AI agents for business: what they are, what they cost, and which ones are worth building (2026)
An AI agent for business holds one job and carries it from start to finish. The three kinds on offer, a cost table from a $20 seat to a custom build, the approval rules that keep agents safe, and how to pick the first job to hand one, from a team that runs five in production.
In September, an AI agent went to the top of the app charts. Three weeks later its maker pointed it at small businesses and plugged it into QuickBooks, Slack and Shopify. Searches for "AI agents for business" went from a few hundred a month at the start of the year to eighteen thousand in August.
Everybody now agrees a business should have agents. Almost nobody agrees on what one is.
An AI agent for business is software that holds one job and carries it from start to finish: it reads what comes in, decides what to do under rules the business wrote, takes the action in the business's own systems, and hands off to a person at the points the business chose. That is the difference from a chatbot, which answers and stops. Agents come in three kinds: personal agents that run errands, agents sold inside products you already use, and agents built for one job in your company. Costs run from $20 a month for a seat to $5,000 to $25,000 for a custom build, plus the model usage underneath, which is often smaller than people fear. We run five in production for about $200 a month.
That's the short answer. The rest of this page is how to tell the three kinds apart, what each really costs, and how to pick the first job.
An agent is a job description, not a product
Here is the test we use before building anything. Write the agent's job the way you would write a posting for a new hire.
What it handles. What it must never do. Who it reports to. How you will know it is doing well.
If you can fill in those four lines, you have an agent worth building or buying. If you can't, you have a demo. Most disappointing agent projects skipped this step and started from the tool.
It also settles the chatbot question. A chatbot has no job. It has a text box.
The three kinds, and when each is the right one
| Kind | What it is | Good at | Stops being enough when |
|---|---|---|---|
| Personal agent | A general assistant working for one person. Meta's Muse is the loud example; Claude and ChatGPT do the same from their own apps. | Errands, research, drafting, anything one person would otherwise do by hand | The work depends on rules, records or approvals that live in your company, not in one person's head |
| Product agent | An agent that ships inside software you already pay for: your CRM, your help desk, your accounting system | The job that product was built around, with no setup beyond switching it on | The job crosses two systems, or your process differs from the product's idea of it |
| Built agent | An agent made for one job in your business, wired into your own systems, working to your written rules | The work only your company does your way: your intake, your follow-up, your reporting, your website | Rarely. The limit is usually messy data underneath, which has to be fixed first |
Start at the top of that table and move down only when you hit the right-hand column. Buying before building is not a consolation prize. It is how you find out which job is actually worth building for.
What agents are doing in real businesses today
Skip the predictions. These are jobs agents hold right now, with numbers from builds we have shipped.
Intake. An estate planning firm's new inquiries waited an average of 52 hours for a reply. An agent now answers at any hour, qualifies the prospect against the firm's criteria, collects documents and books the consultation. Response time: under four minutes. Consultations: up 2.4 times. Staff added: none. The build log has the whole story.
The morning briefing. An agent reads every call, deal and lead in a firm's CRM overnight and sends each of five team members their own action list before 8 a.m. Every open lead has a scheduled next step. Case study.
The inbox. An agent sorts every incoming email into four tiers and drafts replies in the owner's voice. The daily review takes five minutes. Case study.
The website. An agent checks a site every morning, fixes what it is allowed to fix, drafts new pages for approval and reports. For an M&A advisory firm, that put eight pages live in four weeks and cut the time Google takes to find a new article from more than eight days to about thirteen hours. We run our own site the same way.
The reports. An agent rebuilt one firm's marketing attribution from raw records and found $48,000 of revenue credited to the wrong channel. Case study.
Five jobs, and none of them is "be an assistant." Each one would fit on a job posting.
What AI agents cost
Three layers, and most price talk confuses them.
| Layer | What you are paying for | Typical cost |
|---|---|---|
| A seat | One person's access to a general agent | $20 a month for Claude Pro; $100 or $200 for Max; $25 to $125 per seat on a Team plan. Muse has a free tier with paid plans. |
| A platform | An agent builder or an agent feature inside business software | The market quotes $21 to $50 and up per user per month at the entry level, and $500 to $5,500 a month and up for heavier setups |
| A build | An agent made for one job in your business | Our published rates: $2,000 to $5,000 for one agent on clean systems; $5,000 to $25,000 when the data and connections under it need work too; $3,000 to $8,000 a month to run and extend several. Enterprise builds elsewhere are quoted at $40,000 to $100,000 and more. |
Then there is the running cost, which is where the fear usually sits. Every time an agent thinks, it uses a model, and models are billed by usage.
That bill is smaller than most people expect, for two reasons. Prices keep falling: Anthropic's smallest current model costs 10 cents per million tokens of input, a tenth of its predecessor's price. And a well-built agent sends routine work to the cheap model and saves the expensive one for the hard parts. Our model guide shows the routing.
Our own numbers: five production agent loops, including the one that runs this site's search program every day, cost about $200 a month in subscriptions plus a few dollars of data fees. A weekly research pipeline costs $15 to $30 a month to run.
The expensive part of an agent is almost never the agent. It is the cleanup underneath it.
Rules first, agent second
A surprising share of what gets sold as an AI agent is a routing rule wearing a costume.
If a lead came from a webinar, send the webinar follow-up. If an invoice is thirty days late, send the reminder. No judgment required, so no model required. Rules are cheaper, faster and never invent anything.
An agent earns its place where a rule runs out: reading a document, weighing a reply, deciding which of several things to do next. The best systems we build are mostly rules, with an agent at the two or three points that need a mind.
Sometimes the answer has no AI in it at all. A non-profit was losing two hours every other Friday to payroll reconciliation across three spreadsheets. The fix was one browser file. Five minutes now. No model in production.
Anyone selling you agents should be willing to tell you that.
Who approves what: Tier A and Tier B
The real question about an agent is not how smart it is. It is what it may do without asking.
We sort every action into two tiers.
Tier A: the agent does it, checks it and reports it. Routine, reversible work with a clear right answer. Fixing a broken link. Filing a record. Sending the scheduled follow-up. Correcting a date that has passed. Each action is logged and can be undone in one step.
Tier B: the agent prepares it, a person approves it. Anything new that speaks in the company's name, and anything that touches a promise, a price or a person. The agent does the research and the draft. A person says yes.
The test for which tier is simple. Would the owner want to be asked? Nobody wants a meeting about a broken link. Everybody wants a say in a new claim.
Most agent failures you read about are Tier B work running on Tier A permissions. Write the tiers down before you write any code.
What still goes wrong
Messy data. An agent on a CRM full of duplicates will act on the duplicates, quickly. Clean first. This is the most common reason projects stall, and it is why implementation has an order.
No handoff rule. An agent with nowhere to send the hard cases will handle them itself. Badly.
Confidential information in the wrong tool. Consumer agents on personal phones are not where client files belong. Read the terms of anything that touches regulated data. Our guide for law firms goes into this for one regulated industry.
Starting with the tool. See the first section.
How to pick your first one
Look for a job with four properties. It repeats every day or week. It crosses more than one system. It has a number attached, such as hours, response time or leads. And if it is done slightly wrong, a person will catch it before a customer does.
Intake and follow-up usually qualify. Reporting almost always does. Strategy never does.
Then go down the table. Can a seat do it? Can a product you already own do it? If not, that is the one to build.
If you want help finding the job, our revenue audit is free and takes about an hour. If you already have the job description, book a call and bring it.
The app at the top of the charts can book your dinner. The question for a business is which job, in your company, you would hand over on Monday.
Frequently asked questions
What is an AI agent for business?
An AI agent for business is software that holds one defined job and carries it from start to finish. It reads incoming information, decides what to do under rules the business set, takes action in the business's own systems, and hands off to a person at agreed points. A chatbot answers questions and stops; an agent completes the work.
How much do AI agents cost for a business?
A seat on a general agent costs about $20 to $200 a month per person. Agent platforms and product add-ons run from roughly $21 per user per month into the thousands. A custom agent built for one job costs $2,000 to $25,000 at Automaton's published rates, depending on how much data and integration work sits underneath. Model usage is billed separately and for a small business is often a few hundred dollars a month or less.
What is the difference between an AI agent and a chatbot?
A chatbot responds to a prompt and waits for the next one. An AI agent is given a goal, plans the steps, uses tools and systems to carry them out, and keeps going until the job is done or it reaches a point where a person must decide. The practical test: a chatbot has a text box, an agent has a job.
Should a business buy an AI agent or have one built?
Buy first. Use a general agent or the agent features in software you already own, and move to a custom build only when the job depends on your own rules and records, or crosses several systems. Building is worth it for work only your company does your way, such as intake, follow-up, reporting or running your website.
What business tasks are AI agents good at today?
Agents are reliably handling client intake and qualification, lead follow-up, daily pipeline briefings, email triage, recurring reports, and the technical upkeep and content drafting for a company website. They suit work that repeats, crosses systems and can be measured. They are not suited to strategy or to decisions a customer would expect a person to make.
Do AI agents replace employees?
In the builds we have shipped, they replaced tasks, not people. An intake agent answered inquiries no one had time to answer within two days; a briefing agent did the reading nobody was doing. The staff stayed and spent their time on the work that needs judgment. The handoff rules decide where that line sits.
How do you keep an AI agent from making costly mistakes?
Limit what it may do without approval. Sort actions into two tiers: routine, reversible work the agent does and reports, and anything new or sensitive that a person approves first. Log every action, make each one reversible, and give the agent a clear place to send cases it should not decide.