Automate before you AI: the order that actually saves money
Adding AI to a broken manual process just gives you an expensive broken process. Here's the workflow-first approach we use at Jora.
Every other inquiry we get starts with "we want to add AI to our operations." Our first question is always: have you mapped the process yet?
Nine times out of ten, the answer is no. And adding AI to an unmapped process doesn't fix it - it makes the mess more expensive.
Workflow first, AI second
The order that works:
- Map the process by hand. Write down every step, every handoff, every tool. You'll find 30% of the work is pure waste - duplicate data entry, manual status pings, "just checking" emails.
- Automate the deterministic parts. The if-this-then-that steps don't need AI. They need n8n, Make, or 50 lines of Python. This is cheap, reliable, and immediately visible.
- Now add AI for the judgment calls. Classification, extraction, summarization, draft generation - the steps where a human was reading something and deciding. This is where LLMs earn their cost.
A real example
A client wanted "an AI that handles our invoices." We mapped their process and found 7 steps - only 2 of which (reading the PDF, categorizing the expense) needed AI. The other 5 (downloading, validating, posting to the ledger, sending confirmation, archiving) were deterministic automation.
By automating the 5 first and adding AI for the 2, we cut processing time 90% at a fraction of the cost of the "all-AI" approach they originally wanted.
The lesson: AI is a tool in the automation toolbox, not the toolbox itself. Reach for it last, not first.