How much to charge for an AI agent
How much to charge for an AI agent comes down to two questions: what job does it take off someone's desk, and how much is it trusted to do without a person checking? The job sets the value. The trust level sets the tier. Quote a fixed setup fee for building it and a monthly fee for running and supervising it, both sized as an obvious fraction of what that job costs the client today, and put model usage on its own line. Do not quote build hours. With Claude Code, the first working version of an agent can come together fast, and a client who hears how fast will price it like a quick job, even when it is doing the work of part of a person.
This is the CCC Role Pricing Method, and it exists because agents are the offer where new builders get the shape of the price wrong, not just the number. They quote a one-off build like a website, hand it over, and then spend the next six months answering messages about why the agent did something odd on a Tuesday. An agent is closer to a hire than a deliverable. It needs a job description, limits, someone checking its work, and a budget for what it spends. Price all four, or you end up providing the last three for free.
An agent is not a chatbot or an automation, and the price shows it
The difference is who decides the steps. A chatbot answers questions from a fixed source. An automation runs the same steps every time: when a form comes in, add a row, send an email. An agent is given a goal and decides the steps itself, reading, choosing a tool, taking an action, checking the result, and going again. That is what makes an agent valuable, and it is also what makes it riskier, because it can take an action nobody wrote down in advance.
Chatbot, automation, or agent
| Chatbot | Automation | Agent | |
|---|---|---|---|
| What it does | Answers questions | Runs fixed steps | Decides its own steps and takes actions |
| What it touches | A knowledge source, sometimes one system | The systems in its recipe | Any tool it has been given access to |
| What goes wrong | A wrong answer | A broken step when an app changes | A wrong action taken with real access |
| What you price | The tier of what it is wired into | The hours it removes | The job it does and the trust it is given |
Say this out loud on the call. Plenty of clients ask for "an agent" when they need an automation, and the automation is cheaper, safer and faster to deliver. Telling them so is the fastest way to be trusted with the agent they will want next year.
Write the job description before you write a number
You cannot price an agent until you can describe its job the way you would describe a role you were hiring for. Five questions get you there, and the client should answer each one in their own words.
- What is the job? One sentence: triage the support inbox and resolve the routine tickets, qualify inbound leads and book the good ones, reconcile the weekly orders against the invoices. If it takes a paragraph, it is two agents.
- Who does it now, and how much of their week does it take? Let the client say it. That person's time is the number your quote sits next to.
- What is it allowed to do alone? Read only, draft for approval, or take the action. This answer picks the tier.
- What must it never do? Refund above an amount, email a customer about legal matters, delete anything. These become written limits, and the limits are part of the deliverable.
- Who checks its work, and how often? If the answer is nobody, the honest quote includes you doing it, or the agent stays in the lowest tier.
Say the job description back before you quote. "So the agent handles the routine support tickets your two coordinators spend most of their mornings on, it can close a ticket alone but cannot issue refunds, and your ops lead reviews its log on Fridays" is a sentence the client has to agree with. Once they have, your price is a fraction of a number they already accepted. The full question set is in client discovery questions before you quote.
The CCC Agent Price Ladder: three tiers by trust
Present three tiers for the same job, and let the client choose how much they want the agent to do alone. The tiers are not more features. They are more trust, with more limits, logging and review built to earn it.
The CCC Agent Price Ladder
| Tier | What the client gets | How to price it |
|---|---|---|
| Assist | The agent does the work up to the action and stops: drafts the reply, prepares the booking, fills the record. A person approves each one with a click. Every action is logged | A setup fee that is a clear fraction of the job's cost, plus a modest monthly fee. The safest first step, and the easiest yes |
| Operate | Everything in Assist, plus the agent acts alone on the routine cases inside written limits and sends everything else to a person. A weekly review of its log with you, and a written list of what it may and may not do | The anchor tier. The monthly fee is priced against the hours it now saves outright, because nobody is approving every action anymore |
| Own | Everything in Operate, plus the agent runs the whole job end to end, including handing work to other tools or agents, with an agreed exception path and a monthly report on what it did, what it escalated and what it cost | Priced like the part of a role it replaces. This is the retainer, because someone has to own the limits as the business changes |
Most clients land on Operate once they see the table, because Assist still keeps a person clicking all day and Own asks for a level of trust they have not built yet. That is fine. Start them there, show them the log for two months, and the move to Own becomes their idea. The three-tier shape is the same one in how much to charge for an MVP and how much to charge for an AI workshop; the tiers change, the logic does not.
Put model usage on its own line
An agent spends money every time it works, and that spend moves with volume you do not control. Models are billed by usage. Claude Sonnet 5.5, for example, is priced at $2 per million input tokens and $10 per million output tokens on the Anthropic API. A busy month for the client is a busy month for the agent, and if usage is buried inside a flat monthly fee, their best month is your worst one.
- Best: the agent runs on the client's own API account, and they pay usage directly. You never carry the bill, and the key never leaves their control.
- Next best: you pass usage through at cost with a small handling margin, invoiced monthly with the numbers attached.
- If they insist on one flat fee: include a stated usage allowance and a written overage rate, and set a hard spend cap on the account so a runaway loop cannot become a surprise.
Whichever you choose, set a spend limit before the agent goes live and tell the client what it is. An agent with real access and no ceiling is the one thing in this business that can cost a client money overnight. Is Claude Code safe for client work covers how I explain access and limits to a client, and what to put in an AI project contract covers the clause that makes the usage line binding.
What moves an agent quote up or down
- The access it needs. Read-only access to one inbox is a small line. Write access to billing, a CRM and customer email is a large one, because every system it can change needs its own limits and its own tests.
- The cost of a wrong action. An agent that can send a refund or email a customer needs approval paths, logging and a way to stop it. The more a mistake costs, the more of the price is the guardrails, and that is correct.
- Volume. Ten cases a day and a thousand are different agents. Volume changes usage, review time and how carefully the exception path has to be built.
- How messy the inputs are. Clean forms are cheap. Free-text emails in three languages with attachments are not.
- Who reviews it. If the client has someone to read the log, the monthly fee is lighter. If that someone is you, it is in the price.
- Handover. If the client wants to own and change the agent themselves later, the documentation and training are their own line. How to hand off a Claude Code project covers what that includes.
The mistakes that turn an agent into an unpaid on-call job
- Selling it as a one-off build. An agent with no monthly fee is an agent you support for free. The monthly fee is for supervision and limits, not for hosting.
- Unlimited usage in a flat fee. The client's growth becomes your cost. Usage goes on its own line, every time.
- No written limits. If what the agent may not do is not written down, every odd action becomes an argument about whose fault it was.
- No log. If you cannot show what it did and why, you cannot defend it, improve it, or move the client up a tier.
- No off switch. Agree how the agent gets paused and who can do it before it goes live, not after it does something at 2am.
- Starting at Own. Full autonomy on day one means the first mistake happens with nobody watching. Start at Assist or Operate and let the log earn the upgrade.
- No deposit. How much deposit to charge before an AI project covers the number. The setup fee is a real project and gets paid like one.
How to present the agent number so it closes
Lead with the job description you agreed on, then put the three tiers under it, with the usage line and the spend cap written out beside them. "You said the routine tickets take most of your coordinators' mornings; here are three ways to hand that job to an agent, depending on how much you want it doing alone" is a sentence a client can say yes to on the call. Send it as a one-page proposal with the job in one line, the limits, the review rhythm, and the setup and monthly fees. If they push on price, the answer is the Assist tier, not a discount, and the "can you do it cheaper" objection has the words.
The agent that works becomes the client's favourite thing you have built, and the log you have been sending them every month is the best case study you will ever have. How to turn a one-off client into a monthly retainer is what happens next, and usually the next job is a second agent.
Short, practical drops on skills, MCP, agents, prompts, and more. No spam, unsubscribe anytime.
Frequently asked questions
How much should I charge for an AI agent?
Price it against the job it takes off someone's desk, not your build hours. Get the client to say what that job costs them now, then quote a fixed setup fee and a monthly fee that are an obvious fraction of it, in three tiers by trust: Assist, Operate, and Own. Put model usage on its own line.
What is the difference between an AI agent and an AI automation for pricing?
An automation follows fixed steps, so you price the hours it removes. An agent decides its own steps and takes actions with real access, so you also price the limits, logging and supervision that make that safe. That is why an agent always carries a monthly fee.
Should AI agent pricing include the model usage costs?
Keep usage separate. Models are billed by usage, and a busy month for the client means more usage. The best setup is the agent running on the client's own API account; the next best is passing usage through at cost with a small margin. If you must include it, state an allowance, an overage rate, and a hard spend cap.
Should I charge a monthly fee for an AI agent?
Yes. The monthly fee pays for reviewing the agent's log, adjusting its limits as the business changes, and fixing what breaks when connected tools change. Without it, you are supporting the agent for free.
Which tier should a new AI agent client start on?
Assist or Operate. Assist has a person approve every action, which is the easiest yes. Operate lets the agent handle routine cases alone inside written limits. Move to Own only after a couple of months of logs show it can be trusted with the whole job.
Last reviewed by Duncan Rogoff on September 29, 2026
