I replaced my daily admin with three n8n workflows. The AI was the easy part.
Here is the admin that eats a founder's morning. Not the work — the paperwork wrapped around it. Reading a support email twice to decide whether it was billing or a bug. Opening a company's site to remember what they do before replying to their founder. Downloading an invoice PDF, retyping eight fields into a spreadsheet, closing it. Thirty of those a day, about four minutes each. The work is the…
The admin that consumes a founder's morning is not the work itself, but the paperwork surrounding it. Reading a support email twice to discern whether it relates to billing or a bug, opening a company's website to recall their purpose before replying to the founder, and downloading an invoice PDF to manually input eight fields into a spreadsheet are all examples of this tedious process.
On average, a founder spends about four minutes on each task, totaling thirty tasks daily. The initial step is often to utilize a language model, and while AI models can generate decent summaries of the provided input, the surrounding processes frequently fail. These include fetching data that returns JavaScript shells instead of text, parsing prose expecting numbers, processing items incorrectly in a loop, and scheduling tasks at inconvenient times due to instances being set in UTC.
To address these issues, the author built and tested n8n workflows, aiming to create efficient processes that survive on other instances. The three workflows discussed in this post carry the most load, with their underlying schemas and the three most significant bugs responsible for consuming the most time. Despite the model's ability to generate coherent text, it never determines the next steps, and all routing, thresholds, and permissions are managed by the code.
The model's output is solely used to fill fields, never instructing a human action. This ensures drafts are created, alerts are sent to the user, and ledger rows are updated. The human act of sending is left separate from the workflow. The first workflow focuses on lead research, using a webhook trigger with company domain, contact name, and notes from the source.
The model's output includes a company summary, industry, business model, fit score, and reasons for fit or non-fit. A schema is employed to validate the output, ensuring crucial fields like icp_fit_score (a numerical value) and confidence (enum: low, medium, high) are present. The personalization hook, a separate field from the summary, contains a one-sentence emailable statement referencing specific details from the source page.
A system prompt guides the model's behavior, prohibiting fabrication of funding, headcount, or customer names. Any missing or unreadable data prompts the model to provide a "confidence" score and conservatively rate the input. Before processing, the page text must be obtained as clean text, with markup and long strings removed. After processing, the model writes to a sheet, creating drafts, and sending pings to Slack, all while humans perform the final actions of distribution.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
This story
This is one outlet's version. Read the fullest account.