Guide

AI agents for small business: what one is and when to build it

17 min read

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

A card reading "AI agents for small business: what one is and when to build it" beside the Mavrin Labs logo
The card that shows when this guide is shared or linked.

A custom AI agent is software given a goal, a set of tools and permission to choose its own steps. For a small business, one is worth building when the work varies case by case, you can say what a correct outcome looks like, and you can name the hours it would give back.

The guide is written for owners and managers of small firms weighing whether an agent would help. It covers what an agent is and when one is worth building. Choosing an AI model, comparing suppliers and agreeing the investment are left to a later conversation.

Which work can an AI agent take off the week?

An AI agent can take on the work that fills a small-business week without being either difficult or identical. Somebody reads an enquiry and works out which of four things it is, or gathers details from three systems before anyone can decide. Somebody else checks a document against a record and writes to ask about a mismatch.

Plain rules handle none of that well, because the shape changes each time. A person handles it easily and slowly.

The gap between the two is where agents are genuinely useful. The gap is also why they are harder to judge than an automation: the work they take on is exactly the work nobody wrote down.

Most small firms do not use AI yet. A national survey of businesses by the Census Bureau found that fewer than one in five firms with four or fewer employees used AI in the period ending May 3, 2026, against 37% of firms with at least 250 employees (Census Bureau report on business AI use, checked Saturday, October 10, 2026).

What is a custom AI agent?

A custom AI agent is software built for one business that is given a goal, a set of tools and the freedom to choose which tool to use next, within limits the business sets. Unlike a chatbot, it acts as well as answers.

The definition of an AI agent on its own page covers the general term, and this guide covers the custom kind. The usual question of AI agents vs assistants comes down to the same point: an assistant answers when asked, while an agent acts.

The agent works towards the goal across several steps, decides what to do at each one, and stops when the goal is met or when a rule says to hand the work to a person. The word custom matters.

An agent built for your business knows your definitions, works inside the systems you already run, and is limited to the actions you agreed. A general assistant has none of that context, which is why it gives plausible answers to questions about your business and gets the specifics wrong.

The parts are easier to understand separately than together.

The goal
What finished looks like, written precisely enough that somebody else could tell whether it was reached.
The information
The records, documents and systems the agent may look at, listed deliberately rather than granted wholesale.
The tools
The specific actions it may take, such as reading a record, drafting a reply or creating an entry, each one chosen.
The person
Who approves the actions that cannot be undone, who receives what the agent cannot resolve, and who reads the log.
TheagentThe goalWhat finished lookslikeThe informationRecords it may read,listedThe toolsActions on theallow-list onlyThe personApproves, andreceives hand-offs
The four parts of an agent. The goal and the information flow in, the tools act out, and the person works both ways: approving and receiving hand-offs.

Why it matters: agents widen what can go wrong

An agent acts, so its mistakes travel further than a wrong answer on a screen. Stanford's 2025 AI Index recorded 233 AI-related incidents reported in 2024 to a public database that tracks them, 56.4% more than in 2023 (Stanford's AI Index, responsible AI chapter, checked Monday, November 16, 2026).

The National Institute of Standards and Technology (NIST), the federal standards agency, published a generative AI profile in July 2024 that sets out 12 risks generative AI introduces or makes worse (NIST AI 600-1, checked Wednesday, November 18, 2026).

Three of those risks, information security, human-AI configuration, and value chain and component integration, bear directly on an agent wired into your systems.

Each one becomes a design question: what the agent may touch, who approves its actions and what the log records.

What do AI agents do well, and when should you not use them?

Agents do well at work that varies in shape but not in purpose. That means sorting mixed input, gathering information from several places, drafting something a person will approve, and checking one record against another. Do not trust them with judgment calls, with irreversible actions, or with anything a mistake would make expensive and invisible.

The honest version of this is a short list on each side, and the right side of the list is the one that protects you.

What do AI agents do well, and when should you not use them?
Kind of use Examples
Good use Sorting mixed enquiries, gathering information held in several systems, drafting replies and summaries for approval, checking one record against another, chasing a missing piece of information
Poor use Deciding anything that needs human judgment, sending money, changing a contract, making a promise to a customer, or any action nobody would notice if it went wrong
Always with approval Writing to a customer, altering a financial or clinical record, deleting anything, or taking an action that would be awkward to reverse

Nothing in that middle row becomes safe because the technology improved. The table is about consequence, not capability.

How does a custom agent compare with an off-the-shelf AI tool?

A custom agent follows your work and obeys your rules. An off-the-shelf tool follows a general case and lets you configure the edges. The choice turns on how much of your advantage lives in the specifics of how you work, and on how patient you can afford to be when something needs changing.

How does a custom agent compare with an off-the-shelf AI tool?
What you are weighing An off-the-shelf tool A custom agent
Fit to how you actually work close enough, if your process is common shaped to your steps and definitions
Time before it is useful short longer, because the design is the work
Control over what it may do whatever the supplier allows exactly the actions you agreed
Where your information goes the supplier decides you decide, within your own systems
Behavior when your process changes you wait for the supplier you change the rules
Upkeep included, and out of your hands yours, and visible

Most trades are still early, which argues for the simpler column. In the same national survey, AI use as of May 3, 2026 stood at 39.7% of businesses in the information sector and 33.9% in finance and insurance, against 19.8% of businesses overall (Census Bureau figures by sector, checked Thursday, October 15, 2026).

Outside those two sectors there is rarely a reason to start with the harder of the two options.

Start with the simpler option when it fits. The custom software or off-the-shelf tool guide works through the same choice for software in general.

When is building an agent worth it?

A custom agent is worth building when three statements are true. The work varies enough that fixed rules keep breaking, you can describe a correct outcome clearly enough to test it, and you can name the hours it would give back. If any one of them is missing, wait until it is in place.

  • Varied work: the work arrives in a different shape most times it runs.
  • A testable outcome: you can define a correct outcome well enough to check one.
  • A measured task: you can measure the task's duration and how often it runs.
  • Approval without a queue: a person can approve the key steps without becoming the bottleneck.
  • Open systems: the systems involved let other software read and write.
  • An owner: somebody will own the agent after launch and read its log.

If the work repeats along familiar steps instead, running a defined workflow is the simpler and steadier choice, and the AI automation readiness guide covers the checks for it. Either way, the result is agreed before the build through the measurable results method.

What are the risks, and where does a person stay in control?

The risks of an agent come down to four, and design handles them better than vigilance. An agent can be confidently wrong, act on instructions hidden inside content it was asked to read, repeat an action it has already taken, or keep going past the point where it should have stopped. Limits set in advance keep a person in control.

The Open Worldwide Application Security Project (OWASP), a nonprofit foundation for software security, lists excessive agency as the sixth risk in its 2025 list for applications built on a large language model (the kind of AI model behind chat assistants). Its entry traces damaging agent actions to too much functionality, too many permissions or too much autonomy.

OWASP's advice on excessive agency includes requiring a person to approve high-impact actions before they are taken (OWASP entry on excessive agency, checked Thursday, October 8, 2026).

Three of those four risks have a numbered entry you can design against. Prompt injection is LLM01:2025, excessive agency LLM06:2025, and misinformation, which covers an agent that is confidently wrong, LLM09:2025 (the 2025 list of risks for language model applications, checked Thursday, October 15, 2026).

Six rules for the build

Six rules hold an agent to its limits, and each is written down before the build so it can be tested.

  • Fewest tools: the agent holds the smallest set of tools the job needs, and nothing it might one day need.
  • Approval: every action that cannot be undone waits for a person's approval.
  • Content as information: content the agent reads is treated as information, never as instructions it may follow.
  • A full log: each action is logged, so a mistake can be traced and reversed.
  • A stop rule: a rule says when the agent stops and hands the work to a named person.
  • A reading day: somebody reads the log on a set day, and the reading is part of the job, not a favor.

Three of those rules work best written as parameters, and each value here is a starting default you can change. The tool allow-list names every tool the agent may call; anything not on the list is refused by the system, not by the agent.

The stop condition is written as a test, such as handing over after two failed tries at the same step, or at any case it cannot place. The log review starts weekly and moves to every two weeks only after four clean reviews in a row.

A case arrivesThe agent gathers and draftsUsing only tools on the allow-listThe stop rule is checkedA case it cannot place goes to anamed personA person approvesAnything that cannot be undoneThe action is taken andloggedThe log is read on review day
Where approval and hand-off sit in one case. The filled steps are the two points where a person takes over.

What mistakes do businesses make with a first agent?

The most common mistake with a first agent is giving it a vague goal and broad access, then hoping supervision will catch the errors. A second is choosing work that repeats identically, which a plain automation handles better. A third is commissioning an agent without asking how it was built and tested.

  • A goal too vague to test.
  • More tools and access than the job needs.
  • Irreversible actions with no approval step.
  • Repetitive work that a plain automation would handle better.
  • No named owner reading the log after launch.
  • No questions asked about how the agent was built.

Two public references help with that last question. The Secure by Demand guide from the Cybersecurity and Infrastructure Security Agency (CISA) turns secure-by-design principles into questions a buyer can put to the maker, such as whether it provides a software bill of materials and keeps security logs (CISA guide to buying secure software, checked Saturday, October 10, 2026).

The secure software development framework from NIST, version 1.1 in its publication SP 800-218, groups its practices into four sets that a buyer can hold any builder to (NIST secure software development framework, checked Saturday, October 10, 2026).

The word agentic, a label for AI that takes steps on its own toward a goal rather than only answering, now has a reference of its own.

The agentic security initiative at OWASP published a threat and mitigation reference on February 17, 2025, and calls it the first in a series on agentic threats (OWASP reference on agentic threats and mitigations, checked Thursday, October 15, 2026).

Ask a builder which of those threats the design answers, and how.

What shapes the investment in an agent?

Three things shape what building an agent takes. The first is how clearly you can state the goal, because a vague goal turns into weeks of adjustment. The second is how many systems the agent has to reach and whether each one offers an interface for other software.

The third is how much approval the work needs, since every approval point is a piece of design and a change to somebody's day.

You agree the investment in writing after the first step, once those three are clear. Quoting before that is guessing, and the page on engagement stages sets out when the scope and the investment are settled.

How does Mavrin Labs build an agent?

Mavrin Labs builds an agent in five steps: Assess, Design, Build, Launch and Support. Each step ends with something you can see and approve before the next begins, so the limits, the examples and the review date are agreed while changing them is still easy.

  1. Assess the task

    The task is mapped as it runs today, timed and written down, so there is a baseline to judge the result against.

  2. Design the limits

    The goal, the information, the tools and the approval points are agreed in writing before anything is built.

  3. Build on real examples

    The agent is built against real examples from your own operations, including the ones that go wrong.

  4. Launch in stages

    The agent goes live in stages, with your staff shown what changed and how to step in.

  5. Support and review

    On the review date you agreed, the numbers are checked against the target and what was found is shared.

Mavrin Labs delivers this work as a team in mixed roles, so the people who design an agent are the same people who answer for how it behaves once it is live. The custom AI agents page describes the service.

How has guidance on agent risk developed?

Guidance on agent risk is recent, and the two main public references both took shape from 2023. NIST released its AI risk management framework on January 26, 2023, and added a profile for generative AI on July 26, 2024, which helps organizations identify the risks particular to generative AI (NIST AI risk management framework, checked Thursday, October 8, 2026).

The OWASP work on language model risks began with a small group of security professionals in 2023 and has grown into an open project whose current list is the 2025 edition. The list places prompt injection first and excessive agency sixth (OWASP Top 10 for LLM applications, checked Thursday, October 8, 2026).

Sources behind this guide

The references above carry the evidence in this guide: the Census Bureau survey, Stanford's AI Index, the NIST frameworks and generative AI profile, the OWASP list with its entry on excessive agency, the OWASP reference on agentic threats, and the CISA buying guide.

Each is worth reading before you commission an agent. Where they say nothing, the guide relies on the design rules above.

Each claim and its source

The table below pairs each claim the guide makes from outside evidence with that evidence and its publisher.

Each claim and its source
Claim Evidence Source
Very small firms are the least likely to use AI Fewer than one in five firms with four or fewer employees used AI in the period ending May 3, 2026, against 37% of firms with at least 250 employees (survey figures, checked Saturday, October 10, 2026) Census Bureau
Too much freedom is what makes an agent's actions damaging Damaging actions trace to too much functionality, too many permissions or too much autonomy, so a person should approve high-impact actions OWASP on excessive agency
Public guidance on AI risk is recent NIST released its AI risk management framework in January 2023 and a generative AI profile in July 2024 NIST AI risk framework page
A buyer can ask how software was built Secure by Demand gives a buyer questions to put to the maker before buying CISA Secure by Demand
AI use is concentrated in a few sectors Information and finance report far higher use than businesses overall Census Bureau figures by sector
Each agent risk has a numbered published entry Prompt injection, excessive agency and misinformation each carry an identifier in the current list the list of risks for language model applications
Reported AI incidents rose in 2024 The AI Index counted more reported incidents than the year before Stanford AI Index, responsible AI chapter
Generative AI brings a named set of risks The profile lists the risks generative AI introduces or makes worse NIST generative AI profile
Agent threats have their own reference now An agentic threat and mitigation reference was published as the first in a series OWASP reference on agentic threats

Key takeaways

  • A custom AI agent is software with a goal, tools and limited freedom.
  • Build one only when the work varies, the outcome is testable and the hours are nameable.
  • Always require approval for messages to customers, financial records and deletions.
  • Give the agent the fewest tools the job needs, and log every action.
  • Repeat work with fixed steps suits a plain automation instead.

If you are weighing an agent against a plain automation, start with the readiness checks for automation. If the question is whether to build anything at all, the framework for deciding whether to build sets out the four options and the rule for choosing between them.

If an agent will need to reach several of your systems, the checklist for reaching several systems covers what that takes. Owners who would rather meet in person will find how that works on the Sarasota page.

Work handled this way gives hours back to the people you already employ, without replacing the systems they rely on.

Frequently asked questions

What is the difference between an AI agent and an automation?

An automation runs a sequence somebody designed, from the first step to the last. An agent is given a goal and chooses its own steps towards it, deciding which tools to use and in what order. The practical difference is predictability: an automation is easy to test, while an agent handles variety and is harder to verify.

Does an agent replace a member of staff?

An agent does not replace a member of staff. An agent takes over the parts of a job that are legwork: gathering, checking, drafting, chasing. The judgment, the relationship and the accountability stay with the person. A business that removes the person also removes the only thing standing between a confident mistake and a customer.

What can go wrong with an AI agent?

An AI agent can go wrong in four ways. It can be confidently wrong, act on instructions hidden in content it reads, repeat an action it already took, or stop too late. Each risk has a counter: limit its tools, require approval for what matters, log every action, and define when it hands the work to a person.

How do I keep a person in control?

Keep a person in control by limiting what the agent may do rather than by watching it. Give it the smallest set of tools the job needs, make irreversible actions require an approval, and route anything it cannot resolve to a named person with a note on what it tried. Control that depends on somebody paying attention is not control.

Can an agent work with the systems I already use?

An agent can work with the systems you already run if they let other software read and write through an interface. The usual arrangement gives the agent narrow permission to look things up and write in specific places, while staff keep their own screens. A system with no such interface is a real constraint, not a reason to replace it.

How do I know whether building one is worth it?

Building an agent is worth it when you can name the hours it would give back. You also need to measure the task as it runs now and agree what a correct outcome is; without those three, wait. A task that repeats identically suits a plain automation, which is easier to test and steadier once running.

What does an agent need to be told?

An agent needs four things written down: its goal, the information it may see, the tools it may use, and when it must stop and ask. Those four are the whole design, and writing them is most of the work. The more precisely you describe what finished looks like, the easier it is to tell whether the agent works.

Who is responsible when an agent gets something wrong?

Responsibility for an agent's mistake stays with the business that deployed it, so the approval points matter. Decide which actions a person signs off, and keep a record of every action the agent took so a mistake can be traced and undone. Mavrin Labs agrees in writing, before the build, which actions an agent may take without approval.

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