"How much does an AI agent cost?" has three honest answers, because you can be buying three different things. A per-seat copilot is about $20–$30 a user per month. A do-it-yourself agent on an API costs a few dollars a month in tokens. A custom enterprise build runs tens of thousands and up. Pick the row you're actually in.
Why the question has no single number
Most "AI agent pricing" articles quote one scary range — six figures to build, thousands a month to run — because they only cost the third row: a custom, staffed enterprise system. That's real, but it's not what most people mean when they wire up an agent to read their email. The desk-level version of the same job costs less than a lunch. The honest answer is that price tracks which of three things you're paying for, so here are all three, with real numbers.
Row 1 — Buy a seat (the copilot path)
The simplest agent is one someone else built and rents you by the seat. Microsoft's Microsoft 365 Copilot pricing lists it at $30 per user per month (annual) for enterprise, with a lower-priced Copilot tier for smaller businesses. That buys in-app assistance across Word, Excel, Outlook, and Teams, grounded in your own tenant data. On our own desk, Issue #001's Gmail agent runs on nothing more than Claude Pro at $20/month — the Gmail connector is included, there's no API key and no hosting bill. Per-seat pricing is predictable and needs zero engineering; the trade-off is you get the vendor's agent, not yours.
Row 2 — Pay per token (the build-it-yourself path)
If you build the agent yourself against a model API, the meter is tokens — text in and text out — and it's cheaper than most people expect. Anthropic's pricing is per million tokens: Claude Haiku at $1 in / $5 out, Sonnet at $3 / $15, and Opus at $5 / $25. Two levers cut that further, both documented in Anthropic's docs: batch processing is 50% cheaper, and prompt caching cuts the cost of re-sent context by up to 90%. This is why a personal agent is nearly free to run: the same Gmail triage job, built on the API path instead of Claude Pro, cost about $3/month in API usage in Issue #001's own side-by-side. The catch isn't the tokens — it's that you now own the code, the prompt tuning, and the debugging.
Row 3 — Rent a platform (the no-code middle)
Between "buy a seat" and "write code" sits the automation-platform tier, and its cost model is the thing to understand before you commit. Platforms meter differently: Zapier bills per task (each action an automation takes), while n8n bills per workflow execution (one run, however many steps) or runs free on a server you host. We walk the full trade-off in n8n vs Zapier for AI agents, but the cost headline is this: a per-task meter gets expensive fast as an agent takes many small actions, while a per-execution or self-hosted model stays flat. Issue #001's table put a self-hosted n8n build at roughly $5/month in server cost versus $20–$50+ for the same job on a hosted per-task platform.
Row 4 — Commission a custom build (the enterprise path)
This is the row the listicles fixate on, and the ranges are genuinely wide because "custom agent" covers everything from a scoped internal tool to a multi-agent system wired into legacy software. Treat any single quoted figure with suspicion — vendor estimates vary by an order of magnitude — and get your own scoped bid. The one durable point: a custom build carries ongoing cost, not just a one-time price. Retool's argument for hourly, outcome-based pricing of AI agents is worth reading precisely because it separates the build cost from the run cost that keeps arriving every month.
The cost everyone forgets
Whatever row you're in, the sticker price is not the total. The cost people leave out is human review time — the minutes a person spends checking what the agent did before it counts. That's a real recurring line item, and it's the reason we argue for measuring ROI against a baseline instead of a launch-day number: subscription plus tokens plus setup plus review is the true running cost. An agent that's cheap to run but needs ten minutes of checking per task may cost more than a pricier one you can trust to run unattended. Count all of it before you decide the agent is worth it.
FAQ
How much does a basic AI agent cost per month? For a single-person desk agent, roughly $3 to $30 a month. On an API you pay only for tokens — Issue #001's Gmail triage ran about $3/month that way, or $20/month bundled inside Claude Pro. A per-seat copilot like Microsoft 365 Copilot is $30/user/month.
Why do some sources say AI agents cost six figures? Those figures are for custom, staffed enterprise builds — a bespoke multi-agent system integrated into existing software, with ongoing engineering. That's a real cost, but it's a different product from a desk agent you set up in an afternoon. The question only has one answer once you fix which of the three tiers you mean.
What does it cost to run an agent on the Claude or GPT API? You're billed per million tokens. Anthropic's pricing is $1/$5 for Haiku, $3/$15 for Sonnet, and $5/$25 for Opus (input/output). Batch and prompt caching cut that by 50% and up to 90% respectively, which is why a light personal agent often costs only a few dollars a month.
Is a no-code platform cheaper than building my own? It depends on volume and how many actions your agent takes. Per-task platforms like Zapier are simplest to start but climb as the agent gets busy; a self-hosted tool like n8n stays flat but you maintain it. The n8n vs Zapier breakdown walks the crossover point.
What's the hidden cost of an AI agent? Human review time. Every gated action a person has to approve or double-check is recurring labor that rarely shows up in the pricing page. Count it as part of the running cost — it's central to measuring an agent's real ROI.
We document real agents professionals actually run — with the real costs, not demo-day math. Subscribe free for a new build each week.