It is four layers, not one market
People argue about automation tools as if it were one market with one winner. It isn't. It is four layers stacked on each other, and a small business touches all four whether it wants to or not.
- The connector layer. Zapier, Make, n8n, and a long tail of niche tools. This is the plumbing: something happens, something else follows.
- The model layer. OpenAI, Anthropic, Google, Mistral, plus local models. Text goes in, judgement comes out.
- The protocol layer. The Model Context Protocol, which is how the first two layers now talk without bespoke glue. We wrote about what MCP changed separately.
- The rules layer. The EU AI Act's transparency chapter has been applicable since August 2026, and some of it lands on you, not on the model vendor.
Most buying mistakes happen because somebody compares two products from different layers.
The connector layer: one job, three billing units
The three big connectors do roughly the same thing. What actually separates them is what they charge you for — and that decides your bill far more than the sticker price does.
- Zapier bills per task. A task is counted whenever Zapier successfully completes a unit of work, so a Zap with eight action steps burns eight tasks per run. Triggers, polling and built-in tools like Formatter don't count. The free plan gives 100 tasks a month; Professional starts at $19.99 a month for 750 tasks billed annually, or $29.99 billed monthly, and scales up in tiers from there. (Zapier pricing)
- Make bills per credit. Each module action in a scenario counts as one credit — same idea as a task. The free plan gives up to 1,000 credits a month, Core starts at $9 a month for 10,000 credits billed monthly, and higher volumes are bought per tier. (Make pricing)
- n8n bills per execution. "An execution is a single run of your entire workflow. It doesn't matter how many steps are in the workflow or how much data it processes." Starter is €20 a month for 2,500 executions, Pro €50 for 10,000, Business €667 for 40,000 — all three billed annually. (n8n pricing, retrieved 22 September 2026)
Read those three bullets again with a ten-step workflow in mind. On a per-step model that run costs ten units; on a per-execution model it costs one. The gap doesn't widen because you get more popular — it widens because your workflows get more thorough. Which is exactly what happens once you stop building toys. That is most of the reason the Vault is built on n8n, and the long version is here.
All prices above are what the vendors listed on their own pricing pages on 22 September 2026, and the billing period is named because the headline number differs by up to a third between paying yearly and paying monthly. Pricing in this market changes a few times a year, so check before you commit.
Breadth versus depth
The second real difference is catalogue size, and here the ranking flips.
- Zapier's own help documentation puts it at over 8,000 apps — comfortably the widest catalogue in the business. (Zapier help)
- n8n's integrations page currently lists 2,192 integrations, plus an HTTP Request node that reaches anything with a public API and a dedicated set of AI nodes for agents, chains, embeddings and vector stores. (n8n integrations)
- Make sits between the two, with deeper per-connection configuration than Zapier offers.
The practical rule: if the one tool your business depends on exists only in Zapier's catalogue, that decides it, and no amount of per-execution maths beats a missing connector. If your stack is mail, sheets, chat, a CRM and an invoicing tool, all three cover you and the billing unit decides instead.
The layer nobody advertises: where your data sits
Of the three, only n8n can be self-hosted in full, which is the entire argument for people handling client data under EU rules. The Community Edition is free software you run on your own server; the licence allows use for your own internal business purposes and stops at reselling n8n itself as a hosted service. (Sustainable Use License)
Self-hosting is not free, though — it is a server, updates and a backup you actually test. Whether that trade is worth it depends on one question: would you be comfortable if every mail your automation reads passed through a vendor in another jurisdiction?
The model layer got boring, which is good news
Two years ago the model you picked was the whole architecture. Now it is a node. Every Vault workflow ships wired to OpenAI with a one-line note on swapping in Anthropic's node instead, and the prompts work unchanged. The interesting development in 2026 isn't that the models got better — it is that the plumbing stopped caring which one you use.
That is worth money to a small business in a specific way: it means the thing you build this month is not a bet on a vendor.
The bottleneck nobody is selling a fix for
Here is the part the vendor comparisons miss. For a one-person business the hard part was never picking the tool. It is these three, and they show up in the same order every time:
- The blank canvas. A free tier and 8,000 connectors is not a starting point, it is a paralysis machine. People sign up, stare at an empty scenario, and close the tab.
- The second week. The first automation is a thrill. The one that silently stops working in week three is the reason most people quit — and almost nobody's tutorial covers an error handler.
- The upgrade. Self-hosted instances drift. n8n 2 broke old workflows quietly, which is the normal condition of infrastructure, not an exception.
Every one of those is a content problem and a craft problem, not a feature problem — which is precisely why the answer we built is a set of finished workflows rather than another tool.
If you only take one thing
Choose by billing unit and by where your data has to live. Choose by catalogue only if a tool you cannot replace lives in exactly one catalogue. Ignore integration counts otherwise — you are going to use eleven of them.