How much does an AI agent cost in 2026?

An AI agent can cost nothing, a few hundred dollars a month, or six figures, because "AI agent cost" describes three different things. Buying an off-the-shelf agent usually costs from a free tier up to a few hundred dollars a month, or roughly $1 to $2 per conversation. Building a custom agent typically costs $8,000 to $40,000. Running either one adds model usage, hosting and a little maintenance every month.

These are approximate ranges drawn from publicly listed prices and the kinds of quotes that circulate in this market. They are not a formal survey, and vendor prices change often. Use them to sanity-check a plan or a proposal.

#The three costs people mean

Most confusion comes from comparing a number from one row with a number from another.

#What it costs to buy an agent

Many tools you may already pay for now include an agent: help desks, CRMs, e-commerce platforms and workflow tools. There are also general agent builders that let you assemble one without code. They charge in one of four ways.

• Per seat. A monthly fee per user, often as an add-on to a plan you already have. Simple to budget. It can get expensive when many people need access but few use it heavily.
• Per conversation or per resolution. You pay each time the agent handles or fully resolves a request. Published list prices for well-known support agents have sat at roughly $1 to $2. It is fair at low volume and needs watching as volume grows.
• Usage credits. You buy a pool of credits and each action spends some. Flexible, and the hardest model to predict. Ask for an estimate based on your real volumes.
• Platform fee plus usage. A base subscription with metered usage on top. Common with agent builders.

Three things to check before you commit:

1. How a "resolution" is defined. If the customer simply stops replying, does that count?
2. What happens at three times your current volume. Per-conversation pricing that looks small today can pass the cost of a custom build within a year.
3. What you can take with you. Your knowledge content is yours. The agent's configuration, conversation history and tuning often are not portable.

#What it costs to build a custom agent

A custom agent is a software project. The price follows the amount of engineering, not the cleverness of the model.

The bottom of each range is usually a small specialist firm with a narrow scope. The top is usually more systems, more risk, or a larger firm with more overhead.

#What moves the build price

• How many systems the agent touches. Each one needs a connection, permissions and error handling. A system without an API can double the work.
• The state of your knowledge. Clean, current documents are quick to index. Contradictory policies and scanned PDFs are not.
• What the agent is allowed to do. Answering is cheap to make safe. Issuing refunds, changing records or sending external email needs approvals, limits and a way to undo.
• How accuracy is proven. A proper evaluation set, built from real examples and run before every change, takes time. It is also the difference between an agent you hope works and one you know works.
• Channels. A web chat is one piece of work. Adding Slack, Teams, email or a place inside your product adds more.
• Users and permissions. An agent that must respect who is allowed to see what, or that serves many customer accounts, needs careful design.
• Compliance. Regulated data, audit trails and data-residency rules all add scope.

#What it costs to run an agent

A build quote is not the whole cost. Every agent has a monthly bill made of three parts.

#AI model usage

You pay the model provider for the text the agent reads and writes, measured in tokens. Prices range from well under a dollar to tens of dollars per million tokens depending on the model.

As an illustration: suppose a conversation uses 20,000 tokens once the instructions, the retrieved passages and several turns are counted, and the blended price is $2 per million tokens. That is 4 cents per conversation. At 5,000 conversations a month the usage bill is about $200.

Long documents, top-tier models and agents that take many steps cost more. A good build uses a smaller model for routine steps and a stronger one only where it matters, and sets a spending limit with alerts. Ask any supplier for an estimate based on your own volumes before you start.

#Hosting and monitoring

A small server or serverless functions, a database for the knowledge base, and logging. For most small and mid-sized deployments this is a minor monthly line.

#Maintenance

Your policies change, your products change, and new edge cases appear. Someone has to review conversations, update the knowledge and re-run the evaluation. Budget a few hours a month of your own team's time, or a support plan. Market rates for ongoing support typically run from $1,500 to $8,000 a month depending on how much is covered.

We go deeper on this in what it actually costs to run an AI automation in production.

#Build or buy: a simple rule

Buy when a product already does the job, your volume is modest, and the agent does not need to act inside systems the product cannot reach. A support agent on top of your existing help desk is the usual example.

Build when the work is particular to your business, when the agent must read and write to your own systems, when you need to control exactly how it behaves, or when per-conversation fees at your volume would pass the build cost within a year or two.

Many businesses do both: an off-the-shelf agent for the common case and a custom one for the process that makes them different. For the difference between the two kinds of tool, see AI agent vs chatbot.

#A quick payback check

Before comparing quotes, check whether the agent is worth having.

Cost to handle one request by hand = minutes per request ÷ 60 × loaded hourly cost.

Monthly saving = requests the agent fully handles × that cost.

As an illustration: a request that takes 8 minutes of someone's time at a loaded cost of $35 an hour costs about $4.70 to handle. An agent that fully resolves 40 percent of 3,000 monthly requests takes 1,200 of them, worth roughly $5,600 a month. After a few hundred dollars of running costs, a $20,000 build pays back in about four months.

Run the same sum at 300 requests a month and the saving is about $560. At that volume an off-the-shelf tool, or no agent at all, is the better answer. You can try your own numbers in the AI automation ROI calculator.

#How to read a quote for an AI agent

A good proposal makes these easy to find.

1. The job. One or two sentences on what the agent is responsible for, and what it hands to a person.
2. One price, or a cap.
3. How accuracy is measured. An evaluation set of real examples, with a target, run before launch and before every change.
4. Guardrails. What the agent will refuse, which actions need approval, and what is logged.
5. Running costs. Model usage and hosting estimated from your volumes, with a spending limit.
6. Ownership. The code, prompts, knowledge base and accounts are yours.
7. Handover. Documentation and a walkthrough, so you are not tied to the builder.

If a low quote is missing most of these, it is not comparable with the higher ones.

#What we charge

So that you have one concrete reference point, here is our own pricing. Every project is quoted as one fixed price before work begins.

• Custom AI agent: from $9,800, three to five weeks. That covers one well-defined job, a knowledge base, a small set of tools, guardrails, an evaluation set, logging and a 30-day fix window.
• AI Opportunity Audit: $1,500, one week, if you are still working out where an agent would help most.
• Ongoing support: from $2,500 a month, optional.

Model usage and hosting are billed to you by the providers, and we estimate both up front. Full details are on the pricing page.

#Bottom line

If a product already does the job, start there: it will cost little and teach you a lot. If the job is specific to your business, expect $8,000 to $40,000 for a custom agent, plus running costs that are usually small next to the time saved. Whichever route you take, work out what one request costs you today, and only then decide what an agent is worth.

Common questions.

• How much does it cost to build an AI agent?

  Most custom AI agents for small and mid-sized businesses cost between $8,000 and $40,000 to build. A simple agent that answers from your documents sits at the low end. An agent that reads and writes to several systems, handles sensitive actions and needs formal testing sits at the high end. Enterprise programs with many agents, single sign-on and compliance requirements often pass $100,000.

• How much does an AI agent cost per month to run?

  For a custom agent, expect three lines: AI model usage, which is often a few cents per conversation; hosting and monitoring, usually a small fixed amount; and maintenance, either a few hours of your own team's time or a support plan. At a few thousand conversations a month, many businesses spend low hundreds of dollars on usage and hosting combined.

• Is there a free AI agent?

  Yes, several agent builders and help desk tools have free tiers, and they are a sensible way to learn what an agent can do. Free tiers usually limit conversations, knowledge sources or integrations, and they rarely include the controls a business needs, such as approvals, logging and team permissions.

• Is it cheaper to build an AI agent or buy one?

  Buying is cheaper when a product already does the job and your volume is modest. Building becomes cheaper when per-conversation fees add up at volume, when the agent must act inside systems the product does not connect to, or when the work is specific to how your business runs. Compare three-year totals, not the first month.

• Why do quotes for AI agents vary so much?

  Because the name covers very different amounts of work. One quote may be a chat window connected to a model. Another includes a knowledge base, permissioned tools, guardrails, an evaluation set, logging, a cost dashboard and handover. Ask each supplier what is running at the end and how accuracy is measured, then compare like with like.