n8n has put a conversational builder at the centre of its workflow automation platform, and the detail worth noticing is in the billing model rather than the demo. n8n Assistant does not stop once it has drawn a workflow on your canvas; it runs the thing, opens the execution data, decides what broke and tries again — and each of those attempts draws down the AI credits attached to your plan. A request that lands first time is cheap. One that needs four rounds of repair is not.
Key takeaways
- n8n Assistant builds a standard, editable n8n workflow from a plain-language request, then executes and debugs it rather than handing over untested output.
- It fully replaces the older AI Workflow Builder, and its spend is metered separately against each plan's existing AI credit allocation.
- New n8n Cloud instances have it enabled by default; self-hosted Docker users need version 2.36 or later plus their own model keys, and Enterprise is excluded for now.
The pitch is about what you are left holding
Generating an automation from a sentence stopped being a differentiator some time ago. n8n's argument is about durability instead: ask a coding agent to wire up an integration and you inherit code that somebody now has to host and maintain, while an open-ended assistant that simply performs the task leaves nothing behind once the run ends. Neither result is easy to hand to a colleague who was not in the room.
What the Assistant emits is deliberately unremarkable — the same node graph a person would have assembled by hand, in the same project, with the same run history. That makes it inspectable by anyone on the team, versionable, and explainable without appealing to what a model happened to decide. For a platform whose core selling point against hosted no-code rivals has always been that you can see and own the logic, extending that property to generated workflows is a coherent position rather than a marketing line.
Generation is the cheap part
Most of the engineering sits after the canvas fills in. The Assistant executes what it built and reads the same per-node inputs, outputs and errors a developer would open, which is a meaningfully different claim from reporting success because something was produced. Failures trigger a diagnosis, a proposed fix, an applied patch and another run, with no JSON exports or copy-pasting between tools in between.
Ambiguity gets a question rather than a guess: a vague instruction about routing form responses draws a follow-up about which form and which destination. Credentials are requested at the moment a node needs one instead of up front, so nothing has to be configured before you see a first result. Two actions stay behind explicit human approval — granting credential access, and activating a workflow — which keeps the loop from quietly switching something on in production.
Who can run it, and on what terms
Availability splits three ways. New n8n Cloud instances get it switched on without any action; Enterprise Cloud is carved out of that default and broader Enterprise support is described only as on the roadmap; self-hosters need Docker, n8n 2.36 or newer, their own model keys and a handful of extra environment variables.
The credit accounting is the part worth modelling before anyone on your team starts experimenting. Existing allocations are untouched and usage here is tallied apart from the builder it replaces, so nothing already budgeted changes — but consumption scales with how many repair cycles a build needs, and top-up options are still weeks away. Teams handing the Assistant genuinely gnarly integrations may find the meter moves faster than the demos suggest.
What it explicitly will not do
The launch is unusually forthcoming about gaps. Nothing it produces is guaranteed to be production-ready, and reviewing the result stays a human responsibility. It cannot conjure an account on a service you do not already have, so third-party setup has not disappeared — only moved to a better moment. It is reactive rather than proactive, meaning no instance monitoring, no learned preferences and no unprompted suggestions. It also operates against a single instance and does not touch your browser or machine, though browser-assisted credential setup is under exploration.
Outlook
The preview flag signals active development rather than a waiting list, and the feature is already reaching cloud users by default. The real test is whether repair-and-rerun converges on workflows considerably larger than a demo, where an agent can burn credits cycling through fixes that never quite land. If it holds, the more interesting consequence is directional: agent tooling that produces durable, reviewable artifacts instead of one-off runs, a pattern visible elsewhere this year in agents designed to keep working between conversations. Configuration specifics are in the n8n Assistant documentation.
FAQ
Does n8n Assistant replace the AI Workflow Builder?
Yes, it supersedes it entirely. The difference n8n draws is that the earlier builder generated a workflow and stopped, whereas the Assistant also runs what it produced, reads the execution data and iterates on improvements.
Can self-hosted n8n users get the Assistant?
Docker deployments on version 2.36 or newer can enable it. Unlike n8n Cloud, self-hosting requires bringing your own model keys and setting additional environment variables first.
Is n8n Assistant safe to point at production systems?
Treat anything it produces as a draft. Credential access and activation both require your confirmation, and n8n states that a first generated workflow is not guaranteed to be production-ready. Read the nodes, verify the credentials and check the logs before activating anything business-critical.
