How much does it cost to build an AI agent? Build cost, running costs and why projects fail

"How much does it cost to build an AI agent?" is usually the first question, and the honest answer starts with another one: an agent that does what? AI agent development cost depends far less on the model than on the work around it: the systems it connects to, the information it answers from, the approvals it needs and the testing that proves it works. Then come the running costs, which arrive every month whether the project succeeded or not.

This guide is for business owners and product leads budgeting for a first agent. It explains what drives the build cost, what you pay to run one, why so many agent projects get cancelled, and how to start small enough to find out quickly whether it deserves more. You will not find invented price ranges here; you will find the drivers that decide the price.

I build AI agents for businesses and have built AI into products of my own. If you want to know what an agent is and what it can do before talking money, read what AI agents can do for a business first.

Why there is no honest price list for AI agents

"AI agent" covers everything from a chat window that answers from your FAQ to a system that reads orders across three marketplaces and drafts refunds for approval. Published price ranges vary enormously for exactly that reason, and a range wide enough to cover both ends tells you nothing about your project.

What you can know before any quote is which of the drivers below apply to you, and how heavily. They are also what a developer should ask you about before giving you a number.

What drives AI agent development cost

1. Scope. The number of jobs. "Answer delivery questions" is one job. "Answer delivery questions, process returns and update the CRM" is three, each with its own rules, tests and failure cases.
2. Integrations. Every system the agent touches, such as Shopify, Seller Central, a calendar, a CRM or a helpdesk, needs authentication, data mapping, rate-limit handling and a plan for when it is down. Write access costs more than read access, because each action needs checks.
3. Data. If your policies, product facts and past answers are written down and current, grounding the agent in them is quick. If they live in people's heads, writing them down becomes part of the project.
4. Approvals. Anything that commits the business needs a step where a person approves, and a record of who approved what. In Clearday, the AI life organiser I built, the assistant waits for the user's approval before it touches their calendar. Designing that step well is real interface work, not a checkbox.
5. Evaluation. A set of real inputs with known good answers, run every time a prompt or model changes. Without it, nobody can say whether the agent is getting better or worse.
6. Guardrails. Limits on what the agent may say and do, refusal of off-topic or manipulative requests, hand-off to a person, and usage caps so a bug cannot run up a bill.
7. Channels and languages. A website widget, WhatsApp, email and an app are separate front ends on one backend. Each language needs its own testing with real messages.
8. Models. One provider is simpler. Several can be better: ProAI Training uses OpenAI, Gemini and Claude side by side, each for the part of the job it handles best, but that means more to test and maintain.

AI chatbot vs AI agent: why the cost differs

	Answers (chatbot)	Looks things up (read-only agent)	Takes actions (agent with write access)
Example	Shipping and returns questions from your policies	"Where is my order?" from live order data	Starts a return or changes a booking
Main work	Knowledge base, tone, hand-off	API connections, identity checks	Approvals, limits, audit log, undo
Main risk	A wrong answer	Showing data to the wrong person	A wrong change, repeated at scale

Each column to the right adds integration, permissions and testing. Most businesses do best moving left to right over time rather than starting on the right.

Running costs: what you pay every month

Model usage. Providers bill per token, roughly per word read and written, with different rates for input and output and for each model. Rates change often, so check the current pages rather than any article, this one included: OpenAI, Anthropic and Google Gemini.

To estimate, multiply conversations per month by tokens per conversation. Remember what an agent sends on every turn: its instructions, the relevant knowledge, the conversation so far and any tool results. An agent that calls three tools before answering makes several model calls for one reply. Caching repeated content and using a smaller model for simple steps both bring the figure down.

Hosting. A server for the agent, a database, perhaps a search index over your documents, and a queue for slow jobs. Small for a quiet agent, larger for heavy ones, never zero.

Monitoring and maintenance. This is the AI agent maintenance cost people forget: logs and alerts, a weekly read of real conversations, re-running the evaluation set, and updates when something underneath changes. Providers retire older models, and platforms change their APIs; Shopify, for example, releases a new API version every quarter and supports each one for at least 12 months. An agent nobody maintains drifts.

Your team's time. Someone has to own the knowledge base and pick up the hand-offs. Budget for that every week, especially in the first months.

Why AI agent projects fail

In June 2025 Gartner predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. It also warned about "agent washing", older chatbots and automation tools relabelled as agents, and estimated that only about 130 of the thousands of vendors claiming agentic AI offer the real thing.

Each of those three reasons maps to a decision you can make before building:

• Escalating costs: set a budget per conversation and per month, enforce caps in the code, and measure from the first day.
• Unclear value: write down the number the agent should move before you start, such as hours a week spent on order-status emails or time to first reply.
• Weak risk controls: read-only first, approvals for actions, a log of everything and an obvious way to switch it off.

A fourth reason is simpler: trying to do everything at once.

How to start small

1. Pick one job with a number attached, for example the order-status questions that eat part of every day. Measure the baseline first.
2. Collect real examples. A few dozen real messages or tasks, each with what a good result looks like. This becomes your evaluation set.
3. Build it read-only. The agent answers and looks things up; it changes nothing.
4. Run it in draft mode. Staff see its suggested replies and send, edit or discard them. How often they edit tells you how close it is.
5. Go live on one channel, with a clear route to a person.
6. Add one action at a time, each behind an approval step, once the numbers justify it.

This is the same way I approach any product: the smallest useful version first, released in pieces, so you can try something early while changes still cost hours, not weeks.

Build or buy, and how I quote

Buy when the job is standard and a tool you already use does it well: a helpdesk's AI add-on, Shopify's Sidekick, Amazon's Seller Assistant. Many charge per resolution or per seat, which is easy to start with and worth modelling at your real volume. I compare that pricing for stores in AI customer service for online stores.

Build when the agent needs your systems, your rules or several channels at once, or when you want to own the code and the data. For online stores, common first agents are order-status answers, listing drafts for a person to approve and plain-English summaries of ad reports; my e-commerce automation page shows the store systems I connect.

Because AI agent development cost depends so heavily on the drivers above, I quote agents from a short brief rather than a price list. My pricing page lists starting prices for websites, mobile apps, Shopify apps and marketplace tools, and AI agents are not on it for that reason. Every project gets a fixed quote before work starts, payment in milestones that are each something you can open and try, and the code and accounts in your name. For AI already working inside a real product, see Clearday, whose assistant asks before it changes anything.

FAQ

How much does it cost to run an AI agent each month?

It depends on volume and design, not on a fixed fee. The main costs are model usage, billed per token by providers such as OpenAI, Anthropic and Google; hosting for the agent and its data; and the time spent monitoring and maintaining it. Estimate by multiplying monthly conversations by tokens per conversation at current provider rates, then add hosting and review time. Set hard usage caps so a bug cannot surprise you.

Is it cheaper to use a ready-made AI tool than to build an agent?

Often, at first. A ready-made tool costs little to start and suits standard jobs inside a system you already use. It becomes expensive when per-resolution or per-seat fees grow with your volume, or when it cannot reach the data or follow the rules your business depends on. Compare the tool's cost at your real monthly volume with a custom agent's build and running costs before deciding.

What should I prepare before asking for an AI agent quote?

Describe one job the agent should do, in your own words. Gather a few dozen real examples of the messages or tasks involved, with good answers. List the systems that hold the information and who can grant access. Say which actions need a person's approval, which channels matter and which number you want to improve. With that, a developer can give you a fixed quote rather than a guess.

Ready to price one? Send me a short brief with that job, the systems involved and when you need it working, and I will come back with a fixed quote. You can also see the AI agents for business service, or read integrating AI into your app for the technical side of evaluation, fallbacks and cost limits.