How Much Does It Cost To Build An AI Agent? A Practical Breakdown For Growing Businesses

AI agent cost has three parts: the one-off cost to build the agent, the running cost each time it completes a task, and the monthly cost to keep it working. In the worked example later in this guide, the build is the largest cost in year one, maintenance and people's review time come next, and model usage is the smallest line.

Many buyers expect model fees to be the biggest cost. Published price lists rarely help: many give a single range for AI agent development cost, often from a few thousand dollars to several hundred thousand, with no method behind it. A range that wide cannot help you budget. What helps is knowing which decisions move the number, how to calculate the running cost of one task yourself, and which costs arrive after launch.

This guide covers all of that. It starts with a question most cost guides skip, which is whether you need an agent at all. It then explains what drives the cost to build an AI agent, shows the running-cost math step by step using current published model prices, lists the costs that estimates tend to leave out, and works through a full year-one budget for an illustrative business. It is written for owners and managers of businesses with roughly 10 to 500 people who want to judge a quote, not write the code.

What Counts As An AI Agent, And Why It Changes The Price

An AI agent is software that uses an AI model to decide its own next steps and use tools, such as searching a database, reading a document or updating a record, until a task is done. Anthropic separates agents from workflows: in a workflow, the steps are fixed in code, while agents are "systems where LLMs dynamically direct their own processes and tool usage" (LLMs, or large language models, are the AI models behind tools like Claude and ChatGPT) (Anthropic, Building effective agents (opens in a new tab)).

The difference matters for AI agent cost because freedom costs money. An agent that chooses its own steps makes more model calls, needs more testing, and needs more controls to stop it doing the wrong thing. Anthropic's engineers put it plainly: agentic systems "often trade latency and cost for better task performance", meaning they are slower and cost more in exchange for better results, and the same guidance warns that autonomy brings higher costs and the risk of errors that build on each other.

The word "agent" is also used loosely in sales material. Gartner calls this "agent washing", meaning existing chatbots, assistants and automation tools rebranded as agents, and it estimated that only about 130 of the thousands of vendors claiming agentic AI offer real agents (Gartner, June 2025 (opens in a new tab)). Before comparing quotes, check that each one describes the same thing. Our guide to how to vet an AI vendor shows how to check the claims behind a product before you buy it.

First, Check Whether You Need An Agent At All

Before pricing an agent, check whether a simpler tool would do the job. Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls (Gartner, June 2025 (opens in a new tab)). Choosing the simplest option that works is good protection against all three.

There are four common options, from cheapest to most expensive to build and run:

A useful test: if you can write the process as a flowchart that covers nearly every case, a workflow will usually be cheaper and easier to test than an agent. An agent earns its higher cost when cases vary so much that the flowchart would never be finished. Our guide to automating repetitive business tasks explains how to pick between rules and AI for each step, and an AI readiness assessment checks whether your data and systems can support either.

What Drives AI Agent Development Cost

The cost to build an AI agent comes mostly from people's time: working out the workflow, connecting systems, writing the agent's instructions and tools, adding controls, and testing. The model itself is a small part of the build. Six factors move the build effort the most.

1. How Many Decisions The Agent Makes

An agent that answers one kind of question is far cheaper than one that handles ten kinds of request across departments. Each new type of task adds instructions, tools and test cases. Starting with one narrow job and expanding later is one of the most effective ways to keep the build cost down.

2. Integrations With Your Systems

An integration is the connection that lets the agent read from or write to another system, such as your accounting software, customer database or inventory system. Modern systems with documented APIs (ways for software to talk to each other) are quicker to connect. Older systems without them can take much longer and sometimes need a workaround. Integrations are often the largest single item in a build estimate.

3. Data And Document Preparation

If the agent needs to look things up in your documents, such as policies, product manuals or contracts, someone has to prepare those documents and set up a way to find the right passage. This is called retrieval, and it brings its own design and testing work, covered in our guide to building RAG that reduces hallucinations. Messy, scattered or outdated documents add cost before any AI work begins.

4. Guardrails And Human Approval

Guardrails are the limits on what an agent can access and do, such as read-only access to some systems, spending limits, or a rule that a person approves any payment. Designing these, plus a simple screen where a person can approve or correct the agent's work, is a real part of the build. The more an action can cost you if it goes wrong, the more control it needs, as our guide to AI agent guardrails explains.

5. Evaluation And Testing

Evaluation means checking the agent's work against a set of real examples with known correct answers before it goes live, then again after every change. Building that test set takes time from the people who know the work. It is also the step that tells you whether a cheaper model is good enough, which can lower the running cost for the life of the agent. Running an AI proof of concept before the full build is a low-cost way to create this test set and check the approach works.

6. Security, Privacy And Compliance

Agents that handle personal data, payments or regulated information need access controls, audit logs (a record of what the agent did and why) and sometimes a review by your compliance or legal team. These add build time but also protect you from far larger costs later.

Our guide to the anatomy of a production AI agent explains how these components fit together in a working system.

How To Estimate The Build Cost

Whether you hire an agency or build in-house, custom AI agent development cost comes down to one simple formula: hours of work multiplied by the hourly rate of the people doing it. Every quote you receive is a version of this, whether or not it shows the working. Asking a provider to break their price into hours per task is the best way to compare quotes fairly.

Rates vary widely by location and type of provider. As a reference point, the US Bureau of Labor Statistics reports a median annual wage of $135,980 for software developers in May 2025 (BLS (opens in a new tab)). That figure is salary only. An agency or consultancy rate is higher because it also covers benefits, management, tools, testing and profit, so always compare quotes on hours and scope as well as the total.

The table below shows how the hours might split for a narrow agent that handles one workflow and connects to two systems. The hours are planning assumptions for illustration, not a quote, and your workflow could need more or less.

At an assumed average rate of $120 per hour across the team, 300 hours comes to $36,000. Change either number and the total moves with it, which is why you should ask how each line was estimated rather than accepting a single figure.

AI Agent Cost Per Task: What It Costs To Run

Running cost is what you pay each time the agent completes a task. The biggest variable part is model usage, which AI providers charge per token. A token is a small chunk of text, roughly three quarters of a word in English. Providers charge separately for input tokens (what you send the model, including instructions, documents and the conversation so far) and output tokens (what the model writes back).

Here are published prices for two large and two small models from Anthropic and OpenAI, checked on 19 October 2026. Prices change often, so always confirm them on the provider's own page before budgeting. OpenAI charges more for very long prompts, so its rates below are for shorter ones.

Why One Task Means Many Model Calls

This is where many AI agent cost estimates go wrong. A chatbot answer is usually one model call. An agent task is several: the agent reads the request, decides what to look up, calls a tool, reads the result, decides again, and so on until it finishes. Each call resends the agent's instructions and tool descriptions, plus everything gathered so far, so input tokens add up quickly.

To work out the AI agent cost per task, estimate four numbers: the number of model calls per task, the average input tokens per call, the average output tokens per call, and the price per million tokens. Multiply calls by tokens, then by price. Then add an allowance for retries, because agents sometimes repeat a step after an error.

Here is the math behind Figure 2. Early calls send fewer tokens and later calls send more, so the example uses an average of about 8,000 input tokens per call. Six calls at that average is 48,000 input tokens, which costs $0.096 at $2 per million. Six calls of 500 output tokens each is 3,000 output tokens, which costs $0.03 at $10 per million. The total is about $0.13 per task. The same task on a small model priced at $0.10 and $0.50 per million costs about $0.006, roughly twenty times less, if the small model is accurate enough for the job.

Two Ways Providers Lower The Price

Prompt caching lets the provider store the part of the prompt that repeats on every call, such as the agent's instructions and tool descriptions, and charge much less to reuse it. On Anthropic's pricing page, reading from the cache costs 5% of the normal input price on Claude Sonnet 5.5, while writing to it costs 25% more than normal input. In the example, if 5,000 of each call's 8,000 input tokens repeat, the first call writes them to the cache and the next five read them back, which brings the task down to about $0.08.

Batch processing suits work that doesn't need an instant answer, such as overnight document processing. Anthropic's Batch API gives "a 50% discount on both input and output tokens" in exchange for slower results (Anthropic pricing (opens in a new tab)).

The Rest Of The Running Cost

Model usage is only one part of the running AI agent cost. You also pay for hosting the agent's code, any databases it uses, logging and monitoring tools that record what the agent did, and the time people spend reviewing the cases the agent hands over. Hosting and tools are usually a modest monthly amount. Review time can easily be larger than the model bill, as the example below shows.

AI Agent Pricing: How Ready-Made Agent Products Charge

Not every business needs a custom build. Many software platforms now sell ready-made AI agents, especially for customer service, and they price them differently from custom work. Instead of a build fee, the AI agent cost is usually charged per result, per conversation or per unit of usage, on top of any subscription you already have. These are the published prices for four well-known products, checked on 19 October 2026:

Per-result AI agent pricing is easy to budget at low volumes and needs no build, so it often wins for standard tasks inside a platform you already use. The trade-offs are cost at high volumes and limited reach. At $0.99 per outcome, 1,500 resolved requests a month cost about $1,485 a month, or about $17,800 a year, before any subscription. A packaged agent also only reaches the systems and actions its platform supports, and most are built for customer service. A task like invoice matching across inventory and accounting software usually needs a custom build, which is why the worked example below uses one.

The Costs Most AI Agent Estimates Leave Out

Many quotes for AI agent development cost stop at launch. These are the costs that arrive afterwards, and they decide whether the agent pays for itself.

• Human review time. Someone has to check the cases the agent is unsure about or is not allowed to finish alone. Multiply the share of tasks that need review by the minutes each review takes. Our guide to human-in-the-loop AI explains how to keep that review quick and useful.

• Re-testing after changes. Each change to the agent's instructions, tools or model can fix one problem and cause another. Re-running the test set after every change takes developer and reviewer time. Accuracy is a real concern: in the Federal Reserve Banks' 2026 survey of small employer firms, accuracy was the challenge most often reported by businesses using AI, cited by 46% of them (Small Business Credit Survey (opens in a new tab)).

• Model and API changes. Providers release new models, retire old ones and change prices. Systems your agent connects to also update their APIs. Each change means testing and sometimes code changes.

• Monitoring. Someone needs to look at logs, error rates and cost per task each week, and act when something drifts.

• Improvements. Once people use the agent, they will ask for more. New tasks are new development, not maintenance, and should be priced separately.

• The cost of mistakes. A wrong payment, a misrouted order or a poor customer reply has a cost. Guardrails and review reduce it but never remove it entirely.

These ongoing costs follow the same pattern as other software. Our guide to software maintenance cost after launch explains how to budget for them.

A Worked Example: Year-One Cost Of An Invoice-Matching Agent (Illustrative)

This example is hypothetical. The business, volumes and rates are assumptions chosen to show the method, not figures from a Fossilite client. Replace them with your own numbers.

A 40-person wholesale distributor receives about 1,500 supplier invoices a month. A finance assistant matches each invoice to its purchase order and goods receipt (the record that the goods arrived) in the company's inventory system, then records it in the accounting software. Each invoice takes about 6 minutes, or about 150 hours a month. At an assumed staff cost of $35 per hour including benefits, that work costs about $5,250 a month, or $63,000 a year.

The proposed agent reads each invoice, looks up the matching records, compares quantities and prices, and drafts the accounting entry. Invoices that match are recorded after a quick batch check, assumed here to take only a few minutes a day, so it is left out of the figures below; include it in your own estimate. About 15% have a difference, such as a short delivery or a price change, and go to the finance assistant with the agent's explanation. Each of those takes about 4 minutes to review.

Here is the year-one AI agent cost for a custom build, using the figures from earlier in this guide:

Three things stand out. First, model usage is the smallest line, under 3% of the year-one total. Switching to a small model would save about $1,600 a year here, if it proved accurate enough, which matters far less than the build and maintenance hours. Second, year one roughly breaks even against the $63,000 the manual work costs today. In practice the agent goes live partway through the year, so the manual cost continues during the build and year one is somewhat less favourable than this. Third, the picture changes in year two, when there is no build cost.

From year two, the agent costs about $27,660 a year to run and maintain, against $63,000 for the manual process, a difference of about $35,000 a year. That saving only becomes real if the finance assistant's freed time goes to useful work, such as chasing supplier credits or closing the month faster.

Volume matters most here. At 300 invoices a month instead of 1,500, the manual work would cost about $12,600 a year, and the agent would not pay back even in year two. This is why the volume and time per task belong in any AI agent cost estimate before anyone discusses technology.

How To Keep AI Agent Costs Under Control

The following habits lower costs without cutting the controls that keep the agent safe:

• Start with one narrow task. Prove the agent on one workflow with a clear measure before adding more.

• Use the smallest model that passes your tests. Use your test set to compare a small model with a larger one. Some agents use a small model for simple steps and a larger one only for hard decisions.

• Cap the steps. Set a maximum number of model calls per task, so a confused agent stops and hands over instead of looping.

• Keep prompts lean. Every word in the instructions is resent on every call. Remove what the agent doesn't need, and cache what repeats.

• Batch work that can wait. Overnight jobs can use cheaper batch pricing where the provider offers it.

• Track cost per completed task. Divide the monthly bill by the tasks finished correctly, not just tasks attempted. Set a budget alert with your model provider.

• Review the review rate. If people are checking far more cases than planned, the agent needs work or the task is a poor fit.

Questions To Ask Before You Accept A Quote

Use these questions to compare quotes on equal terms:

1. 1

   Is this an agent that decides its own steps, or a workflow with fixed steps? Why is that the right choice for our task?

2. 2

   How many hours are in each part of the build, and what rate applies?

3. 3

   What is the expected model cost per task, and what assumptions about calls and tokens is it based on?

4. 4

   Which model will you use, and how will you test whether a cheaper one is good enough?

5. 5

   What share of tasks do you expect a person to review, and how long will each review take?

6. 6

   What guardrails will limit what the agent can do, and which actions need human approval?

7. 7

   How will you test the agent before launch and after every change?

8. 8

   What does maintenance cover each month, and what counts as new work?

9. 9

   Who owns the code, the instructions and the test set if we part ways?

A good provider will answer these clearly and break the AI agent cost into build, running and maintenance lines. SensViz's guide on how to choose an AI agent development company (opens in a new tab) describes what good answers to questions like these sound like. For a broader checklist, see our guide on how to choose an AI consulting partner.

How Fossilite Approaches AI Agent Cost

We start with the workflow, not the technology. Before estimating any AI agent cost, we look at how the work runs today, how often it happens, how long it takes and what a mistake costs. That tells us whether an agent is worth building, or whether rules-based automation or a single AI step would do the job for less.

When an agent is the right fit, we estimate build, running and maintenance costs separately and show the assumptions behind each, including the expected cost per task. We build the test set early, compare models on your own examples, and put human approval where the consequences call for it. The aim is an agent whose cost you can predict and whose results you can measure.

Frequently Asked Questions

How Much Does It Cost To Build An AI Agent?

The cost to build an AI agent depends mainly on how many tasks it handles, how many systems it connects to and how much testing and control it needs. Estimate it as hours of work multiplied by an hourly rate. In this guide's illustrative example, a narrow single-workflow agent takes about 300 hours, or $36,000 at $120 per hour.

How Much Does An AI Agent Cost Per Month To Run?

Monthly running cost combines model usage, hosting and tools, people's review time and maintenance. Model usage depends on task volume, calls per task and token prices. In this guide's illustrative example of 1,500 invoices a month, running and maintaining the agent after launch costs about $2,300 a month, and model usage is only about $140 of that.

What Is The Most Expensive Part Of An AI Agent?

For most narrow business agents, the build is the largest cost in year one, driven by integrations, testing and guardrails. After launch, people's time for maintenance and review usually costs more than model usage. Model fees become the main cost only at very high volumes or when tasks use very long documents.

Is A No-Code Agent Builder Cheaper Than A Custom AI Agent?

A no-code agent builder is usually cheaper and faster to start, and it suits simple tasks within the tools it supports. Costs rise when you need connections it doesn't offer, strict controls, detailed testing or high volumes priced per task. Compare the total cost over two or three years, not just the first month.

How Long Does It Take To Build An AI Agent?

A narrow agent for one workflow can often move from idea to a monitored pilot in a matter of weeks, depending on integrations and how quickly test examples are available. Agents spanning several departments or regulated data take longer. Our guide on moving from AI demo to production in six weeks describes one such path.

Can A Small Business Afford An AI Agent?

A small business can afford an AI agent when the task is frequent enough for the savings to cover the build and running costs. Calculate what the work costs today in hours, then compare it with the total AI agent cost over two years. For low-volume tasks, simpler automation or a single AI step is usually the better value.

Will AI Model Prices Go Down?

Providers change model prices and release new models often, and newer small models are frequently much cheaper than older large ones. Prices can also rise for the most capable models. Build the agent so the model can be swapped and re-tested, and recheck the provider's pricing page before each budget review.

If you're still deciding where an agent could help, our AI readiness assessment is a good next step.