Building Business Process Automation with n8n, Claude API, and Supabase
By Automation Systems Group · 5 min read · 1 September 2026 · Updated 10 November 2026
SaaS subscriptions for point-to-point automation often lead to fragmented tooling and high recurring costs. A self-hosted n8n instance backed by PostgreSQL provides infinite flexibility and zero vendor lock-in.
n8n, the Claude API and Supabase make a practical stack for business process automation: n8n moves data between systems and runs the steps, Claude reads messy inputs like emails and attachments and returns structured data, and Supabase stores clean records in a Postgres database. The parts that make it reliable are less exciting: a measured baseline before you start, validation before anything is saved, an error workflow that alerts a person, and a human check wherever a mistake would be costly.
Core Architecture
By deploying n8n on a lightweight VPS with Docker, our clients route inbound email attachments, webhooks from Stripe, and customer inquiries through automated classification models that extract line-item data and commit clean records into Supabase in under 1.2 seconds.
What each part does
| Component | Job in the workflow | Notes |
|---|---|---|
| n8n | Triggers, routing, calling APIs, retries, notifications | Can be self-hosted or used as a cloud service |
| Claude API | Classifying, extracting and summarising unstructured text | Ask for structured output and validate it before use |
| Supabase | Storing records, auth, file storage, simple dashboards | Built on a full Postgres database |
According to n8n’s hosting documentation, you can self-host n8n on your own infrastructure, on-premises or in a private cloud, using Docker Compose and other methods, and without a licence key it runs as the free Community edition. Supabase’s documentation says every Supabase project gets a full Postgres database, which is what Auth, Storage, Realtime and Edge Functions are built on. That matters for lock-in, your data sits in standard Postgres you can export.
Which processes to automate first
Good first candidates share a few traits: they happen often, follow mostly the same steps, involve copying data between systems, and have a clear right answer. Typical examples in Singapore SMEs:
- Checking supplier invoices against purchase orders.
- Syncing new customers or orders between an online store, CRM and accounting system.
- Sorting customer enquiries by type and urgency, and routing them to the right person.
- Building a daily or weekly summary report from several spreadsheets.
- Sending alerts when stock, payments or deadlines need attention.
Before building record the baseline: how many hours a week the task takes, how often it goes wrong and how long it takes from start to finish. Without that you cant show the automation worked.
Worked example: an invoice-check workflow (illustrative)
This is a simplified, made-up example of how the pieces fit together.
- Trigger: n8n watches an accounts mailbox for emails with PDF attachments.
- Extract: the attachment text goes to Claude with instructions to return supplier name, invoice number, date, line items and totals as structured data.
- Validate: n8n checks the output: required fields present, numbers add up, date is valid. Anything that fails goes to a review queue.
- Match: the invoice is compared with the purchase order stored in Supabase. Matches are marked ready; differences are flagged with the reason.
- Store: the clean record is written to Supabase, with a link to the original file.
- Notify: the finance team gets a daily summary of matched, flagged and failed invoices, and only handles the exceptions.
Making workflows recover from failures
APIs time out, emails arrive in odd formats, and someone changes a spreadsheet column. Plan for it:
- Use an error workflow. n8n lets you set an error workflow for each workflow; it starts with an Error Trigger and can send alerts by email or chat when an execution fails.
- Retry sensibly on temporary errors, with a limit, and send the item to a review queue after that.
- Make steps safe to run twice, for example by checking whether an invoice number already exists before inserting it.
- Keep the raw input so a failed item can be reprocessed after the fix.
- Watch for silent failures: a daily count of items processed, so a workflow that quietly stops is noticed.
Security and personal data
- Give each workflow its own credentials with the least access it needs.
- Store keys in n8n's credential store or environment variables, never in workflow notes or spreadsheets.
- Strip personal data that the AI step doesn't need, and check the AI provider's terms on data retention and training. See keeping company data out of model training.
- Turn on row-level security for Supabase tables that hold customer data.
- Under the PDPA you remain responsible for personal data you process this way. This is general information, not legal advice.
Self-hosted or cloud?
| Self-hosted n8n | Managed cloud | |
|---|---|---|
| Control over data location | Full | Depends on the provider's options |
| Maintenance | You patch, back up and monitor it | Provider handles it |
| Best for | Teams with someone responsible for the server | Teams that want less to run |
When not to automate
Skip or delay automation when the process changes every week, when it happens only a few times a month, or when every case needs judgement. Fix the process first and then automate it. Automating a messy process just produces mess faster.
We build these workflows for clients, starting with a workflow audit and agreed baseline measures. See business process automation, read about connecting AI bots so they hand work to each other, or tell us about the task.
Frequently asked questions
According to n8n’s documentation, a self-hosted instance without a licence key runs as the free Community edition, and paid licences enable Business or Enterprise editions. You still pay for the server and the time to maintain it. Check n8n’s current terms for your use.
A spreadsheet is fine to start, but it breaks down with many records, several workflows writing at once and access control. Supabase gives you a full Postgres database with authentication and row-level security, and the data stays exportable.
It handles messy text well, but no AI step should be trusted blindly. Ask for structured output, validate every field, and route anything that fails or looks unusual to a person.
Record the baseline before you start: hours spent per week, error rate and turnaround time. Measure the same things after launch, plus how many items needed a person to step in.