Running an AI agent costs the same platform plan you may already pay for — Make from $9 a month, n8n Cloud from €20, Zapier Professional from $19.99 — plus a usage bill you cannot predict in advance. That unpredictability, not the headline price, is the thing to plan around.
The reason is structural. A normal automation runs a fixed number of steps, so you can multiply and budget. An agent decides at run time how many steps its goal needs, which turns a known monthly cost into a range. This guide shows how each platform meters that, what a realistic month looks like, and how to cap it before you get a surprise.
Key takeaways
- Entry cost is your existing plan. Agent features ship on standard paid tiers on all three major platforms, not as enterprise add-ons.
- The variable is step count, and the agent chooses it. Budget a range, never a figure.
- Per-workflow billing suits agents best. n8n counts one execution however many steps an agent takes; Zapier bills every step.
- Set a step cap before the first live run. An agent looping on a misunderstood goal is the main way people overspend here.
- Free tiers cannot run a production agent. Zapier’s 100 tasks and Make’s 1,000 credits go in days.
What Does an AI Agent Cost to Run?
The plan price is the floor and it’s modest. Every major automation platform includes agent capability on its standard paid tiers, so if you already automate, your marginal cost to start is zero.
| Platform | Entry paid plan | Included usage | Billing unit |
|---|---|---|---|
| Make | Core $9/mo | 10,000 credits | One module run |
| Zapier | Professional $19.99/mo annual | 750 tasks | One action step |
| n8n Cloud | Starter €20/mo annual | 2,500 executions | One whole workflow run |
| n8n self-hosted | Free software | Unlimited | — |
The self-hosted row deserves a caveat: free software still runs on a paid server, and one to two hours a month of maintenance costs more than the €20 plan at any professional hourly rate. The full arithmetic is in the n8n pricing breakdown.
Why Can’t You Predict an Agent’s Cost?
Because the agent decides how much work to do. It takes a goal, plans an approach, calls a tool, looks at the result, and decides whether to continue — and that loop runs as many times as the model judges necessary. Nothing in your configuration fixes the count in advance.
A worked example makes the spread concrete. Take one agent researching an inbound lead, run 200 times a month:
| Behaviour | Steps per run | Make credits | Zapier tasks | n8n executions |
|---|---|---|---|---|
| Efficient — finds it fast | 4 | 800 | 800 | 200 |
| Typical | 9 | 1,800 | 1,800 | 200 |
| Thorough — checks everything | 18 | 3,600 | 3,600 | 200 |
| Looping on a bad goal | 40+ | 8,000+ | 8,000+ | 200 |
Read the last column. On n8n the cost is identical in all four rows — 200 executions, comfortably inside the 2,500 Starter allowance — because n8n’s documentation is explicit that “it doesn’t matter how many steps are in the workflow or how much data it processes, it’s still a single execution.”
On Make and Zapier the same agent ranges from 800 to over 8,000 units. Zapier’s 750-task Professional tier can’t even cover the efficient row, and the looping row would push you several tiers up. This is the single most important cost fact about agents, and it is the one no pricing page states.
The rule: if you plan to run agents seriously, per-workflow billing removes your biggest cost risk. If you’re running one research agent a few times a day, it won’t matter.
How Do You Cap the Cost?
Four controls, all set before the first live run. None is optional, and the first one is what actually prevents a surprise bill.
| Control | What it stops | Set it to |
|---|---|---|
| Maximum steps per run | Loop-until-bankrupt | 2–3× your typical run, no more |
| Maximum runs per day | A trigger misfiring in a loop | Slightly above real volume |
| Usage alert on the platform | Silent drift | 50% and 80% of allowance |
| Tool allow-list | Expensive tools being called unnecessarily | Only the tools the goal needs |
The tool allow-list is underrated as a cost control. An agent granted access to a paid enrichment API will use it whenever it thinks that would help — which may be every run. Grant the minimum set of tools the goal genuinely requires, and add more only when you’ve watched it work without them.
What Does a Realistic First Month Look Like?
Cheaper than most people expect, because a sensible first agent runs on low volume against a task where being wrong is cheap. Here’s a first month for a single research agent on a small business’s lead flow.
| Line | Amount | Note |
|---|---|---|
| Platform plan | $9–20 | Probably already paying it |
| Agent usage, 120 leads × ~9 steps | ~1,080 units | Inside Make Core; over Zapier’s 750 tier |
| Shadow-mode double running | +0 | Shadow mode is the same runs, just unused output |
| Your review time, 2 weeks | 2–4 hrs | The real first-month cost |
| Build and configuration | 1–2 hrs | One-off |
That fourth row is the honest answer to “what does an agent cost.” Three to six hours of your time in the first fortnight, reading output to find out how often it’s wrong in ways you wouldn’t have noticed. At a $50 hourly value that’s $150–300 — more than a year of the platform plan.
What Are the Costs Nobody Quotes You?
Three of them, and together they usually exceed the platform bill in the first quarter. None appears on a pricing page because none of them is charged by the platform.
| Hidden cost | When it hits | Rough scale |
|---|---|---|
| Verification time | Every week, forever | 15–60 min/week |
| Third-party tool calls | Whenever the agent chooses to use one | Whatever that API charges |
| Cost of a confident wrong answer | Rarely, unpredictably | Unbounded |
Verification time never fully goes away. An agent that produces work you act on needs spot-checking indefinitely — you can drop from reading everything to sampling 10%, but not to zero, because an agent that reasons badly reports success exactly like one that reasoned well.
Third-party tool calls are billed by whoever owns the tool. If your agent can call a paid enrichment or search API, it will when it judges that useful — and that spend appears on a different invoice entirely, which is how it goes unnoticed.
The cost of being confidently wrong has no formula. It’s the reason to point your first agent at research and drafting rather than at anything that moves money or sends an external message. Keep a human on the outbound edge and this cost stays near zero.
Is an Agent Cheaper Than Just Building an Automation?
Almost never, for the same task. A rule-based automation runs a fixed number of steps with no reasoning overhead, so if the task can be specified, automation wins on cost, speed and predictability every time.
| Automation | AI agent | |
|---|---|---|
| Steps per run | Fixed, known | Variable, decided at run time |
| Monthly cost | Predictable | A range |
| Speed | Seconds | Seconds to minutes |
| Cost of complexity | Linear in steps | Unpredictable |
| Handles unforeseen cases | No | Often — the reason to pay more |
So the cost question and the architecture question are the same question. If you can write the logic as a flowchart on one page, build an automation and spend nothing extra. If the right next step genuinely depends on what the last step found, an agent earns its variable bill. The full decision framework is in AI agents for business.
Frequently Asked Questions
How much does an AI agent cost per month?
The platform plan — from $9 on Make Core, $19.99 on Zapier Professional, €20 on n8n Cloud Starter — plus variable usage that depends on how many steps the agent takes. Budget a range, and set a step cap before the first live run.
Why is my AI agent bill unpredictable?
Because the agent decides how many steps its goal needs, and that varies per run. On per-step billing like Zapier’s, a run that takes 18 steps instead of 4 costs four times as much for the same task.
Which platform is cheapest for AI agents?
n8n, because it bills one execution per whole workflow regardless of step count — so an agent taking 40 steps costs the same as one taking four. On Make and Zapier the same agent can range from 800 to 8,000 billable units a month.
Can I run an AI agent on a free plan?
Not in production. Zapier’s free tier is 100 tasks a month and Make’s is 1,000 credits — an agent doing real work consumes those within days. Free tiers are fine for testing the setup, not for running it.
How do I stop an AI agent running up a bill?
Set a maximum step count per run at two to three times your typical run, cap runs per day, add usage alerts at 50% and 80% of your allowance, and grant only the tools the goal actually needs. Do all four before going live.
Is an AI agent cheaper than hiring someone?
For bounded, recoverable tasks, yes — a platform plan is $9–20 a month against thousands for a hire. But agents handle narrow tasks with no accountability, so the comparison only holds for the specific work they can do, not for a role.
The Bottom Line
Agent pricing is not expensive, it’s uncertain — and uncertainty is the thing to engineer around. The plan you already pay for covers the entry cost; what varies is how many steps the agent decides to take, and on per-step billing that variance lands directly on your invoice.
So do three things before the first live run: pick a platform whose billing unit suits variable step counts, set a hard step cap, and add usage alerts. Then accept that your real first-month cost is three to six hours of reading output to find out how often it’s quietly wrong.
For platform-by-platform detail, see the breakdowns of Make’s credit math, Zapier’s per-step tasks and n8n’s per-execution model.
Want one practical automation you can set up in 15 minutes, delivered twice a month? Join the free AI automation newsletter.
Sources
- n8n — Pricing plans and execution definition
- Make — Pricing plans
- Make Help Centre — Operations and how they are counted
- Zapier — Pricing plans
- Zapier — Agents product overview
- Amazon Web Services — Agents vs Automation: A Strategic Guide
All sources retrieved 26 July 2026. Usage tables are worked arithmetic from each platform’s documented billing unit, not vendor benchmarks. Agent features and pricing change frequently.
