AI Automation Systems

The brief

Automation pipelines built with n8n and AI tooling: webhook glue between tools, document generation and notification flows for small teams. Each workflow is self-hosted and documented so the client can run it without an engineer in the loop.

Problem

The teams these pipelines were built for were not short of tools, they were short of connections between them. Someone was copying data out of a form, pasting it into a document, saving it to a folder and then messaging a colleague that it was ready. None of those four steps required judgement, and all four consumed a working week every month.

My role

I designed and built the workflows: mapping the trigger and hand-off points, writing the n8n flows, adding AI steps where a decision needed language rather than a rule, and writing the runbook the client uses afterwards.

Process

• Watched the process being done by hand first, and noted every place a human made a decision worth keeping
• Automated only the steps with no judgement, and left a human checkpoint wherever the output goes to a customer
• Built each flow as a small set of webhook-triggered n8n workflows instead of one large pipeline, so a failure is contained
• Added retries and a failure notification, because a silent automation is worse than a manual process
• Documented the flow, the credentials it needs and the way to re-run a failed item
• Added a dry-run mode so a flow could be exercised against sample data before it ever touched a real customer record

Result

Working n8n pipelines handling document generation, data hand-offs and notifications, with the runbook handed to the client. I removed the 'roughly 60% less manual work' claim that used to sit on this page, because it was never measured. The defensible outcome is that the flows run in production, they are self-hosted, and the client can re-run a failed item without calling a developer. The lesson that carried into client work is that the human checkpoint is not the compromise, it is the feature: every flow that touches a customer ends with a person approving the output, and that is why the teams kept using the pipelines after the novelty wore off. Everything here is deliberately small: a pipeline that fails loudly in ten steps beats one that fails quietly in a hundred.

What it had to do

• Webhook pipelines connecting tools the client already pays for
• Document generation and notification flows with human checkpoints
• Self-hostable workflow definitions with written runbooks

Stack

• n8n
• AI APIs
• Webhooks
• REST APIs
• TypeScript