"How much does an AI agent cost" is almost always an incomplete question. The build is a one-off payment and it is the number you see in the proposal. After that, model usage, infrastructure and maintenance tick over every month — and it is that second part that decides whether the project pays for itself. We at SUDUS·TEAM have been building AI agents for business since 2023, and we have seen enough budgets to count honestly.
Below is the full cost structure of an AI agent in 2026: what the build costs at three levels of complexity, how to calculate the monthly model bill, which costs never appear in proposals, and the conditions under which an agent will not pay for itself at all.
What an AI agent budget is actually made of
Every AI budget has four parts. Two are one-off, two are permanent. The mistake in most calculations is looking only at the first.
1. The build — one-off
Process audit, agent design, system prompt, assembling the RAG knowledge base, integrations, guardrails, testing, launch. This is the fixed figure in the contract. With us it starts at $900 for a single-purpose agent.
2. Model usage — a monthly variable
Every call to the language model is billed by volume of text: separately for what the agent read (input tokens) and what it wrote (output tokens). This is the least predictable line, because it depends on the number of conversations and on how carefully the RAG base was designed.
3. Infrastructure — monthly
Vector database for RAG, hosting for the service layer, queues, logging, monitoring. For a small agent that is $15–50 a month; for a corporate system with an on-premise model it is a separate story with its own servers.
4. Maintenance and development — monthly or in packages
An AI agent is not a static product. Prices change, new services appear, customers ask questions nobody asked before. The prompt and the knowledge base need regular updating — otherwise within six months the agent starts confidently telling people things that are not true.
What the build costs: three levels
The price depends not on "number of pages", as with a website, but on the number of integrations, the size of the knowledge base and how many agents have to work together.
Single agent — from $900, 14–25 days
One agent, one job: answering customers on your site, triaging inbound email, qualifying leads, or an assistant inside a Telegram bot. A RAG base of up to 100 pages of your documentation, integration with one or two systems, a handful of tool functions. The cheapest way to find out whether AI works in your process at all.
Multi-agent — from $3,000, 35–60 days
A team of specialised agents handing work to each other: one qualifies the lead, another drafts the proposal, a third books the meeting in a manager's calendar. An extended base of up to 1,000 pages, MCP integrations with CRM, ERP and accounting systems, an evaluation harness on your real cases, a full analytics dashboard. This is the level at which AI stops being a clever FAQ and becomes part of the process.
Enterprise — from $7,500
Dozens of agents across departments, private on-premise model deployment, fine-tuning on your corpus, voice agents, a security audit, role-based access. Relevant where the data cannot leave the building: fintech, medical networks, companies under strict regulation. The annual figure here is driven by the private infrastructure and the dedicated support team, and is agreed case by case.
The full package comparison is on the pricing page.
Tokens: how to calculate the monthly bill
This is the part where most contractors answer "oh, not much". In fact it can be worked out fairly precisely.
What a token is and how many are in one conversation
A token is a chunk of text of roughly 3–4 characters. One A4 page is around 700–900 tokens. On every turn the agent receives the system prompt, the conversation history and the retrieved RAG fragments. Those fragments make up most of the volume.
A realistic band for one full customer conversation of 6–10 turns:
- Input tokens: 20,000–60,000
- Output tokens: 1,000–3,000
Example: a support desk handling 1,500 conversations a month
Take an average online store: 1,500 conversations a month, 70% of them typical and short, 30% complex with deep retrieval.
- On a top-tier commercial model — around $110–220 a month
- On a compact model with the same RAG — around $15–40 a month
- A hybrid arrangement (compact model for typical questions, top-tier for complex ones) — around $30–60 a month
The gap between the first and second option is up to sevenfold, at identical quality on 70% of enquiries. Which is why the choice of model is not a technical detail — it is a budget line. Our own published band is $40 to $400 a month depending on volume.
What actually brings token costs down
- Model routing. A simple classification of the incoming question: typical goes to the cheap model, complex to the expensive one. The largest effect for the least effort.
- Careful chunking of the RAG base. If the agent pulls twelve fragments of 1,500 tokens where three of 400 would do, you pay for that every single day.
- Caching the system prompt. The stable part of the context should not be paid for again on every request.
- Capping history length. After a certain number of turns, older messages should be summarised rather than carried in full.
In practice those four together cut the bill by 60–80% with no loss of answer quality.
Hidden costs nobody puts in a proposal
Preparing the data for RAG
The most common cause of a missed deadline. The company "has all the documentation", but in reality that is 200 files across three different clouds, half of them out of date, some contradicting each other, and the current price list lives in the sales director's head. Cleaning that up is genuine work — 10 to 40 hours — and somebody on your side has to do it.
Integrating with old systems
Connecting to a modern CRM with a sane API is a day or two. Connecting to an in-house accounting system from 2012 with no documentation is two weeks and a separate line in the budget. If that is your situation, put it in the budget up front rather than discovering it mid-project. We often solve it by building a custom CRM as an intermediate layer.
Your experts' time
For the agent to answer correctly, somebody on your team has to read its answers on real cases and say where it is wrong. That is 15–30 hours from a person who genuinely knows the subject. It is never free — you simply pay your own employee rather than the contractor.
What twelve months actually costs: three scenarios
Year-one total cost of ownership is the figure that belongs next to the expected saving.
Scenario A: small business, support agent
Build $900 + model usage about $40 a month + infrastructure about $20 a month + maintenance from $80 a month.
Year one: from about $2,600.
Scenario B: mid-sized business, multi-agent in sales
Build $3,000 + model usage about $150 a month + infrastructure about $50 a month + maintenance from $175 a month.
Year one: from about $7,500. Payback here is not measured in salary saved but in additional qualified leads — with an average deal of $3,000, seven extra deals cover it.
Scenario C: enterprise ecosystem
Build from $7,500, plus your own infrastructure for locally hosted models and dedicated support. Here the deciding factor is not the ROI of a single task but the combined effect across dozens of processes, and security requirements that cloud solutions simply do not meet.
AI agent, outsourced support, or a new hire
AI is almost never compared with the alternatives honestly. The comparison is easy to make yourself, and it works in any market, because the only number that changes is the salary.
Take the same job — handling 1,500 enquiries a month — over the first year:
- A new hire. Take the fully loaded cost: gross salary plus employer taxes, workspace, equipment, training, holiday and sick cover. Covers roughly 100% of enquiries during working hours and 0% at night and at weekends.
- An outsourced call centre. Billed per handled enquiry. Works around the clock, but the operator does not know your product deeply and quality swings from shift to shift.
- An AI agent. From about $2,600 in year one, as calculated above. Covers 70–85% of enquiries around the clock and passes the rest to a human with the context already gathered.
In most markets the agent comes out several times cheaper, but there is an important caveat: the 15–30% that remain still need a person. So the real saving is not "instead of an employee" but "one employee instead of three", plus round-the-clock availability. That is how the ROI should be calculated.
How to read a proposal for an AI project
AI proposals differ wildly in completeness, and comparing them on the bottom-line figure is pointless. Here is what should be written down — and what to ask about if it is not.
- Who pays for model usage. The most commonly hidden item. Find out whether the provider bills you directly, whether it is inside the maintenance package, and what happens if the volume is exceeded.
- Size of the RAG base, in pages. A "knowledge base" with no number attached is not a scope of work. Ours is up to 100 pages on the starter package and up to 1,000 on the professional one.
- Integrations, named individually. Not "integration with your systems" but "CRM X over REST API, mailbox Y over IMAP".
- Number of tool functions. How many actions the agent can actually perform, as opposed to merely advising.
- Acceptance criteria as numbers. Accuracy on a control set of questions. Without it you cannot prove the result is unsatisfactory.
- What free support covers and how long it lasts. With us the first two months after launch are part of the work, plus a 6-month warranty on what was delivered.
- Who owns the prompts, the base and the code. If you cannot take the system and hand it to another contractor, you are on a hook.
A proposal missing points 1, 2 and 5 almost guarantees additional charges at the stage where backing out is no longer an option.
When an AI agent will not pay for itself
The honest part. There are situations in which we turn clients down ourselves:
- Fewer than 200 repetitive enquiries a month. The saving will not even cover maintenance. Hiring someone part-time is cheaper.
- No structured knowledge base and no intention of creating one. An agent without data is a generator of confident fiction.
- Every case is unique. With no repetition, there is nothing to automate.
- The expectation of "replacing the whole department". AI removes routine and the first line; it does not take responsibility for decisions.
If your task is simpler than it looks and comes down to scripted answers and buttons, what you may need is not an AI agent but an ordinary Telegram bot at a fraction of the cost — the difference is set out in AI agent or Telegram bot.
Frequently asked questions
How much does an AI agent cost for a small business?
A realistic entry budget is $900 for the build and $140–200 a month to run it, so from about $2,600 in the first year. The second year is cheaper, because the build does not repeat.
Can we start with a smaller budget?
You can start with a narrow pilot on one process: not "an agent for the whole company" but "an agent that answers the five most common customer questions". It is faster, cheaper, and produces numbers you can actually base a scaling decision on.
How much does monthly maintenance cost?
For a single agent, $140–200 a month including model usage and infrastructure. For a multi-agent system, $400–800. The first two months after launch, when we read real conversations and keep tuning the knowledge base, are part of the build.
Which is cheaper, an AI agent or a new hire?
Over a year the agent is several times cheaper in most markets, provided the volume of enquiries is stable and the task is repetitive. Over three months the hire is cheaper, because there is no build stage. AI makes sense where you are confident about volume years ahead.
Is the model billed separately?
Yes, model usage is billed by the provider separately from the build — typically $40 to $400 a month depending on volume. We set up metering and limits so the bill is predictable and an overrun is technically impossible.
In short
The right question is not "how much does an AI agent cost" but "how much does owning one cost per year, and what do I save for that money". For a small business with a real flow of enquiries the answer is usually in the agent's favour — from about $2,600 in year one against the fully loaded cost of one person doing the same job. For a company with twenty enquiries a month the answer is the opposite, and an honest contractor will say so at the first meeting.
Want the figure for your own process? Write to us — the free consultation takes 30 minutes and ends with numbers, not a slide deck.