What an AI Agent Actually Is, and What One Costs to Run

Every vendor in America now sells AI agents. Most of them are selling a chatbot with a longer sales cycle. Below is the dividing line, four agents that actually pay for themselves at small-business scale, and what the monthly bill looks like when you do the token math instead of guessing.

The dividing line, in one test

Ask one question about any system a vendor calls an AI agent:

If the goal succeeded a different way than expected, would the software notice and change course?

If yes, it is an agent. If it would run the same seven steps regardless, it is a workflow, and there is nothing wrong with that. Workflows are more predictable, cheaper, and easier to audit. Most businesses need three workflows and one agent, and are being sold four agents.

Here is the same distinction across the four things people currently call AI agents.

We wrote the longer comparison in AI chatbots vs AI agents. The terminology fight over the word "agentic" specifically is in what is agentic AI, with the head-to-head in agentic AI vs AI agents and the separate generative question in agentic AI vs generative AI. Read one of them if a vendor is leaning hard on the vocabulary. None of it changes the test above.

Who Google quotes for this today

Pulled live from Google US, logged out, on 25 Aug 2026. The interesting finding is not the ranking list. It is that the definitional query and the commercial query are two completely different competitive situations, and almost nobody writing about AI agents seems to have noticed.

Search "what is an ai agent" and an AI Overview fires. It cites six sources, in this order: YouTube, IBM, Google Cloud, AWS, a Substack newsletter called The AI Engineer, and Snowflake. Organic underneath is a similar cast:

• Reddit, a thread titled "What even is an AI agent?"
• IBM.
• McKinsey.
• Google Cloud.
• LangChain.
• AWS.
• The AI Engineer, on Substack.
• Salesforce.

Look at who that is: a model and infrastructure vendor, three cloud platforms, a data warehouse, a consultancy, an orchestration framework, and a CRM. Every one of them sells something an agent runs on, is built with, or is bought through. The definition on page one is therefore the definition that makes the seller look necessary. That is not a conspiracy, it is just what happens when a dictionary term is also a product category. Reddit sits at number one because it is the only result written by somebody with nothing to sell.

Now search "how much does an ai agent cost" the same day, in the same logged-out browser. Completely different world. The AI Overview cites seventeen sources and the first is Reddit, followed by bakedwith.com, softteco.com, thecrunch.io, braincuber.com, a Medium engineering publication, appinventiv, savibm, powercodegroup, aissist.io and several more in the same vein. The closest thing to a recognized name anywhere in that list is CloudZero, a cloud cost-monitoring company. Organic: Reddit again at number one ("Been running my businesses on AI agents for months"), then softteco, Retool with a piece arguing for hourly pricing, bakedwith, an AWS builder guide, YouTube, Medium, and azilen.

IBM is gone. McKinsey is gone. Google Cloud, AWS as a marketing site, Salesforce, all gone. The moment the question turns into "what will this cost me," the platforms stop answering and the field is left to one Reddit thread and a long tail of development shops with no consistent authority between them. Seventeen citations in a single AI Overview is Google saying, in public, that it could not find a source it trusted enough to lean on.

factoryjet.com appears in none of it, on either query. We are writing this from outside the results, not from inside them. That gap is the entire reason the article exists: the definition is over-served by companies selling infrastructure, and the price is under-served by everybody. So the second half of this page does the arithmetic.

One more detail from the same pull, because it tells you who is actually searching. The People Also Ask box under "what is an ai agent" carries "Do AI agents get paid?", "How to make $1000 a day using AI?" and "How much can I make selling AI agents?" Under the cost query it carries "Can you really make money with AI agents?" A meaningful share of the traffic on this term is not businesses trying to buy an agent. It is people trying to sell one. Worth remembering the next time a cold email arrives from an agency that discovered the category last month.

The three parts every real agent has

1. A goal and a stopping rule

"Resolve this ticket or escalate it to a human within four exchanges." Not "be helpful."

The stopping rule is the part that gets skipped and the part that causes the incidents. An agent without a budget, a retry limit and an escalation path will loop, spend, and confidently do the wrong thing at 3am. Every agent we ship has a hard ceiling on steps, a hard ceiling on spend, and a named human it hands to.

2. Tools that change something

A tool is a function the agent can call: look_up_order(order_id), issue_refund(order_id, amount), create_quote(line_items), book_slot(date, time).

This is where the real engineering lives, and it is why agent projects cost what they cost. The model is a commodity you rent by the token. The issue_refund function that talks to your payment processor, respects your policy, writes to your ledger, is idempotent so a retry does not refund twice, and logs enough to survive a chargeback dispute: that is the product.

3. A check on its own work

The agent proposes an action, something verifies it, and it retries or escalates on failure. In practice that means schema validation on every tool call, a permission gate on anything that moves money, and a confidence threshold below which it hands to a person.

An agent without step 3 is a demo. Demos are why the category has a trust problem.

Four agents doing real work in US businesses

Not hypotheticals. These are the four shapes that actually pay for themselves at small and mid-size scale. The wider survey of what is landing in US small businesses right now is in AI agents for small business.

Support deflection. Reads the ticket, retrieves the order and the policy, answers or acts (refund, reship, address change), replies in your voice, tags and closes. Escalates anything above a dollar threshold or below a confidence threshold. Lives inside Zendesk, Gorgias, Freshdesk or Help Scout. The architecture detail is in AI customer support agent architecture and the service is AI customer support agents.

Quote and RFQ handling. Reads an inbound RFQ, usually a PDF or an email with a spreadsheet attached, extracts line items, matches them to your catalog and pricing rules, flags the ones it cannot match, and drafts the quote for a human to send. This is the highest-return agent in manufacturing and distribution and almost nobody has built it, because it requires knowing the customer's part numbering. See RFQ and bidding agents.

Lead follow-up. Watches the CRM for a new inbound, researches the company, drafts a first-touch email that references something real, sends, waits, follows up on a schedule, and books the meeting straight into the calendar. Stops the moment a human replies. Detail in how to build a custom AI SDR; the service is AI sales agents.

Back-office reconciliation. Matches purchase orders to invoices to receipts, flags the mismatches, and files the clean ones. Boring, unglamorous, and the one with the clearest payback, because you can count the hours it replaces. The plumbing under it is ordinary workflow automation with a judgment step in the middle.

A fifth shape, voice, is the same architecture with speech in and speech out, and stricter latency requirements that change both the model choices and the cost. That one has its own page: AI voice agents.

Two patterns run through all of them. Each one owns a queue with a countable backlog, and each one hands off rather than guessing when it is unsure.

What one actually costs

Three separate numbers, and vendors love to quote you whichever one flatters them.

1. The model bill, smaller than you expect

Worked example: a support agent handling 2,000 tickets a month. Across roughly two model calls per ticket it reads about 16,000 input tokens (system prompt, your policy docs, product data, the ticket thread) and writes about 1,200 output tokens.

The lever that matters is prompt caching. Your system prompt and policy documents are identical on every ticket, so they can be cached and re-read at roughly a tenth of the input price. Say 12,000 of those 16,000 input tokens are cached and 4,000 are fresh.

Run that against Anthropic's published list prices as of August 2026. Model prices move, so check the current published rates before you build a budget on them.

The Sonnet 5 row worked out in full, so you can check it: 12,000 cached input tokens at roughly $0.20 per million is $0.0024, plus 4,000 fresh input tokens at $2 per million is $0.008, plus 1,200 output tokens at $10 per million is $0.012. That is $0.0224 a ticket, and $44.80 across 2,000 tickets. Those totals are arithmetic on the assumptions above, not a measurement of your business. Change the token counts and the answer changes.

Read the table again anyway, because it reframes the whole decision. The model bill in this example is tens of dollars a month. Anyone quoting you a five-figure monthly fee "because of AI compute costs" is not describing this workload. Ask what the number is per unit of work, and watch what happens.

Two things do move it. Skipping prompt caching roughly triples the input cost, and agents that loop pay for a full round trip on every retry. This is the second reason for a hard step limit. The first reason is safety. The second is your invoice.

2. The build, where the money actually goes

The bands we see quoted in the US market for a single production agent, integrated into live systems, with permissions, logging and a human handoff:

• A single-queue agent on systems with clean APIs (Shopify, Zendesk, HubSpot): low to mid five figures.
• Anything touching an ERP (NetSuite, SAP, an on-prem system, or a database somebody's cousin designed in 2009): meaningfully more, and the estimate is mostly a guess until someone has looked at the data.
• A pilot on one narrow queue, deliberately scoped to prove or kill the idea: four figures, two to four weeks. This is what we recommend first, every time.

The variable is never the model. It is how many systems have to be touched and how bad their data is.

3. Running it

Model bill, plus monitoring, plus the human who reviews escalations, plus the maintenance when a vendor changes an API or your policy changes and the agent needs to know. Budget for the reviewer. An agent with nobody watching the escalation queue is an agent that quietly stops working and nobody notices for six weeks.

The full comparison of doing it yourself against hiring it out, including what a competent internal developer can realistically ship, is in build AI agents, DIY vs agency.

When not to build one

Straight answers, because this is the part the category never says out loud.

Your queue is under about 200 items a month. The build will not pay back. Hire a part-timer or fix the process.

The decision rule is genuinely fixed. If the answer is always "if X then Y," that is an if statement. It is cheaper, it is faster, it never hallucinates, and it does not need a reviewer. n8n vs Zapier vs Make covers that tier honestly.

Your data is a mess. An agent reading a catalog where the same part has four names will produce four answers. Fix the catalog. That project is less exciting and worth more.

You cannot tolerate a wrong answer and cannot afford review. Regulated advice, clinical decisions, anything where the error is unrecoverable. Use retrieval to help a human decide, rather than an agent to decide.

What you actually want is a chatbot. If the job is answering questions on your website and nothing more, buy a chatbot. It costs a fraction and it will not disappoint you, because you have not asked it to do anything it cannot do.

How to tell whether an agent proposal is serious

Six questions. The answers are diagnostic.

• Which queue, and how many items are in it this month? No number means no business case.
• Which systems does it write to, and through which API? "We integrate with everything" means they have not looked at yours.
• What is the step limit and the spend cap per run? If there is not one, they have not run an agent in production.
• What happens when it is unsure? The answer has to be a named person and a route, not "it asks for clarification."
• What does one unit of work cost in model calls? They should be able to do the arithmetic above for your workload in about five minutes.
• Who reviews the escalation queue, and how often? If the answer is "you will," that belongs in the plan and in the budget.

An honest agency will fail some of these on your project and tell you which ones. If you want it scoped against your actual systems, AI agent development is the page to read next. The first conversation should be about the queue, not the model.