AI Agents for Business: What They Do, With Examples

An AI agent is a system where an AI model decides for itself which step to take next and uses tools, such as reading files or searching the web, to finish the job you give it. Five systems on this site show it at work. One agent reads job applications and scores them against the listing. A team of agents splits a weekly content job among specialists, and an add-on worker scans competitors and leaves a short report. Another group tests explanations for a result against your numbers, and one agent spots the tasks you keep repeating. Where they produce drafts, you approve them, and in the CV Screening Agent, sending emails starts switched off.

What makes something an AI agent

The line between an agent and a fixed workflow is who chooses the steps. In a fixed workflow, code written in advance sets the order. In an agent, the model chooses the steps itself. Anthropic's engineering team drew this line in a December 2024 post, describing agents as systems where the model directs its own process and tool use, typically using tools in a loop and reacting to what they return.

OpenAI's practical guide to building agents names two traits. An agent uses an AI model to run a workflow and make decisions, and it has access to tools for dealing with outside systems, choosing the right one for the current state of the job.

In the systems below, a tool is an ordinary thing: reading a file, searching the news, writing a draft into a folder.

Reading and scoring job applications

An agent that screens job applications repeats the comparison a recruiter makes for each CV (resume): read it, check it against the listing, write down why. The CV Screening Agent reads the CVs sent in for a listing, whether they are PDF, Word or plain text files. It scores each one against the listing out of 100 and drops anyone who crosses a red line you set, a rule applied before scoring, with the reason written down. It then writes the shortlist into a table and drafts the invitation and rejection emails. A missing requirement lowers the score instead, unless you turn it into a red line.

Sending those emails is a setting that starts switched off, so invitations and rejections stay drafts you read and send yourself. You can switch sending on later.

Age, gender and photo stay out of the score. If a CV mentions them, the agent reads them and ignores them. It gives no credit for a skill the CV does not mention, and it does not guess one from a job title or a company name. When a CV contradicts itself, the agent leaves the candidate in and asks you in the report.

The page states its limits. The score compares what the CV says with what the listing says, so it does not measure real ability, and a badly written CV can push a good candidate down. Reading the candidates just under the cutoff and the report's list of questions is your job.

Splitting a weekly job among a team of agents

A group of agents can divide one job among specialists the way a small company would. The Agent Company template makes Claude Code, the AI assistant it runs on, work like a company. Claude takes the role of the boss. It reads the job, lists what is needed, hires one worker for each need from a roster of writer, illustrator, composer, researcher and reviewer, and gives each a desk of its own. It keeps a log of who did what and when.

Picture a weekly job with a newsletter, an announcement post and a picture. The writer and the illustrator start together because neither waits for the other, and the reviewer goes last. Each worker uses its own copy of the files, so nobody overwrites anyone else, and the copies merge into the main folder when the job ends. Finished work lands in an approval folder. Claude does not send mail, publish posts or make payments. You do that.

The page sets limits up front: at most five workers and two helpers per job, no helpers hiring helpers, and a usage cap per job. If a job would pass the cap, Claude stops and asks.

Adding a worker that scans competitors

Lookout is a worker you add to this company. You hand it the list to follow, and its sources are free and need no key. Each scan leaves a short report in three parts: what the competitors did, which of it landed, and what you should make today. Lookout reads and posts nowhere. Where a source gives no view numbers, it writes "no view data" instead of inventing a count. When it spots an opening, it drafts a post, has the reviewer read it and leaves it in the approval folder. According to the page, nothing goes out before you approve.

Testing why a video, product or email did what it did

When a result surprises you, agents can argue about the cause and check the arguments against your numbers. The Agent Team system puts one question in front of five agents. Four of them each defend a theory. The fifth, a mathematician who acts as the referee, defends none and tests each theory against the figures in a data file built from your own numbers. In round two, an agent whose theory the numbers contradict withdraws, and the theory left standing becomes the answer, written up as a one-page report.

When the data is too thin to test a theory, the referee writes "cannot be separated" and does not turn the gap into weak evidence. The page calls this the most important behavior in the system.

The page ships four ready scenarios. They ask why a video spread, why a product did not sell, why an email went unopened and why an ad got no clicks.

Spotting the work you keep repeating

Agents can also notice your own habits. Skill Hunter reads the last seven days of your Claude Code chat history on your computer and takes only the messages you wrote. It counts which requests came back, on how many days and in how many chats. For each strong repeat, it drafts a skill, a small file that describes how a job gets done. Once a skill is installed, one word is enough to start that job.

According to the page, a job counts as a skill after at least eight repeats, on at least three different days and in at least three separate chats. Debates about decisions such as pricing or strategy do not count, because those are worth thinking through fresh each time. The steps in a draft come from your own corrections during the week. Claude installs no draft you have not approved, and according to the page nothing goes to the internet.

How to tell whether a job needs an agent

A practical way to start is to take one task you already repeat each week and write down its steps. If you can list all the steps in advance, a fixed workflow is the simpler choice. If the steps depend on what the work turns up, the job looks like the five above.

Anthropic recommends finding the simplest solution possible and adding complexity only when needed, which can mean not building an agent at all. It notes that such systems often trade speed and cost for better results, and that agents suit open-ended problems where the number of steps is hard to predict and no fixed path can be written down, as the post explains. OpenAI's guide says agents suit workflows where traditional rule-based approaches fall short.

Keep the draft and approve step for anything that leaves your business. All five systems sit in the free systems library on this site.

FAQ

Is a chatbot the same as an AI agent?

No. OpenAI's practical guide lists simple chatbots and single-turn models among applications that use an AI model without letting it control how a workflow runs, and it says those are not agents.

How many agents should a small business start with?

The Agent Team page suggests three: two theories and the referee that tests them. Its own run used five agents at the same time, and the page says that costs roughly five times what one agent costs.

What if the roster has no role for my job?

In Agent Company, Claude writes a new worker from a hiring template when no role on the roster fits. You can also add one yourself with the hire command, for example hire translator. Each worker gets a written job description, limits and a list of the moments when it must ask you.

Does the CV Screening Agent upload CVs to a website?

No. According to its page, CVs are personal data, so they stay on your computer and are uploaded to no site. When the role closes, a cleanup command deletes the candidate data after it shows you the list. The page adds that these are the system's own rules and do not replace legal advice.

How do I set one up?

You need Claude Code. Agent Company is set up from a chat conversation, and its template comes with a fictional sample company so you can watch it run before you adapt it to your work. The full setup guide for each system opens on its page after you leave an email address.

Next step

If you want one of these six jobs running in your own business, there are two ways to get there.

If you want to bring AI into your work and daily life and learn to build agents yourself, One-on-One Training fits. We meet for two live lessons a week for six weeks. You bring a task from your own work to each lesson and build the system on your own screen while I walk you through every step. By the end, you know how to build the next agent on your own.

If you want your company's repeat work handed over to AI together with your team, Advisory fits. Over three months we meet four times a month. In the first month we list the work that repeats every week in your company, rank it, and automate the top of the list. I also give your team two lessons on using these systems day to day.