Guides
How to reduce AI costs in n8n workflows
Use NuRoute's n8n node to test automatic model selection, inspect token usage and control retries. Includes setup steps and a summary-quality rubric.
Published 3 min read
On this page
When an n8n workflow gets expensive, first check how many model calls it makes for each completed item. A loop, a retry or an unnecessarily long response can cost more than the choice of model.
Once you understand that request chain, you can test whether automatic model selection reduces cost while keeping the result useful. This walkthrough uses a text-summary workflow and NuRoute’s official community node. Its gateway configuration follows the Platform documentation. Multi-provider controls in that node are Enterprise features; standard NuRoute access remains Artemis-only. The setup follows the production documentation; it has not been executed in a live n8n instance for this article.
Start with a summary task you can judge
Use a workflow that turns a support ticket into a short internal summary:
Manual Trigger → Edit Fields → NuRoute → Validate summary → ReviewFor a synthetic test ticket, use: "Invoice INV-1042 appears twice. The customer asks for a corrected invoice by Friday. No refund has been approved."
The summary must preserve the invoice number, the Friday deadline and the fact that a refund has not been approved. It must not claim that money has been returned. These checks are more useful than asking whether the response sounds good.
Install the NuRoute community node
The documented installation path is for self-hosted n8n: open Settings → Community Nodes → Install, enter n8n-nodes-nuroute, then restart n8n. The previous n8n-nodes-aicp package is deprecated. Check community-node availability for your deployment before following these steps on n8n Cloud. NuRoute n8n integration
Create a NuRoute API credential with your gateway base URL, without a trailing slash, and a Platform API key from your project's or organization's API Keys page. The node's credential takes the gateway URL; do not append the chat-completion endpoint path.
For cost isolation, use a project-scoped key for this workflow. A project key automatically applies that project's settings to its chat requests. Project-scoped keys
Configure one model call per item
Add the NuRoute node and set:
| Parameter | Initial setting |
|---|---|
| Model | auto |
| Messages | A system instruction and the ticket as a user message |
| Simplify Output | On |
| Provider | Leave unset for automatic model selection |
| Max Tokens | A small limit appropriate to a short summary; check for truncation |
A suitable system instruction is:
Summarise the ticket for an internal support agent in at most
three short sentences. Preserve identifiers, deadlines and unresolved
questions. Do not claim an action has happened unless the ticket says so.Bind the user message to the current item's ticket field with n8n's expression picker. Confirm the binding with two different tickets; each output should describe its own input.
For Enterprise gateway users, the production docs warn that setting Provider while leaving Model at auto bypasses the router: the pinned provider serves the first model it lists. Leave Provider unset for this Artemis test. Node parameters
Inspect usage and failed summaries
With Simplify Output enabled, the node returns content, model, finishReason and usage. Save those fields with an item ID and the review result. Check whether the summary passed all three factual requirements, not just whether the node completed.
If you need JSON, parse and validate it before downstream actions. NuRoute's routing reference currently says response_format is ignored. Asking for JSON in the prompt does not guarantee schema compliance. Structured-output limitation
Compare against the current workflow
Replay the same representative tickets through your current model and the routed version. Include short, long, multilingual and ambiguous tickets. Measure spend per accepted summary, error rate and end-to-end completion time.
Keep n8n execution costs separate from model costs. NuRoute gateway plans and applicable request overages also sit outside Artemis token charges. NuRoute pricing
Contain repeated attempts
n8n documents Retry On Fail and batching with Loop Over Items plus Wait for handling rate limits. Use a bounded attempt policy and test the actual error branch. Do not retry bad credentials or a project budget error as if they were transient failures. n8n rate-limit handling
If this workflow later sends a reply or updates a ticket, keep that action after validation and prevent duplicate execution. A repeated model call should not create a second customer action.
Use the production connector reference for parameter changes and the project budget guide to contain unexpected volume.