Most n8n office automation starts the same way. Someone watches a YouTube tutorial, connects Gmail to Google Sheets, and by Friday invoices are being logged automatically. Three weeks later the sheet has duplicate rows, one vendor got paid twice, and nobody knows which of the forty executions failed or why. The tool did exactly what it was told. The trouble is that nobody told it what to do when things went sideways.

Office work is mostly exceptions wrapped around a simple happy path. An attachment arrives as a scanned image instead of a PDF. A manager approves by replying "ok" instead of clicking the button. Someone sorts the spreadsheet and the row IDs move. If your workflow only handles the clean case, it isn't automation. It's a faster way to make mistakes. The fixes are not exotic, and n8n already has most of the parts.

Start with the task someone does every Monday

The best first workflow is boring, frequent and owned by one person. Invoice intake from a shared inbox is a good one. So is turning leave requests into calendar entries, or compiling the weekly numbers from three systems into one report. These have clear inputs, a known output and someone who will notice right away if the result is wrong. That last part matters more than it sounds, because a workflow nobody checks will drift for months before anyone notices.

Avoid the tempting one-off that touches everything, like "automate onboarding" across HR, IT, payroll and Slack in a single canvas. That workflow has twenty failure points and no single owner, and it runs maybe twice a month, so bugs hide between runs. Write down the manual steps exactly as they're done today, including the ones people do without thinking, like checking whether the invoice number has already been entered. Those unspoken checks are the logic your workflow is missing.

The spreadsheet is not your database

Google Sheets is a fine place for people to read results and a poor place for a workflow to keep state. People insert columns, rename tabs, sort ranges and paste over formulas, and every one of those changes can quietly break a lookup that worked yesterday. When a workflow has to remember things between runs, like which emails it processed, which approvals are pending or which invoice IDs exist, that memory belongs in a real table.

A small Postgres database next to a self-hosted n8n instance costs almost nothing and removes a whole class of bugs. The workflow writes to Postgres, and a separate step mirrors a read-only view into the sheet people actually look at. If someone mangles the sheet, you rebuild it from the table. If the workflow misbehaves, you can query exactly what it did. This one split is behind most of what we cover in building n8n automations that don't break.

Make every run safe to repeat

Workflows get re-run. A trigger fires twice, a node times out and retries, or someone hits "execute" while debugging. If running the same input twice creates two payments or two calendar entries, you will eventually get both. The fix is a dedupe key: something stable about the input, like the email's Message-ID or the vendor plus the invoice number, which you check before doing anything with side effects.

Put a unique constraint on that key in the database and let the insert fail loudly when it's a duplicate. Then branch on that result instead of guessing. Do the same with outbound actions. Record that you sent the approval request before you send it, and record the response when it arrives. A workflow that can be safely re-run from the top turns most incidents into a two-minute fix instead of an afternoon of reconciling rows by hand.

Put the human in the flow, not around it

Most office processes need a person to approve something. The fragile version emails a manager and hopes someone updates the sheet afterward. The sturdy version uses n8n's Wait node to pause the execution and resume it through a webhook, sending the approver two links that carry the decision back into the same run. The workflow then knows who approved, when and on which version of the request, without anyone retyping it.

Set a timeout on every wait and decide what happens when it expires. Usually that means a reminder, then escalation to a second person, never a silent approval. Deciding which steps a person must touch and which can run unattended is a design choice, not a detail, and it's worth thinking through properly. We wrote about that trade-off in attended versus unattended automation.

Own the failure before it owns you

Every production workflow should point at an error workflow built on the Error Trigger node. It should post the workflow name, the failing node, the error message and a link to the execution in a channel someone actually reads. Silent failure is the default in most setups, and it is the reason office teams lose trust in automation and drift back to doing things by hand.

Credentials are the other quiet risk. If the Gmail connection is authorised with an employee's personal account, the workflow dies the day that person leaves or changes their password. Use shared service accounts, keep credentials scoped to what each workflow needs, and export workflow JSON to version control so a bad edit can be rolled back. If you'd rather have this built properly from day one, that is the kind of workflow automation work we do: fewer workflows, each one owned, logged and safe to re-run.

Questions, answered.

Is n8n good for office automation?

Yes, especially for inbox processing, approvals, reporting and syncing data between business tools. It needs the same discipline as any software, though: real state storage, dedupe keys and an error workflow. Without those, it fails quietly.

Can n8n use Google Sheets as a database?

It can read and write Sheets, but using a sheet as the workflow's memory is fragile because people edit, sort and restructure it. Keep state in a proper database like Postgres and mirror a read-only view to Sheets for people.

How do I add approvals to an n8n workflow?

Use the Wait node set to resume on a webhook call, and send the approver links that carry their decision back into the paused execution. Always set a timeout with a reminder or escalation path so requests don't stall forever.

How do I get alerted when an n8n workflow fails?

Build a separate workflow that starts with the Error Trigger node and set it as the error workflow in each production workflow's settings. Have it post the failing node, error message and execution link to a channel your team watches.

What office tasks should I automate first with n8n?

Pick something frequent, well defined and owned by one person, such as invoice intake, leave requests or a weekly report. Avoid large cross-department workflows like full onboarding until the basics are proven.