Connector counts decide nothing. The billing unit decides everything — tasks, API requests, or executions — because the billing unit is also the failure unit, and each one breaks at a different, predictable point.
The invoice that started this comparison was for one month of a Zap that a junior dev could rewrite in forty lines of Python. Six action steps: webhook in, enrichment lookup, formatter, two API calls, Postgres write, Slack post. It fired 9,400 times during a lead-gen push and billed 56,400 tasks. Nothing broke. The Zap did exactly what it was built to do, and so did the meter. The client had modeled the run count and never multiplied it by the step count.
All three tools do HTTP, branching, cron, retries, and error handlers. All three will take a webhook, write a row to Postgres, and post to Slack, and the demo takes about twenty minutes in each. What actually separates them is the billing unit, because the billing unit is also the failure unit. Zapier meters successful action steps. Power Automate meters API requests against a rolling 24-hour budget per licensed seat. n8n meters whole workflow executions on cloud, and your server bill if you host it. Each has a volume where it stops being cheap and becomes a migration.
A task is one action that runs successfully. Filters and paths that halt early cost nothing, which reads as generous until you count nodes in a workflow that survived contact with a client. Six actions across 10,000 records is 60,000 tasks for one campaign. Worse, the Looping by Zapier app charges a task per iteration per step, so a loop over five line items inside a four-step flow is twenty tasks per order, not four. The plan tiers look enormous next to a two-step demo and evaporate next to a nine-step production flow with a loop in it.
The second ceiling is compute. Code by Zapier caps around ten seconds and a few hundred megabytes on paid plans, so a real HTML parse, a batch transform, or an LLM call with two retries either gets split across steps — multiplying tasks again — or gets pushed into a Lambda you now own, monitor, and deploy. Zapier is then a scheduler wrapped around your code, billed per invocation. It is still the right call for a three-step flow firing a few thousand times a month that someone in marketing needs to edit on a Tuesday without opening a ticket. That category is real and we still recommend it.
Power Automate counts API requests, not tasks — roughly 40,000 per 24 hours on a per-user premium plan, and about 6,000 on the seeded Office 365 entitlement most teams start with. Every action, every loop iteration, and every connector poll draws down that budget. An apply-to-each over 8,000 SharePoint records with four actions inside is 32,000 requests, which is why nightly syncs start throttling at 3am and finish at 9am with no error anyone can point at. Connectors throttle independently on top of the tenant budget, so SharePoint can return 429s while your request allowance looks fine. The quietest failure is Get items defaulting to 100 rows: pagination off, no error, the flow reports success and processed an eighth of the queue.
Then licensing. The HTTP connector — the one you need the instant you talk to something outside Microsoft — is premium, as are SQL, Dataverse, and nearly every third-party connector worth wiring up. A free Office 365 automation quietly becomes a per-seat premium license for everyone who triggers it. That is not a reason to avoid it. If the data lives in SharePoint and Dataverse, if DLP policies are enforced at the tenant boundary, and if audit trails have to land in Purview without a side integration, Power Automate wins on governance before anyone builds a flow. Enterprises pick it for that, and they are right to.
n8n cloud counts full executions. A forty-node workflow with three loops and an LLM agent costs the same as a two-node one, which inverts the economics of anything branching or agentic. Self-hosted, marginal cost per execution is your server bill. This is our default for client work with real volume — and the savings hold only if you accept what you just signed up to operate.
n8n passes items between nodes in memory. A Postgres node selecting 400,000 rows will try to hold every one, and the container dies with a JavaScript heap out of memory that surfaces in the UI as an execution that simply never finishes. The fixes are unglamorous and non-optional: LIMIT and OFFSET on every query, Split In Batches with a bounded size, N8N_DEFAULT_BINARY_DATA_MODE set to filesystem or S3 so PDFs never become node data, and no unbounded array reaching an LLM node. Execution history writes every run to Postgres; at 12,000 runs a day the execution_entity table will eat a 50 GB volume in weeks unless you set a prune max-age. Past a few executions per second, move to queue mode with Redis and separate workers — the single-process default blocks on any long-running node, including the ones waiting on an API.
Nobody demos this and it decides more migrations than pricing does. An n8n workflow is a JSON document with stable node IDs. Export it, commit it, open a pull request, and read exactly which node someone edited at 11pm before the sync started dropping records — credentials stay as references, so nothing secret lands in the repo. Power Automate has a genuine answer in solutions and ALM pipelines across dev, test, and production environments, but the setup cost is high enough that most teams skip it and edit production flows live. Zapier gives you in-product version history and no meaningful export, so your rollback plan is a screenshot and whoever built it.
The rule we use: short, stable, owned by marketing — Zapier, and stop thinking about it. Data never leaves the Microsoft tenant and compliance owns the roadmap — Power Automate, with the premium seats budgeted honestly on day one. Branching logic, custom code, an LLM in the loop, or volume that makes per-step billing absurd — n8n, staffed like the infrastructure it is. The expensive mistake is not choosing wrong at the start. It is choosing Zapier for five steps, letting it reach fourteen over a year, and rebuilding it under deadline because the invoice finally got a director's attention.
A software studio that ships and maintains its own products — KeepChats, Gwora and MoveProof — and builds the same way for clients. Founded and led by codewithumar.
Talk to the studio →Send the problem. You get one fixed number and a plan back within a business day.
We build software worth keeping — for clients, and for ourselves.
Founded & led by codewithumar