Agent or workflow? How to pick the cheaper automation for each task in 2026

Most automation decisions get made on sticker price, and that is the wrong number. The number that matters is cost-per-run, and it is set by the billing unit each platform uses: a task, a credit, or an execution. You see that unit before you compare plan tiers.

This guide gives you a decision rule, the arithmetic behind it, and a lock-in test. If a task’s steps are known and repeatable, run it as a deterministic workflow. Reserve AI agents for steps that are genuinely unpredictable, because a fixed sequence has a cost you can calculate in advance.

The decision rule in one line

Run known, repeatable steps as a deterministic workflow, and reserve AI agents for steps that genuinely change run to run. Ask two questions: can I write the steps down, and do they change from run to run? n8n’s pricing page states the cost logic directly: n8n says it lets you “build freely and only pay when a workflow runs from start to finish,” unlike tools that charge per step or user, according to n8n pricing.

If you can write the steps down, the task is deterministic, and a fixed sequence of actions reproduces it every time. If the steps change because the input is unstructured, the next action depends on a judgement call, or the path branches on content, that is the narrow slice of work where an agent earns its variable bill. Everything else belongs in a workflow.

This rule is about cost structure, not vendor preference. A workflow has a known step count, so its cost-per-run is knowable before you run it. An agent has an unknown step count, so its cost-per-run is a range. When the work is stable, that range is pure downside.

What you are billed for: tasks, credits, executions

Zapier bills tasks, Make bills credits, and n8n bills executions. The billing unit, not the sticker price, drives cost-per-run. Zapier’s pricing page states that AI steps, code, and SDK now follow the same task-based model. Make’s pricing page defines most actions as one credit.

Zapier has a Free plan at $0 per month with 100 tasks. Professional starts from $19.99 per month and adds multi-step Zaps, premium apps, and webhooks. Per-task rates are listed in a separate task-rate table. That table was not retrieved for this guide, so no per-task dollar figure appears below.

Make bills per credit. Its FAQ says most actions consume 1 credit, while router modules and error-handler modules (Rollback, Break, Resume, Commit, Ignore) consume none. Features that use Make’s AI Provider may consume more, and the operations help article documents consumption. The Free plan includes 1,000 credits per month. Extra credits come in bundles of 1,000 or 10,000 and expire after one month on a monthly subscription. Make’s startup program offers 50% off its Business plan to companies under 20 employees, and n8n runs a similar Start-up Plan.

n8n bills per execution. Starter is EUR 20 per month billed annually with 2,500 workflow executions and unlimited steps. Pro is EUR 50 per month billed annually with 10,000 executions, up to 50 concurrent executions, and 3 shared projects. Business is EUR 667 per month billed annually with 40,000 executions, and the pricing page lists it as self-hosted only. Agents, MCP, and AI nodes are included even on Starter. n8n’s FAQ says hosted plans store data in Frankfurt, Germany, per its security FAQ, and n8n Cloud docs cover the hosted option.

The table below compares the billing unit and exit path side by side. Prices are shown only where the fetched pages state them.

Criterion	Zapier	Make	n8n (hosted)	n8n (self-hosted)
Billing unit	Task (AI steps included)	Credit (most actions = 1; routers and error handlers = 0)	Execution (unlimited steps)	No execution billing on Community Edition
Free tier	$0, 100 tasks/mo	1,000 credits/mo	Trial only	Community Edition, free on GitHub
Entry paid price	From $19.99/mo (Professional)	Not retrieved for this guide	EUR 20/mo (Starter, 2.5K executions)	Server cost only; Hetzner from EUR 6.49/mo excl. VAT
Cost with an agent step	Billed as tasks; per-task rate depends on usage type	May use more credits via Make’s AI Provider	Agents and AI nodes included; execution count unchanged	Model and server costs depend on your setup
Self-host option	None listed on the pricing page	None listed on the pricing page	Not applicable	Yes, per installation docs

Currency note: Zapier prices in USD; Make, n8n, and Hetzner prices in EUR. No exchange-rate conversion is applied in this guide.

Where agents break the math

Agent runs are variable. An agent may loop, retry, or call tools an unpredictable number of times, and on Zapier and Make those calls consume tasks or credits. A deterministic workflow has a fixed step count, so its cost-per-run is knowable in advance. An agent buys flexibility at the price of an unpredictable bill.

An agent given a goal can take three tool calls or thirty, depending on what it finds. On a per-step or per-credit platform, that variability shows up directly on your invoice. On n8n’s per-execution model, the steps inside one run do not add to the execution count. AI model usage can still carry its own cost, so the bill does not move only when the model cost is zero or already budgeted.

This is why the hybrid pattern matters. You do not have to choose all-workflow or all-agent. The common shape is a deterministic workflow for the predictable steps, with one agent step for the part that cannot be scripted: classifying a message, drafting a reply, or extracting fields from a messy document. The workflow keeps the cost floor low, and the agent handles the exception.

Worked example: 10,000 runs a month

Scenario: 10,000 runs per month of a five-step deterministic job, such as new form submission, validate, enrich, update CRM, notify.

Assumptions, stated up front:

Zapier: each step consumes 1 task; no AI steps.
Make: each action consumes 1 credit; no router or error-handler modules; no AI Provider usage.
n8n: one execution equals one run from start to finish; steps are unlimited per execution.
Hetzner: one small instance is enough for this volume. That is an assumption, not a capacity claim.

Arithmetic, line by line:

Zapier: 10,000 runs x 5 steps = 50,000 tasks per month. The Free plan’s 100 tasks cover none of it. The paid price for 50,000 tasks was not retrieved for this guide, so no dollar figure appears.
Make: 10,000 runs x 5 actions = 50,000 credits per month. The Free plan’s 1,000 credits cover 2% of it. The paid price for 50,000 credits was not retrieved, so no dollar figure appears either.
n8n: 10,000 runs = 10,000 executions per month, which is exactly the Pro allowance at EUR 50 per month billed annually. Starter (2,500 executions) is too small, and Business (40,000) is oversized for this volume.
Self-hosted n8n: the Community Edition software is free. The Hetzner Regular Performance instance starts from EUR 6.49 per month excluding VAT, which is the server floor for the same 10,000 runs.

The takeaway: the same job consumes 50,000 billing units on Zapier and Make and 10,000 on n8n. Because Zapier and Make paid prices at this volume were not retrieved, this comparison is in billing units, not dollars. Adding an agent step raises Zapier and Make consumption; on n8n it does not change the execution count.

The lock-in axis: who owns the workflow

Cost is not only the monthly bill. The Zapier and Make pricing pages reviewed for this guide list no self-hosted edition. n8n’s Community Edition is free and can be self-hosted, and the installation docs confirm that without a license key n8n runs as the Community edition. The edition comparison lists what the free edition includes.

Lock-in is the risk that a vendor’s pricing change or plan cap forces your hand. If your automations run only on a managed platform, your exit path is rebuilding them elsewhere. If they run on a self-hosted platform, the workflow logic sits on infrastructure you control, and the software is free under the Community Edition. The trade is operational: you own updates, backups, and capacity planning.

The dollar figure for that insurance is the Hetzner floor of EUR 6.49 per month excluding VAT, compared with EUR 50 per month for n8n’s hosted Pro tier at the same 10,000-execution volume. Hetzner advertises a 99.9% uptime SLA for its cloud services, according to its cloud page. When lock-in risk outweighs managed-hosting convenience, self-hosting is the concrete alternative.

The flip test: when to move a step from workflow to agent

Use a three-question test for each step. Is the output deterministic? Does the input arrive in a fixed format? Would an error at this step cost more than the agent’s variable billing? If the answer to all three is yes, keep the step in the workflow. If any answer is no, the step is a candidate for an agent.

The pattern that falls out is a deterministic workflow wrapping one agent step. The workflow handles validation, enrichment, CRM updates, and notifications, all fixed and cheap per step. The agent handles the one step that genuinely cannot be scripted. On Zapier and Make, that agent call adds to the task or credit count. On n8n it stays inside the same execution.

FAQ

When is an AI agent actually cheaper than a workflow?

An agent is cheaper than a workflow when the alternative is a person doing the task by hand, or when building and maintaining the deterministic version costs more than the agent’s variable billing. For fixed, high-volume, repeatable steps, a deterministic workflow is usually the cheaper run, because its step count is known in advance.

Does self-hosting n8n really cost just EUR 6.49 a month?

EUR 6.49 per month excluding VAT is the Hetzner Regular Performance starting price for one small instance, according to the Hetzner cloud page. The Community Edition software is free. Whether one instance handles your volume, and how much operations time you spend, depends on your setup. This guide makes no performance or capacity claims.

Can I mix agents and workflows in the same automation?

Yes. A common pattern is a deterministic workflow with one agent step for the unpredictable part. On Zapier and Make, that agent call adds tasks or credits on top of the workflow’s steps. On n8n it stays inside the same execution, so the execution count does not change.

The bottom line

Run known, repeatable steps as a deterministic workflow, and reserve AI agents for steps that are genuinely unpredictable. Compute cost-per-run from the billing unit (task, credit, or execution), not from the sticker price. When lock-in risk outweighs managed-hosting convenience, self-host the Community Edition on infrastructure you control. Ask the same two questions for every new task: can I write the steps down, and do they change run to run?

How this guide was built

This guide is based on official pricing pages and vendor documentation; we did not run these platforms hands-on. Sources: Zapier pricing, Zapier task rates, Make pricing, Make operations help, Make startup program, n8n pricing, n8n Start-up Plan, n8n docs, n8n hosting installation, n8n GitHub, and Hetzner Cloud and its Regular Performance plan. Figures reflect pages fetched on 2026-10-20. Zapier’s per-task rate table and Make’s paid-tier prices at 50,000 units were not retrieved, so those figures are omitted rather than estimated.

HERO_IMAGE_PROMPT: Dark near-black background with a single glowing green accent line tracing a branching automation path from a fixed node into a looping node, abstract geometric shapes, no text, no logos

NiteAgent — AI agent development, frameworks, and production patterns

Cross-links automatically generated from NoCode Insider.