Short answer: an AI agent is software that takes a goal, works out the steps, uses tools to carry them out, checks the result, and keeps going until the job is done or it needs a person. A chatbot answers you. An agent does the work, and finishes it in your systems.
I have built AI products from the inside: an AI engagement platform as CEO of Product Genius, a generative AI product used by Fortune 500 engineering teams as COO of Blameless, and New Relic AI as EVP and GM of the New Relic platform. I also run two inference servers of my own and work hands on with the models and how they are trained. So this is the version I give to CEOs: plain, and grounded in what the tools can do this quarter.
One example to keep in mind as you read: a client’s agent runs on a $4,000 server in their own office and takes 35 percent of the workload off four employees. If you already have a job in mind, custom AI agents explains the free call where I tell you whether an agent can take it.
What is an AI agent?
Think of a capable new hire who does not tire of the routine. You give them a job, such as “match each incoming invoice to its purchase order and flag what does not match”. They read what arrives, decide what to do with it, use the systems they have access to, check their own work, and come to you with the exceptions. An AI agent does the same, in software.
There is one difference from a new hire that matters more than the rest. A new hire who does not know how to do something says so. An agent tends to say “no problem” and produce something that looks right. That is why the way an agent is set up, and checked, matters as much as how clever it is.
The way I explain it: the model is the core knowledge, and the agent is the instructions and code in front of it that point that knowledge at one job. Whatever the vendor calls it, an agent is built from six parts:
- A model, the brains: it reads, reasons and decides the next step.
- Instructions and limits: what the job is, what done looks like, what it may do alone, what needs a person, and what it may not touch.
- Tools and connections, the hands: email, documents, websites and your business systems, so it can act rather than describe.
- A trigger: what starts the work. A new email, a form, a call, a change in the CRM, or a time of day.
- Knowledge: your documents (policies, price lists, past examples) and read access to the live records it needs (the order, the account, the invoice).
- Checks and a record: it confirms each step worked, logs what it did, and hands the odd case to a person.
The test: does it finish the job?
Here is the simplest way to tell a real agent from a demo. One tool reads a customer’s email, writes a good summary, and waits for someone to update the CRM and tell accounting. The other reads the email, updates the CRM, starts the accounting step, and flags only the exception for a person. Both use the same model. The difference is not intelligence; it is reach.
An agent that cannot write its result back into the system where the work lives is an advisor that someone else has to operate. The job is done when the record is updated, and a good agent reads the record back to confirm it. Put another way, a chat assistant helps you think and an agent helps you operate. If a system only searches and summarizes, it is an assistant. It becomes an agent when it makes a decision that changes what happens next, and carries it out.
How an AI agent works: read, decide, act, check
Agents run in a loop rather than answering once:
- Read. It takes in the work that arrived: an email, a form, a call, a new row in a system.
- Decide. The model works out the next step toward the goal, using its instructions and what it knows.
- Act. It uses a tool: opens a website, updates a record, drafts a reply, places a call.
- Check. It looks at the result instead of assuming it worked, compares it to what done looks like, and either moves on, retries, or hands the case to a person.
I have seen the same loop at very different scales. At New Relic, New Relic AI watched two billion events a minute so customers heard about a failure before it happened; in a small office, the same read, decide, act, check loop matches invoices. The scale changes; the loop does not.
AI agent vs. chatbot vs. automation
A chatbot answers
A chat assistant answers a question in a conversation. You still copy the answer, paste it into your systems and do the next step yourself. If a person moves the result from the chat window into the system, the process is still manual, with a faster typist. And plenty of website chatbots are not agents: you type a detailed question and it asks whether you want option one or option two. That is a phone tree.
Automation follows a fixed path
Scripts, workflow tools and robotic process automation follow steps someone wrote in advance. They are fast, cheap and predictable when the input does not change, and they break when it does: a new invoice layout, a customer who writes the order in a sentence instead of a form.
An agent works toward a goal
An agent decides its own steps at run time, handles variation in what arrives, and routes the cases it cannot handle to a person. The best systems mix the two. The fixed, rule-bound steps are built as ordinary automation, which runs the same way each time and costs almost nothing; the model is used only for the steps that need reading and judgment. A good builder will often use the AI to write that fixed automation once, rather than asking the AI to repeat the step each time.
An assistant or copilot helps; an agent drives
Assistants and copilots help one person inside one application, with the person driving: they suggest, draft and summarize, and the person decides and clicks. An agent is given the job and the limits, and drives itself; the person reviews the exceptions and the results. When a system moves from recommending an action to taking it, the stakes change, and so do the controls it needs.
AI agents vs. agentic AI vs. generative AI
The three terms get used as if they meant the same thing. They do not:
- Generative AI creates material on request: a draft, a summary, an image, a block of code. You ask; it produces.
- Agentic AI is the capability to plan, use tools and take actions toward a goal with limited supervision. Generative AI is often one of the tools it uses along the way.
- An AI agent is one system built on that capability to do one defined job, such as invoice matching or call answering, inside the limits you set.
For a business, the useful question is not which label a vendor uses. It is what the system is allowed to do on its own, in which of your systems, and who checks it.
What an AI agent can control
These days an agent can run in the cloud or on a computer in your own office, and work almost any application or website the way a person would. It can receive and make calls, read and send texts and email, and speak and listen in many languages. Within the permissions you grant, that covers most of what a person does at a keyboard or on the phone:
- Any application or website: log in, search, click, type and fill in forms, and move data between systems, including older desktop software with no other way in.
- Calls, texts and email, both ways: answer and place calls, receive and send texts and email, sort, draft and follow up on a schedule.
- Documents: invoices, contracts, forms, PDFs and scans, turned into data.
- Audio: listen to calls and meetings, transcribe and summarize them.
- Voice and languages: hold a natural conversation, and speak, listen, read and write in many languages, including the ones your customers use.
- Data: spreadsheets, the CRM, accounting, order and support systems.
Permissions are where the care goes, and they sort into three levels. Reading and looking things up is low risk. Changes that can be undone, such as a draft or a label, can run on their own as long as each one is logged. Anything that cannot be undone or that the outside world sees, such as sending, paying, deleting or changing who has access, gets a person’s approval, and the approval screen shows the actual email or amount, not a summary of it.
How an agent knows your business
Out of the box, a model knows the general internet. Ask it your leave policy and it gives you an industry average. Point the agent at your own documents and it answers from those instead: your policy, your price list, your way of handling a refund, the same answer for each person who asks. Engineers call this retrieval. In plain terms, the agent reads your folder before it answers.
Most jobs need two kinds of knowledge: documents, and live records. A support agent needs the returns policy and the customer’s actual order. Many wrong or empty answers come from wiring up only one of the two. And keep the business facts in one maintained document the agent reads, with an owner, so a price change is made once rather than hunted down across a dozen prompts.
Memory is a third kind, and it works differently. Within one task, the agent keeps track of what it has read and done so far, the way you keep notes while you work. Across tasks, some products also remember preferences and history: how you like a reply worded, what happened with this customer last time. That helps, and it needs care. Find out what the product remembers, where it is kept and whether the provider trains on it. And keep what matters most, your policies and how the job is done, in documents you own rather than in one tool’s memory, so you can change tools without losing it.
Two build habits make the answers trustworthy. Tell the agent to say what it could not find instead of filling the gap with a guess. And let ordinary software do the arithmetic: totals, counts and dates come from the data, not from the model adding numbers in its head.
The brains: frontier models and open models
Each agent has a model doing the reasoning. Frontier models, from companies such as Anthropic, Google, OpenAI and xAI, are the strongest reasoners. You reach them over the internet and pay for each piece of work.
Open models are published openly, free to download and run. On a server in your own office there is no per-use fee and your data stays in the building; an agent built this way is often called a local AI agent. In the cloud, open models cost a fraction of frontier prices. The choice is renting versus owning. Renting is simple and someone else maintains it. Owning gives you control and a fixed cost, and you look after it. Owning tends to pay when the volume is steady and high, or when the data cannot leave the building. It also gives you a fallback when a provider’s service is down.
When we built New Relic AI, I had PhD scientists in Israel building our own models from scratch. Today a business can take an open model, give it its own documents, and get much of that value for a fraction of the effort. For a great deal of business work, a smaller model is smart enough; you do not need a race car to go to the grocery store. The practical rule: try the job on a strong model to learn what good looks like, then run it on the cheapest model that still passes your tests, and keep the frontier model for the cases that earn its cost. The client agent I mentioned at the top works this way: open models on their own server, a one-time $5,000 build, and over $150,000 a year saved.
One design trick keeps the cost down further. Many of the judgment calls inside a workflow are really multiple choice: which team should handle this, how urgent is it, does it need a person? Small, fast models answer questions like that cheaply and consistently, which leaves the expensive model for the open-ended reasoning.
How much an agent should do on its own
Autonomy is not a switch. It is a ladder, and an agent earns each rung on evidence:
- It suggests. The agent proposes; a person does the work.
- It drafts for approval. The agent prepares the reply, the record or the payment run; a person approves each one.
- It acts and escalates. The agent handles the routine cases on its own and sends the exceptions to a person.
- It acts and reports. The agent runs the job and a person reviews a daily summary instead of each item.
Decide in advance what evidence moves it up a rung, such as a long run of items with no corrections, and what moves it back down. One wrong payment should put it back on one-by-one approval the same day.
A good first job for the third rung is sorting and routing. The agent decides where each item goes and a person makes the fix, because a wrong route costs a minute to correct and a wrong fix can cost a customer.
Types of AI agents you will meet
- Personal agents work for one person: the inbox, the calendar, research, drafting, a morning briefing of what needs attention.
- Workflow agents run a back-office process end to end: intake, matching, follow-up, reporting.
- Voice agents answer and place calls, book appointments and log the conversation.
- Background agents watch a stream of work and flag what needs a person: a mismatch between two systems, a refund above someone’s limit, a customer who sounds ready to leave.
- Research agents gather, compare and summarize information on a schedule.
- Coding agents write, test and review software.
Textbooks also sort agents by how they decide (rule-following, goal-based, utility-based, learning). For a business, the useful question is not the textbook type but the job, the systems it touches and who owns the exceptions.
One agent or several?
A single agent asked to do too much gets confused, the same way an overloaded employee does. Reliable systems use narrow agents, each with one job and one set of permissions, passing work between them. Start with one. When you do chain them, put a check at the point where a wrong value would travel furthest, because small errors compound as work passes from one agent to the next.
AI agent examples: what agents do in a business
Across the material I reviewed for this guide, the business functions where agents come up most are compliance checks, scheduling, HR and onboarding, customer support, document intake, lead follow-up and reporting. In plain terms:
- Matching invoices to purchase orders and flagging the mismatches.
- Gathering last month’s receipts, renaming each by date and vendor, setting the travel ones aside, totaling them and sending the total to the office manager.
- Chasing overdue payments, missing paperwork and pending approvals by email and text.
- Moving orders and customer details from email into the CRM, accounting or order system.
- Sorting a shared inbox and drafting replies to the common questions.
- Answering the phone after hours, booking appointments and confirming them.
- Answering new employees’ questions from the handbook, and handling routine HR and IT requests.
- Checking that a file is complete before it reaches the specialist who reviews it.
- Scoring each customer overnight as green, yellow or red for payment risk, so collections start the day with the right calls.
- Transcribing inbound calls to show where staff could not answer and where a ready buyer waited on an unwritten pricing rule.
A few I have seen up close. My first AI work, at Cox Automotive, used the trading history between car dealers to connect dealers who already trusted each other, so more of them traded electronically. At Blameless, generative AI turned the running notes of an outage into separate updates for executives, sales and support, so the engineers could keep fixing instead of rewording. At Product Genius, a digital employee is set up by scanning a business’s website and documents and giving it a name and a personality; at continuing-education events, attendees scan a QR code and ask, in plain English, which open sessions earn the credits they need. In restaurants, capturing customers’ questions in real time can show an owner a staffing gap, such as Wednesdays from 3:00 to 4:00.
What agents are good at, and what they are not
Good at: work that repeats each week, where language sits at the center (emails, documents, forms, calls), the rules can be written down, and a person can look at the result before it has consequences.
Not good at: situations they have not seen before and have no examples for; decisions that could end up in front of a judge, such as hiring, firing or credit, which a person should make and be able to explain; and original thinking. They are good at the derivative, not the original. And they do not reliably tell you when they do not know.
Four kinds of call stay with a person however good the agent gets: the high-stakes ones, the ones that cannot be undone, the ethically sensitive ones and the ambiguous ones. One more habit to know: models tend to agree with whoever is asking and hand back a better-worded version of what you already think. Ask it to push back, and do not take its agreement as proof.
Why the same question can get two answers
Ask a model the same thing twice and you can get two different answers. It chooses each word from likely options, so a small change in the wording, in what it read or in the model version changes the result. In a chat that is harmless. In a business process it means three things. Write the method down once, with examples, so the agent follows the same steps each time instead of working them out afresh. Put the fixed steps, such as a calculation or a format check, in ordinary code that gives the same answer each time. And test more than once, because one good run is a demo, not evidence. Several agent products now package this as a skill: the instructions, examples, scripts and test cases for one job, loaded when that job comes up.
Agents can be wrong, and good ones are built for it
Some vendors claim their agents do not make things up. Do not rely on it. A model can produce a confident, well-written answer that is wrong, and the good writing makes the mistake harder to spot. When work took effort, the effort was a rough quality filter; a model produces polished work instantly whether it is right or not. The answer is design, not trust: the build habits above, plus a check at each step, a fallback to the fixed rule or a person when it is unsure, a separate check whose only job is to find mistakes, and a person on what matters.
Where AI agents go wrong
- The island. It summarizes or recommends but cannot finish the job in your systems, so someone still re-types the result.
- Too much access. An agent does what it is allowed to do. One that runs on a person’s own login inherits whatever that person can reach, which is usually more than people realize. Give it its own account with only what the job needs.
- Instructions hidden in what it reads. An email, a web page or a document can contain text telling the agent to do something. What the agent reads is data, not instructions, and the limits have to be enforced in the software, not just written in the prompt.
- “Done” that is not done. An agent reporting success is the agent declaring itself finished. Check the output, not the status.
- The running bill. An agent that loops, retries and checks uses far more model capacity than the demo did. Routing the routine work to a smaller or open model is what keeps the bill sane, and a retry loop with no limit can burn a month’s budget overnight.
- A broken or unclear process. An agent follows instructions exactly, so it shows you fast where the process ran on someone quietly filling the gaps. If the people doing the work cannot agree how it should go, the agent makes the confusion faster and better-looking. Write down what done looks like first.
- No owner. A demo runs on clean examples with someone watching. Real work brings the odd case, and without a named person to catch it, trust collapses on the first mistake.
- The model changes under it. Providers update models on their own schedule, and the work arriving changes too, so an agent that worked well in spring can slip by autumn with no warning.
- The review pile. Agents move work as well as remove it. What is left is direction and review, and a confident mistake passes a quick glance.
The fixes for these, from caps on attempts and spending to named owners and a test file you rerun when the model changes, are in the 30-day plan and the safety checklist in AI agents for business.
Try one yourself this week
You do not need a developer to learn how agents behave. With a current AI assistant and an afternoon:
- Pick one task you repeat each week and dread a little, such as a weekly summary, a follow-up email or sorting a folder of documents.
- Write a one-sentence definition of done a stranger could check, then tell the assistant why you want it, not only what.
- Ask it to interview you: “Ask me the questions you need to do this well.” Answer out loud if that is faster.
- Give it past examples, such as last month’s emails or reports, so it learns how you actually do the work.
- Test it on this week’s real items, correct what it gets wrong, and feed the corrections back. Where it goes wrong, look at your instructions first; it usually shows you what you left unsaid.
- Keep it to drafts. Let it prepare; you send. Do not give it the ability to send, pay or delete until it has earned it.
If it saves you time on a drafted task, you have found a candidate for a real agent: the same job, wired into your systems, with tests and limits.
How to start in your business
My operating rule is simple: “Any process you can model can be mostly automated, with the exceptions routed to humans.” Start with one workload in the back office, the work your people re-type, chase or check day after day. Ask where its data lives and who owns it, whether it can be written down as steps with exceptions, what it costs today, what happens to the exceptions, and who will own it once it runs.
AI agents for business turns that into a 30-day plan: how to pick the first job, what it costs, how to test it and how to keep it safe. If you have an idea and want it built in weeks, custom AI agents explains how I work, including the free call where I tell you whether an agent can take the work.
What people ask about AI agents
What are AI agents, in simple terms?
Software that takes a goal, decides the steps, uses tools such as websites, email and business systems to carry them out, checks the result, and hands the cases it cannot handle to a person.
How is an AI agent different from ChatGPT?
A chat assistant answers in a conversation and you act on the answer. An agent acts: it uses your systems to do the work and checks its own result. Several chat assistants now include agent modes that can browse and take actions, which is the same idea built into the chat product.
What is the difference between agentic AI and an AI agent?
Agentic AI is the capability to plan, use tools and act toward a goal. An AI agent is one system built on that capability to do one defined job inside limits you set.
Is Alexa an AI agent?
Voice assistants such as Alexa and Siri carry out short commands, which makes them a narrow kind of agent. A business AI agent works toward a goal across several steps and systems, decides the steps itself, and routes exceptions to a person.
What are examples of AI agents?
Invoice matching, payment follow-up, inbox sorting and replies, after-hours phone answering and booking, data entry between systems, weekly reporting, and answering HR and IT questions from your own documents.
Can an open-source AI agent run on my own server?
Yes. Open models can run on a server in your own office with no per-use fee, so data stays in the building, or in the cloud at a fraction of frontier prices. A common design uses an open model for routine volume and a frontier model for the hard cases.
Are AI agents safe for business data?
They can be, when each agent has its own account with only the access its job needs, a person approves anything that sends, pays or deletes, each action is logged, and sensitive data stays on your own hardware where that matters.
How much does an AI agent cost?
From a monthly subscription for an off-the-shelf tool to thousands or tens of thousands for a custom agent built for your systems, plus the running cost of the model. AI agents for business breaks it down.