The price of an agent is not the model. It is the integration you pay for once, the usage you pay for monthly, and the waste that hides between the two. Here is the honest anatomy.
What an AI agent costs: the short answer
An AI agent has two cost lines. The build is a one-time project, usually a few weeks of integration work, that connects a model to your calendar, CRM, WhatsApp or invoicing system and writes the rules for what it may do. The run is a monthly figure made of model usage, hosting and a small share of someone's time to own it, and for a small business it typically lands in the range of a software subscription rather than a salary. The number that decides whether the agent was worth it is neither of those; it is the value of the task it takes over, measured before you sign.
Vendors who quote a single price are usually hiding one of the two lines. Ask which one.
The build cost: integration, not intelligence
The model is the cheapest part of an agent. What you pay for in the build is everything around it: an API connection to each system the agent must read or write, a definition of the tasks it owns and the ones it hands to a person, test conversations for the awkward cases, and a log so you can see what it did. Every extra system adds roughly a unit of work, so an agent that only answers questions from a document costs a fraction of one that books, quotes and updates records.
The other build cost never appears on the invoice: your time. Someone inside the company has to explain how enquiries actually get handled, which is rarely how the process document says. Budget a few hours of the person who knows, early, or pay for it twice later.
The running cost: tokens, platform and an owner
Model usage is priced per token, which in practice means per conversation. A typical customer exchange with a mid-tier model costs a few cents, and even a busy small business rarely spends more on tokens in a month than it spends on coffee. The platform that hosts the agent, keeps its memory and connects the channels is a fixed subscription. The line businesses forget is ownership: someone must read the escalations, correct wrong answers and update the agent when prices or policies change. An hour or two a week, not a job, but not zero.
Costs rise with volume and with reasoning depth. An agent that thinks through a multi-step task uses many times the tokens of one that answers a question, and an agent left to retry a failing tool call can burn a day's budget in an afternoon. Spending caps are part of the build, not an optional extra.
Where the money is wasted
Gartner predicted in June 2025 that more than 40 percent of agentic AI projects would be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls; according to the same analysis, most current projects are early experiments driven by hype, which blinds companies to the real cost of running an agent at scale. Waste has a recognisable shape. Scope: an agent built to handle everything, which means it handles nothing reliably and the team stops trusting it. Overkill: a rule that sends a reminder three days after a quote does not need a model. The pilot that never ends: a demo impresses, nobody names an owner or a metric, and the monthly fee continues without a result attached to it.
The fourth is paying to integrate a system you were about to replace. If the CRM is a shared spreadsheet, the agent inherits its mess. Fix the source of truth first, then connect the agent to it.
AI agent cost by scope: what the number depends on
Five variables move the price more than anything else. Channels: WhatsApp only, or WhatsApp plus web chat plus email, each with its own connection and edge cases. Systems: how many tools the agent reads from and writes to, and whether they have a usable API. Autonomy: whether the agent drafts for a person to approve or acts alone, which changes the testing and the guardrails. Languages: an agent that must hold Malay, English, Arabic and German needs more test conversations, not more model. Volume: running cost scales with conversations, so two hundred enquiries a day and twenty are different purchases.
A bounded first agent for a small company, one channel, two systems, drafting with approval, sits at the low end of every variable. It is also the version most likely to pay back, because it replaces a task you can measure. Add autonomy and channels once the first one has earned trust.
A payback calculation you can do on one page
- Name the task, not the technology. One sentence: the agent answers new WhatsApp enquiries within a minute, qualifies them and books a call. If you cannot write that sentence, you are not ready to price anything.
- Count what the task costs today. Hours per week of the person who does it, multiplied by their loaded cost, plus the enquiries that went nowhere because nobody replied in time. The second number is usually larger and usually unmeasured.
- Put the build cost against the monthly saving. A build that costs three months of the saving is a strong case; one that costs eighteen months of it needs a smaller scope before you start.
- Add the running cost honestly: model usage, platform and the owner's two hours a week. If it exceeds a third of the saving, the task is too small or the agent is too heavy.
- Set a metric and a date. Response time, booked calls, quotes issued. Review at sixty days and decide to extend, cut or stop, on numbers rather than on how the demo felt.
AI agent cost: common questions
- How much does an AI agent cost for a small business? — A bounded first agent, one channel and one or two connected systems, is a one-time build of a few weeks of specialist work plus a monthly running cost comparable to a business software subscription. The build dominates in year one; the running cost dominates from year two and scales with conversation volume.
- Is an AI agent cheaper than hiring someone? — For a narrow, repetitive, high-volume task such as first response and qualification, the running cost is usually a fraction of a salary. For judgement-heavy work it is not a replacement; the useful question is what the person could do with the hours the agent gives back.
- What are the hidden costs of AI agents? — Internal time to define the process, integration to systems without a clean API, the owner's weekly maintenance, token spend from runaway retries, and a pilot that continues without a metric. Spending caps and a named owner remove most of them.
- How long does an AI agent take to pay for itself? — A well-scoped agent that takes over a measurable task typically shows its return within the first few months, which is why we recommend a sixty-day review against a metric agreed before the build. An agent without a metric never pays back, because nobody can tell whether it did.
Keep reading
Build something
worth trusting.
Tell us what's slowing your business down. We'll show you the system that fixes it — and how fast.