Zapier vs Make vs n8n vs custom integration: an operator's guide

Zapier has the largest app library and the shallowest learning curve. If the two systems you want to connect are both popular SaaS products with a Zapier connector, and the workflow is "when this happens, do that," Zapier will be running in an hour and will keep running.

A real example. A new lead lands in a HubSpot form; Zapier creates the contact in QuickBooks as a customer and posts a message in a Slack channel. Three steps, low volume, no judgment. Zapier is the correct tool and paying for anything more is waste.

Where it turns. Per-task pricing punishes multi-step flows at volume. A 10-step Zap that runs 2,000 times a month is 20,000 tasks, and you are now in a plan tier that costs more per month than the hosting for a custom integration. And when a Zap fails silently on step seven, you find out from the customer.

Make (formerly Integromat) gives you a visual canvas with branching, routers, iterators, and error handlers, and it prices per operation at a lower rate than Zapier. It is the tool for the person who has outgrown two-step Zaps but does not want to write code.

A real example. An estimate is approved in Jobber; Make pulls the line items, builds an invoice in QuickBooks, attaches the CompanyCam photos, emails the customer, and logs the result to a Google Sheet, with an error path that posts to Slack if the QuickBooks call fails. Make handles this well and the credits are affordable at contractor volume.

Where it turns. Credits are consumed per module operation, so loops and lookups over large lists get expensive and hard to predict. The 40-minute execution cap bites on long-running jobs. And the scenario becomes the one thing in the company only one person understands.

n8n is the developer's choice. It is priced per execution rather than per step, so a complex workflow costs the same as a simple one. It can be self-hosted for free under n8n's license, which means unlimited runs on your own server and your data never leaves your infrastructure. It has proper execution logs, error workflows, version control on the Business plan, and it lets you drop into JavaScript when a node does not exist.

Cloud or self-hosted. Cloud is fine for a small team that does not want to run a server. Self-hosted is where n8n shines: predictable cost at volume and control over data. The trade is that someone has to patch, back up, and upgrade the server, and if that someone leaves, so does the knowledge.

A real example. A trucking company pulls a nightly export from its TMS, enriches each load with telematics data through an API, writes the result to a Postgres database, and generates a lane-profitability report. n8n on a $20 server does this well, and the developer who set it up can read the logs when it fails.

Where it turns. It is a developer tool. An office manager will not maintain it. And when a vendor has no API at all, or the job needs judgment rather than rules, n8n needs custom code inside it, at which point you are building custom software with extra steps.

Brittle multi-step flows. Every additional step is another place to fail, and no-code tools make it easy to build twenty-step flows that nobody can debug.

Rate limits and quotas. Your TMS or accounting system allows a certain number of API calls per minute. No-code tools handle this poorly, and the failure mode is silent: some records go through and some do not.

No API. A large share of the systems operations companies depend on, port portals, older TMS installs, vendor extranets, have no connector and no API. No-code tools stop there.

Sensitive data. If the workflow touches health data, payroll, or customer financials, you need to know where it is processed and stored, and you need an audit trail. Cloud automation tools can meet this on enterprise plans; most people are not on enterprise plans.

Version control and testing. Changing a live Zap or scenario changes production. There is no staging, no diff, no rollback in the way developers mean it, so the person editing it is one typo from breaking billing. n8n's Business plan adds Git-based versioning; custom code has it by default.

Judgment. Rules handle the predictable 90 percent. The remaining 10 percent, the invoice that does not match the PO, the port date that conflicts with the line's, needs a model and a person. No-code tools are adding AI steps, but governance of those steps is thin.

Who fixes it at 6 a.m. This is the one that matters. When the workflow that feeds dispatch fails, who is accountable, what is their response time, and is it in writing?

Custom integration pays back when the work crosses your core systems, when a system has no connector, when the data is sensitive, when volume makes per-task pricing absurd, when you need a person in the loop on exceptions, or when you want to own what you run.

What custom looks like in our practice: an integration or agent built in your cloud account, in a repository you own, with least-privilege access to each system, logging on every read and write, a shadow period before it replaces the manual process, and approval steps on anything that moves money. AI is used where the data is unstructured, extracting fields from an invoice, deciding which of two dates to trust, and plain code everywhere else.

What it costs. Build workstreams are fixed-fee, $10,000 to $70,000 depending on the systems and the data. Hosting is usually tens of dollars a month. A managed plan, if you want someone on call with a response target, runs $600 to $2,200 a month. We start with a two-day assessment, from $10,000, credited against the build, that ranks every candidate workflow by P&L impact and risk so the first thing built is the thing that pays. Prices verified September 2026. Custom software and integrations → · AI agents →

What it does not cost. A platform fee to us. Everything transfers to you, and there is nothing to migrate off.

Do not rip them out. Inventory them first: what each one does, who built it, what account it lives in, how often it runs, and what breaks if it stops. Most companies find a third are dead, a third are fine where they are, and a third are load-bearing and fragile. The fragile third is the candidate list for n8n or custom, ranked by what it costs when it fails. Rebuild those, run old and new in parallel until the outputs match, then switch. It is a common first workstream and it usually pays for itself in the first year of avoided plan fees and avoided failures.

The workflow: a lead comes in by web form or phone. It becomes a customer record in QuickBooks, a job in the field-service system, and a slot on the dispatch board, and the customer gets a confirmation.

In Zapier. Form → create QuickBooks customer → create Jobber client → create Jobber job → send SMS. Five tasks per lead. Works. At 300 leads a month that is 1,500 tasks, comfortably in a Professional plan. Phone leads are not captured, because the phone system has no connector; someone types them in.

In Make. The same flow with a router: if the customer exists, update instead of create; if the requested slot is taken, pick the next; on any error, post to Slack. Around eight operations per lead, well within a Core plan. Still no phone capture.

In n8n. Same flow, one execution per lead regardless of steps, self-hosted for the cost of the server. A developer adds a node that queries the phone system's call log API, so phone leads are captured too. Someone has to maintain the server and the code.

Custom. The phone is answered by an AI receptionist that qualifies the caller and books the job directly, with a transfer to a person for emergencies. Web and phone leads flow through the same code into QuickBooks, the FSM, and the dispatch board. Duplicates are matched on phone number and address. Every action is logged; the office manager sees a daily digest of what was booked and a queue of the three calls the agent was unsure about. Fixed fee to build, managed plan to run. The difference is not the plumbing, which all four can do; it is the phone, the judgment, and the accountability.