One in-house system crews use every day, two in-house automations, and a set of working example builds. Every workflow diagram below is drawn from the real thing.
A phone checklist and a live office board that replaced a 3-page paper form at the concrete floor coating company where Max runs operations.
Read the case study
Built for the same concrete floor coating company, where the paperwork is real and mistakes cost real money.
Supplier invoices arrive by email as PDFs. Each one is read, and every line item, plus shipping and tax, becomes a row in the invoice sheet.
The detail that matters: Nothing is written unless the rows add up to the invoice total. If they don’t, or the PDF can’t be read, a review email goes to the office instead.
Runs on: Gmail, Claude, Google Sheets
Someone uploads the month’s stock movement report in a simple form. A new tab named for the month appears in the inventory sheet.
The detail that matters: AI reads the lines, but plain code does all the adding, so the AI never totals anything. If a line is missing it refuses to write, and re-uploading a month makes a new tab instead of overwriting.
Runs on: n8n form, GPT-5 mini, Google Sheets
Working demos built to show a pattern. They are examples, not client results, and they run on sample data where a live account would go.
Emails asking the usual questions get the answer the business already wrote. Anything else stays in the inbox for the team.
The detail that matters: The AI only picks which approved question an email matches. The reply is sent word for word from a table, so it can never invent an answer.
Runs on: Gmail, n8n Data Table, AI label only
Each new support email is checked, de-duplicated and drafted into a ticket, then sent to Slack for approval.
The detail that matters: Nothing is created until a person clicks approve. Unreadable emails and off-spec AI answers are parked for manual triage, never lost.
Runs on: Gmail, Claude Haiku, Slack
Invoices and documents arriving as PDFs or pasted text are read into one operations table.
The detail that matters: Subtotal plus tax must equal the total, and scanned PDFs with no text are stopped before AI can guess at them. Anything that fails is held, not written.
Runs on: Gmail, Claude Sonnet, n8n Data Table, Slack
Form leads are checked, matched or created in HubSpot, and the owner hears about them by Slack and email straight away.
The detail that matters: If the CRM is down, the form gets a clear error back instead of the lead quietly disappearing.
Runs on: Web form, HubSpot, Slack, Gmail
Each new lead is scored and gets a drafted first email based on what they actually wrote. Cold leads stop there.
The detail that matters: Sales gets the draft in Slack to check and send. Nothing goes to a customer automatically.
Runs on: AI scoring, Slack
Every morning a short status message is written for each client from their numbers, in their own language, ready to send by WhatsApp.
The detail that matters: A failure for one client is logged and the rest of the run carries on.
Runs on: GPT-4.1 mini, WhatsApp (switched off in the demo)
Tell us what eats your week on a free 30-minute call.
