and to prefer anonymised and generalised wording where sharing is needed. If an agent is going to read customer records, take the legal assessment to your own lawyer. In our automation work we start by drawing up a data map that shows which data passes through which tool.

Questions to ask whoever offers to build one

If an offer comes in, ask these in writing.

  • If the steps in this job were known in advance, would a fixed workflow not be enough?
  • Which tools will the agent connect to, with what permissions; can it only read?
  • Which operations depend on a person’s approval, and how is it given?
  • Will I be able to see the log of every step?
  • How is a wrongly done piece of work undone?
  • If the agent reads customer emails or web pages, what is done about instructions hidden in the content?
  • Which services does my data pass through, and in whose name are the accounts opened?

The last question is covered in more detail in how to choose an automation agency; our approach is on the automation and AI service page.

Common mistakes

  • Taking that it worked in a demo to mean it works in real work. Do not run the first trial on real customer records.
  • Taking a confident tone as proof of accuracy. Made-up information is written fluently too; checking the source and asking a person on risky steps can catch it.

Common questions

Does a small business need an agent?

Not necessarily; according to Anthropic, the simplest solution may be not to build an agent at all. A workflow is enough for jobs with fixed steps; try an agent when a job whose steps are not known in advance and which needs judgement is really troubling you.

What happens if an agent sees customer data?

That needs a legal assessment, and your own lawyer should make it. The KVKK guide advises avoiding sharing identifying data with these systems.

Does an agent notice and fix its own mistakes?

The sources say so in part. Anthropic describes agents moving forward by checking tool results at each step to see what actually happened, and pausing for human feedback when needed. OpenAI writes that an agent can correct its actions and, on failure, halt and hand control back to the user. The same Anthropic article also says errors can compound. That is why the log and the way back are not left to the agent’s self-correction.

The smallest thing you can do today: write down five jobs you do by hand and, next to each, note who chooses the next step: you, a rule, or judgement every time. The jobs that say “rule” are candidates for a fixed workflow. Of the ones that say “judgement”, pick the one where a wrong result would do the least harm and take the first of the six steps above.

Sources

This article is general information, not legal, financial or technical advice. Platforms, their rules and the law change over time, so some details may have changed since 14 November 2026 or contain small errors. Before you make a decision for your business, check the current official sources or write to us.

Tell us about
your project.

Fill in the form or write to us directly, and we will agree on a day to talk.

Your first meeting is with Deniz Sezer, founder of Detmore.

First meeting free · Reply within one working day · Written proposal in 2 to 3 days · NDA ready

What do you need? pick as many as you like

Name, email and phone are required.

Contact

We reply within one working day.