Guide

What is an AI agent?

10 min read

By . Published . Last updated . Reviewed by Mavrin Labs on . How this guide was researched and checked.

An AI agent is software given a goal, a set of tools and permission to choose its own next step, within limits a business sets and with a person approving what matters. This page defines the term for owners and managers of small businesses and nonprofits, explains how an agent works and sets out the limits that keep it safe.

It leaves out whether to build one, which the guide to custom AI agents answers, and choosing a model or the investment, which depend on the task and are settled after a look at the work, as how an engagement works explains.

What is an AI agent?

An AI agent is software that reads a request, decides its next step toward a goal within set limits, and takes it with the tools it may use, while a person approves what matters. The limits are part of the definition, not an extra: without them, the same software is simply acting on its own.

Every agent has four parts, and each one is a decision someone makes on purpose:

  • The goal. What finished looks like, written so another person could check it.
  • The information. The records and documents the agent may read, listed one by one.
  • The tools. The actions it may take, such as looking up a record or drafting a reply.
  • The person. Who approves what cannot be undone and who receives what the agent cannot resolve.

The word agentic describes AI that takes steps on its own toward a goal rather than only answering. Related terms have their own pages: the plain definition of AI automation, and how a CRM (the program that holds client and prospect records) connects to other programs, on the page defining CRM integration.

Why it matters: reported AI incidents keep rising

AI use is growing from a low base, and so are the problems reported with it. The Census Bureau's business trends and outlook survey put business AI use between 17 and 20 percent from December 14, 2025 to May 3, 2026 (Census Bureau report on business AI use, checked Tuesday, November 17, 2026).

Stanford's 2025 AI Index reports that AI-related incidents in the AI incident database it tracks rose to 233 in 2024, a 56.4% increase over 2023 (Stanford AI Index, responsible AI chapter, checked Monday, November 16, 2026).

An agent can act, so a mistake does more than produce a wrong answer on a screen. The limits described below decide how far one mistake can travel.

How does an AI agent work?

An AI agent works in a loop. It reads the goal and what has arrived, looks up what it needs, chooses one tool, uses it, checks the result against the goal, and either takes another step, stops because the goal is met, or hands the work to a person because a rule says so.

Each turn of the loop follows the same order:

  1. Read the request. The agent starts from what it was asked to finish.
  2. Look things up. It reads the records and documents it is allowed to see.
  3. Choose a tool. It picks the next action from the tools it was given.
  4. Act and record. It uses the tool and writes the step to a log.
  5. Check the result. It compares where it is with the goal.
  6. Stop or hand over. It finishes, loops again or passes the case to a person.
Read the requestLook things upChoose a toolAct and recordA person approves anything thatcannot be undoneCheck the resultStop or hand overA rule sends the case to apersonIf the goal is not met, the loop runsagain from the look-up
The agent's loop. A person approves any action that cannot be undone, and the hand-over sends a case to a person; both are filled.

The choosing in the third step is usually done by a large language model (the kind of AI model behind chat assistants, trained on large amounts of text to predict and write language). That is what makes an agent flexible, and also why its behavior has to be tested against agreed cases before anyone relies on it.

How is an AI agent different from a chatbot and from an automation?

An AI agent differs from a chatbot because it acts as well as answers, and from an automation because it chooses its own steps instead of following a fixed sequence. A chatbot replies to the person typing; an automation runs the same steps every time; an agent decides which step comes next, within the limits it was given.

How is an AI agent different from a chatbot and from an automation?
Chatbot Automation AI agent
What starts it A person typing A trigger, such as a form or an email A goal and a request
Who sets the steps Nobody: it replies The builder, in advance The agent, step by step, within limits
Can it act in other systems Rarely Yes, in fixed places Yes, with the tools it is given
How it is tested Sample conversations Each step against examples The goal, the limits and every tool
Best for Answering questions Identical, repeat steps Work that varies case by case

The common question of AI agents vs chatbot tools comes down to that one difference: action. The same holds for AI agents vs assistants, since an assistant waits to be asked. A task that repeats identically is better served by an automation, which is steadier and easier to check.

What can an AI agent do well, and what should it not do alone?

An AI agent does well on work that varies case by case but has a known purpose, such as gathering facts from several places or preparing a case for a person to decide. It should not act alone on anything that moves money, changes a legal record, or speaks for the business in a way that cannot be undone.

Work an agent can take on:

  • Reading an enquiry and working out which kind of request it is.
  • Collecting details from two or three systems before a decision.
  • Checking a document against a record and listing mismatches.
  • Drafting a reply or a summary for a person to approve.
  • Chasing missing information with a standard request.

Work that always waits for a person:

  • Sending money or changing payment details.
  • Changing a contract, a legal record or a filed figure.
  • Telling a client something new about their account.
  • Deleting records or anything that cannot be undone.

The guide to custom AI agents covers when building one is worth it, and the guide to testing an AI agent before clients see it covers the acceptance tests.

What limits keep an AI agent safe?

The limits that keep an AI agent safe are the smallest set of tools its job needs, approval before any action that cannot be undone, a log of every step, and a clear rule for when it stops and hands the work to a person. Public guidance reached that list in three steps, each narrower than the last.

  • How the guidance developed

    NIST's generative AI profile, AI 600-1, from July 2024, sets out 12 risks generative AI creates or worsens, including confabulation (NIST AI 600-1, checked Wednesday, November 18, 2026). Confabulation is the profile's name for what many call a hallucination: an AI answer that sounds confident but is not true.

    The OWASP Top 10 for large language model applications, 2025 edition, replaced a list for 2023 and 2024 (OWASP Top 10 for LLM applications, checked Wednesday, November 18, 2026). It lists prompt injection (hidden instructions in text the AI reads that try to make it ignore its rules) first.

    It names excessive agency, an agent given more tools or permission than its job needs, as a separate risk.

    A guide from CISA and international partners on agentic AI, published Friday, May 1, 2026, advises low-risk first uses and no broad access to sensitive data (CISA announcement of the agentic AI guide, checked Wednesday, November 18, 2026).

    Read in date order, the guidance moved from the risks of AI that writes, to the risks of AI that acts, to how a business should adopt agents at all.

  • Defaults for the limits

    In practice the limits become defaults. Lookup work gets read-only access: OWASP's entry on excessive agency uses a read-only scope as its example, and advises the minimum permissions and a person's approval for high-impact actions (OWASP on excessive agency, checked Wednesday, November 18, 2026).

    The log is the last default. OWASP advises logging what an agent and the systems it touches do, but sets no period, so this page's default is to keep each step log as long as your retention policy keeps the records the agent read or changed.

How does Mavrin Labs build AI agents?

Mavrin Labs builds AI agents as one of its services, starting with a workflow assessment that maps the work and agrees the result before anything is built. The agent's goal, information, tools and approval points are written into a design plan, and acceptance tests are agreed before the build begins.

By default nothing is sent without the client's approval, and client data is not used to train AI models. The client owns its custom work once paid for, and all of its data.

A care plan after launch includes monitoring, fixes and a monthly review against the baseline.

The custom AI agents service describes the work in full, the AI automation service covers repeat work with fixed steps, and how an engagement works explains each stage.

Key takeaways

  • An AI agent works toward a goal with tools and permission to choose its next step within limits.
  • It differs from a chatbot because it acts, and from an automation because it chooses its own steps.
  • It suits work that varies case by case, and it never moves money or changes records alone.
  • The limits that keep it safe are few tools, approval for what matters, a full log and a clear stop.
  • Reported AI incidents keep rising, by the count in Stanford's AI Index.

Where to go next

Before building one, read how to test an AI agent before it reaches your clients. Every guide in the set is listed in the resources library. To talk through one workflow, start a conversation.

Whether a first project needs an agent at all is what the six-question project finder helps settle, from a few questions about where the week goes.

Frequently asked questions

Is an AI agent the same as agentic AI?

An AI agent and agentic AI describe the same idea from two sides. Agentic describes AI that takes steps on its own toward a goal rather than only answering, and an AI agent is one piece of software built that way. A business buys or builds an agent; agentic is the quality the agent has.

Can an AI agent use the systems a business already runs?

An AI agent can use the systems a business already runs when they offer an API (a published doorway one program offers so another can read or send its data). The agent is given narrow permission for the records and actions its job needs. A system with no such doorway limits what an agent can do there.

What can go wrong with an AI agent?

An AI agent can go wrong in four ways: it can state something untrue with confidence, follow instructions hidden in text it reads, repeat an action or act beyond its job. OWASP's 2025 list of risks for language-model applications ranks the hidden-instruction attack, prompt injection, first. Fewer tools, approvals and a full log limit each risk.

Who is responsible when an AI agent makes a mistake?

Responsibility for an AI agent's mistake stays with the business that runs it, which is why the approval points and the log matter. The business decides which actions a person signs off and keeps a record of every action the agent took, so a mistake can be traced and put right.

Does an AI agent learn from a business's data?

An AI agent built for a business uses that business's records to do its work, but using a record is not the same as training a model on it. Whether data is used for training depends on the terms of the AI provider. Mavrin Labs does not use client data to train AI models.

The Mavrin Labs mark

Who writes these guides

Client work at Mavrin Labs is delivered by a team in mixed roles, led by our founder. AI tools help with research and drafting, and a person at Mavrin Labs checks each fact against its named source before a guide is published.

Enquiry form

Start with one workflow

Tell Mavrin Labs what is stuck and which systems it touches. A person at the firm reads every enquiry and replies by email.

Start a conversation

Fields marked with an asterisk are the ones a reply needs.

So a reply can address you properly.

The address the reply goes to.

Optional. Give one only if you would rather talk than write; the reply goes by email either way.

Optional, and useful when it says what the work has to fit around.

The result you want, and when you would want it working. Two or three sentences are plenty.

The task as it actually happens, roughly how often it runs, and which systems it touches. Two links at most, please.

Optional.

What you send goes to the Mavrin Labs inbox and is used to answer you, and a phone number, if you give one, is used for the same thing and nothing else.

A person reads it and replies by email.

If the form does not send, write to info@mavrinlabs.com instead; it reaches the same inbox.

Which hours would you take back first?

Name the one workflow that eats the most of your week. Mavrin Labs reads it, and the reply comes from the people who would build the fix.

Start a conversation