Decision guide

n8n vs Make vs a custom agent: which should your first AI agent be built on?

n8n and Make are workflow builders with an AI-agent node bolted on; a custom agent is code that calls a model with your tools, your rulebook and your tests. Use Make when a non-technical person will own a simple flow and the AI step is a summary or a classification. Use n8n when someone technical will own it, you want to self-host, and the agent is internal or low-stakes. Build a custom agent when it talks to customers, writes to a system of record, or has to pass a test set before every change, because the approval step, the audit log and the sixty-message test are the product, and the builders make those hard. Most good setups use both: the builder for the plumbing and a custom agent for the part that needs reading and deciding.

All three can run something that looks like an agent in a demo. They differ on who can change it, what happens when a message does not fit, what you can prove about it afterwards, and what it costs at your volume. This page puts them side by side on those, says which fits which business, and shows the arithmetic for a month of a clinic’s bookings on each.

Side by side

Maken8nCustom agent
What it isHosted visual workflow builder, per-operation or credit billingVisual workflow builder, hosted or self-hosted, per-execution billingCode calling a model API with your tools, rulebook and tests, on your own hosting
Who changes itAn operations person, no codeA technical person; some JavaScript helpsA developer; the rulebook and facts list are edited by the business in a sheet
AI agent supportAn AI agent builder on the same canvas; good for a summary or a classification stepAn agent node with tools, memory and model choice; capable, with careWhatever the model API supports: strict tool schemas, structured output, caching
Tool permissionsEach module does one thing; "may do alone / with approval" is yours to wireTools are workflows or nodes; permissions are what you wireThe tool list is the permission list, by construction
Approval stepPossible with a human-in-the-loop module or an email; clumsy at volumePossible with a wait node and a form; workableBuilt in: draft, approve, act, with the queue in your console
Audit logExecution history per scenario, retained by planExecution data per run, retained by setting; self-hosted keeps everythingEvery message, tool call and decision in your own table, forever
Test set before a changeManual re-runsManual or scripted re-runsThe sixty-message set runs in CI before every change
Multi-turn WhatsApp conversationAwkward: state lives in a data store moduleWorkable with a database nodeNative: the conversation is the unit
When it breaksA renamed field, an unexpected branch, an operation capThe same, plus the server if you self-hostRules nobody wrote; data it cannot reach
Cost modelPer operation or credit, monthly plansPer execution on cloud; free to self-host, you pay for the serverModel usage plus hosting; the build is a fixed fee
Cost at 2,000 agent conversations a monthA mid plan, likely tens of dollars, plus model usage; worked belowA starter cloud plan or a small server, plus model usage; worked belowLow hundreds of dollars at most, usage and hosting; worked below
Data locationMake’s cloud regionsYour server if self-hostedYour hosting and your model vendor’s region
Lock-inScenarios do not export to anything elseWorkflows export as JSON; self-hosting is portablePlain code and schemas; the model is swappable
Plan prices are described rather than quoted because both vendors change them more than once a year; the vendors’ pricing pages are the only source to budget from.

Which to pick, and when

Pick Make

When an operations person will own the flow, the AI step is a summary or a sort, and nothing talks to a customer unsupervised. It is the fastest way from a form to a CRM to a Slack message, and the visual canvas is the right tool for that.

  • No developer in the business, and none wanted
  • The flow is form-to-CRM-to-message plumbing with one AI step
  • Volume is a few thousand operations a month

Pick n8n

When someone technical will own it, you want the data on your own server, and the agent is internal: a morning summary, a document check, a draft for a person to send. It is also the right place to prototype an agent before deciding whether it deserves code.

  • Someone in the business is comfortable with a server and some JavaScript
  • Self-hosting matters, for cost or for data
  • The agent drafts for a person rather than acting on a customer

Build a custom agent

When the agent is customer-facing, writes to a system of record, or has to pass a test set before every change. The approval queue, the audit log and the test run are easy in code and awkward on a canvas, and they are what makes a clinic or a CA firm safe to run one.

  • It books, qualifies, chases or reconciles with real customers
  • A regulator, a partner or a doctor will ask to see the log
  • The rules are long and the lines it must never cross are the product

What the AI-agent node does, and does not, give you

Both builders now have an agent node: give it a model, a prompt and some tools, and it loops until it is done. For an internal task that is enough, and it is a genuinely good way to find out whether a workflow is agent-shaped before spending on a build. The reference builds on this site were all prototyped that way.

What the node does not give you is the scaffolding that makes an agent safe on real customers. An approval queue someone works from a screen. A log you can filter by patient. A test set of sixty awkward messages that runs before every rule change and fails loudly. Strict tool schemas so a booking call cannot arrive without a date. A rulebook the practice manager edits in a sheet rather than in a prompt box only the builder can see. Each of those can be wired on a canvas; together they are more work than the code, and they break in ways a canvas does not show.

What a month of bookings costs on each

Take the clinic reference build: about 2,000 booking conversations a month, four turns each. On a builder each turn is a handful of operations (receive, load rules, call the model, call a tool, reply, log), call it six, so a conversation is roughly 25 operations and a month is roughly 50,000. On a custom agent the same month is model usage plus hosting. The build fee for the custom option is a one-off and is shown separately so the comparison is honest.

OptionWhat the month consumesMonthly running costOne-off build
MakeAbout 50,000 operations or the credit equivalentA mid-tier plan: described on Make’s pricing page; tens of dollars at this volume, more if AI steps cost extra creditsYour time, or a freelancer’s day rate
n8n cloudAbout 2,000 executions (one per conversation)A starter plan covers it: described on n8n’s pricing pageYour time, or a freelancer’s day rate
n8n self-hostedThe same, on your own serverA small VPS, typically under $20, plus someone’s hours for updates and backupsYour time, plus the server setup
Custom agentAbout 28M input and 4M output tokens before caching, plus a database and a host$120–$600 a month: the estimator’s mid-volume band, model usage before caching plus hosting$4,500 for the pilot; $8,000–$42,000 for one workflow in production
Builder plan prices are deliberately not quoted; both vendors have changed their models (operations, credits, executions) in the last two years. The custom-agent figures are the pricing page’s bands. Model usage is the same on a builder: the agent node calls the same API at the same rates, on top of the plan.

The pattern that usually wins: builder for plumbing, code for the brain

The setups we are happiest with are not one of the three. Make or n8n watches the form, the inbox and the calendar and moves records between them, which is what they are superb at. The agent is a small service behind a webhook that reads the message, applies the rulebook, calls its five tools and writes the log, and the builder calls it like any other node. The business keeps a canvas it can see; the part that needs tests and an approval queue lives in code; and if the builder is ever replaced, the agent does not notice.

It also keeps the agent small, which is what keeps it auditable. An agent asked to do everything on one canvas is the thing nobody can explain to the doctor who asks what it did last Tuesday.

When you need none of them

  • The process is not agreed yet. A canvas makes a disagreement faster, not smaller.
  • It happens a few times a week. A checklist and a person are cheaper and better.
  • The data is not reachable. Fix the export or the API first; no builder can read what it cannot see.
  • The decision belongs to a professional by law. Build tools that prepare the file, not agents that decide it.

Decide in 5 questions

Answer them in order. Most people are done by the third.

  1. 1

    Will the agent talk to customers or write to a system of record without a person in between?

    Yes: Custom agent, with the builder for plumbing if you have one. Continue to check the rest.

    No: A builder is fine. Continue.

  2. 2

    Is there someone technical who will own a server and a workflow?

    Yes: n8n, self-hosted if data location matters.

    No: Make, hosted.

  3. 3

    Will someone ask to see the log, by customer, months later?

    Yes: Custom agent, or n8n self-hosted with a database node and retention set to keep everything.

    No: Any of the three.

  4. 4

    Do the rules change often, and should a non-developer change them?

    Yes: Custom agent with the rulebook in a sheet, or a builder with the rules in a sheet the node reads. Not a prompt box.

    No: Continue.

  5. 5

    Is the volume above a few thousand conversations a month?

    Yes: Run the cost table with your numbers; operation-based plans climb fastest.

    No: Cost is tens of dollars on all three. Decide on ownership and auditability instead.

FAQ

Common questions

Can we start on n8n and move to a custom agent later?

Yes, and it is a good path. Prototype the agent on n8n with the rulebook in a sheet, learn the edge cases from a few weeks of real messages, then move the brain into code with the test set those weeks produced, keeping n8n for the plumbing around it. The pilot can start from exactly that point.

Is a custom agent more expensive to run?

Not to run. Model usage is the same on a builder, because the agent node calls the same API; a custom agent adds a database and a host that cost tens of dollars at small-business volume, and skips the plan fee. The difference is the build, which is a fixed fee against a brief, and the maintenance, which is a person either way.

Which is safest with customer data?

The one where you can say, in two lines, where the data goes. Self-hosted n8n and a custom agent on your own hosting make that easy. Make and n8n cloud put it in the vendor’s region under the vendor’s agreement, which is fine for many businesses and not for some. The model vendor’s region applies to all three.

Do you build on n8n or Make?

We use both for plumbing in client setups and recommend them by name; we build the agent itself in code because the approval queue, the audit log and the test run are what we are being paid to make safe. If a workflow does not need those, we will say so and point you at the builder.

What about Zapier?

Zapier belongs in the same column as Make: hosted, visual, task-based pricing, an AI step available, the easiest for a non-technical owner. The agent-vs-chatbot-vs-Zapier page covers the broader question of whether you need an automation or an agent at all.

Is n8n better than Make?

For a technical owner who wants to self-host and build an internal agent, n8n; for a non-technical owner who wants hosted plumbing with one AI step, Make. n8n’s per-execution billing is cheaper on multi-step workflows and its workflows export as JSON; Make’s canvas is easier to pick up and its per-operation or credit billing climbs faster at volume. Both prices change often; read the vendors’ pages on the day.

Is n8n free?

The self-hosted community edition is free to run on your own server; you pay for the server and for someone’s hours to keep it updated and backed up. n8n cloud is a monthly subscription billed per execution, described on its pricing page. Model usage for an agent node is billed by the model vendor on top in either case.

Can n8n build AI agents?

Yes. Its agent node takes a model, a prompt and tools and loops until done, which is enough for internal agents and a good way to prototype a customer-facing one. What it does not give you out of the box is the approval queue, the per-customer audit log and the test set that runs before every change, which is why customer-facing agents usually end up with the builder doing the plumbing and the brain in code.