Everybody is buying AI agents — by PwC’s count, 79% of organizations already use them somewhere. Yet almost nobody publishes prices. Vendors say “it depends,” demos end before the invoice appears, and a small business owner is left guessing whether an agent costs a hundred a month or a hundred thousand to build.
This article puts real numbers on the table: the three ways agents are priced, honest budgets for the use cases small businesses actually buy, the running costs that surprise buyers, and a way to check the payback before you spend a dollar.
Why nobody quotes you a straight price
An agent’s price has three moving parts, and vendors often quote only one:
- The tool itself — the AI model and the platform around it
- The build — wiring the agent into your systems, data, and workflow
- The running — usage fees, maintenance, and improvements over time
A $99-a-month chatbot and a $40,000 custom build can both be called “an AI agent for support.” Both prices are honest for different problems. The useful question is not “what does an agent cost” but “what does MY task cost to automate” — which is why the sections below are organized by use case, not by technology.
The three ways to pay
Subscribe: ready-made agents ($20–$500/month)
The fastest route: AI features already built into tools you may use. Helpdesks with AI resolutions, CRMs with AI email drafting, no-code platforms that chain an LLM into your workflow.
- Cost: typically $20–$500 per month, or $0.50–$2 per AI-handled conversation
- Time to value: days
- The catch: the agent works the way the vendor built it. If your process has quirks — and every small business process does — a subscription agent bends only so far.
Subscribing is the right first move when your task is common (FAQ answering, email drafts) and your volume is modest. It is also the cheapest way to learn what agents can and cannot do for you.
Build: a custom agent ($5,000–$60,000)
When the task is specific to your business — your data, your exceptions, your systems — a custom agent is the honest option. Realistic phases and prices:
- Prototype: $5,000–$15,000, 2–4 weeks. Real data, one task, a measurable accuracy number, an honest go/no-go.
- Production build: $15,000–$60,000, 4–8 weeks. Integrations, guardrails, monitoring, fallbacks.
- Total for most small-business agents: $20,000–$80,000 end to end.
We break these phases down in detail in Custom AI Model Development: A Guide — the budgets there and here are the same, because it is the same work.
The hybrid middle ($5,000–$25,000)
The most common real-world path: a capable off-the-shelf model, wrapped in a thin custom layer — your prompts, your data connection, your output formats — without building a full platform. Most “agent projects” small businesses actually need live here.
Budgets by use case
Customer support agent
- Ready-made: $100–$500/month, or per-resolution pricing ($0.50–$2 per solved conversation)
- Custom (answers from your real help center, cites articles, escalates cleanly): $10,000–$30,000 build
- Watch for: per-seat pricing that scales with your team, and “resolutions” that are really just auto-replies
Lead follow-up and CRM hygiene
- Ready-made: $50–$300/month inside existing CRMs
- Custom (scores leads, drafts follow-ups, updates the CRM from replies): $5,000–$20,000 build
- Watch for: the agent that drafts beautiful emails nobody measures. If it does not tie to reply rates or booked calls, it is a demo.
Document and back-office processing
Invoices, applications, quotes, reports — the unglamorous work where agents pay back fastest.
- Custom: $10,000–$40,000 build, driven almost entirely by how many systems the agent must touch
- Watch for: integration cost. Two systems is a project; five messy systems is two projects.
The running costs buyers forget
- Usage fees: $50–$500/month in model/API costs for a typical small-business agent — noticeable at scale, trivial at pilot size
- Maintenance: 10–20% of the build cost per year. Agents drift as your data, prices, and policies change; models update underneath them
- Someone’s attention: an hour or two a week from a person who knows the process — checking exceptions the agent flagged
If you want the full picture of what AI-built systems cost to own over years — not just agents — we cover that separately in The Real Cost of AI-Built Software.
Build, buy, or subscribe?
- Subscribe when the task is common, your volume is low, and you want to learn. Budget: one month’s fee.
- Build custom when the task is specific, repeats daily, and the data trail exists. Budget: prototype first, always.
- Hybrid when you need your data and formats but not a whole platform. Most small businesses land here.
Check the payback before you buy
The math takes ten minutes:
- How many hours a week does the task eat? Say 10.
- What does an hour cost you, loaded? Say $40. That task costs $20,000+ a year — before rework, delays, and the work nobody gets to.
- Price the agent honestly: $8,000 hybrid build + $100/month running ≈ $9,200 year one.
- Payback lands around month six, then it prints.
If step 3 is bigger than step 2 — and sometimes it is — the honest answer is “don’t build yet.” A good vendor will tell you that before you ask. If you want those numbers for YOUR task rather than an example, our free AI automation assessment does exactly this arithmetic: nine questions, two minutes, and you get the task worth automating first, what it costs you a year, and what the build would run.
Start smaller than you think
The pattern that works is never “sign a big contract”: pick one task, prototype it in weeks for $5–$15K, measure on real data, and scale only what proves itself. Which task to pick first is its own question — we walk through it in AI Agents for Small Business: First Steps — and when the prototype says go, custom AI development is where the building happens.
Agents are not cheap. Done in the right order, they are cheap to find out about.