Journal · Short film
Your chatbot has never seen your books
What it takes for AI to work with a company's real numbers, using ERPNext as the example, and the two new rules that come with it.
Captions are on. The described version stops at each scene while a voice says what's on screen; that voice is synthetic, and the narration is ours. There's no sign-language version yet. The full transcript is below.
Why it guesses
Ask an AI chatbot how much your company sold last month. It will either admit it doesn't know, or worse, make something up. A language model learns from text, and your invoices aren't in that text.
Connecting it to the numbers
For AI to help run a company, it has to connect to where the numbers live: the ERP. ERPNext is a good example of how this can be done. It's open source (GPL-3.0), built on a framework called Frappe (MIT).
In Frappe, every kind of record is a DocType: customers, invoices, stock entries. Frappe generates a REST API for every DocType automatically, at addresses like /api/resource/Customer. So an AI can read the real invoice instead of guessing.
Two new problems
Memory. If the AI remembers last Tuesday's stock number and repeats it today, it's confidently wrong. It should re-read every time. A stale financial number is worse than no number.
Permissions. An AI that can read your books is useful. One that can edit them is a risk. Start read-only, and let it earn the rest.
What this doesn't prove
That connected AI gives right answers. If your stock count is wrong, the AI will be just as wrong, only faster and more politely.
The underlying data still has to be correct, and so does the process around it.
Transcript
Narration as spoken, with what's on screen described in the indented lines. Company names and amounts in the film are made up.
On screen: A generic mocked-up chat. The question: How much did we sell last month? First reply: I don't have access to your sales figures. Second reply: Last month you sold $1,284,000, stamped MADE UP in red.
Ask an AI chatbot how much your company sold last month. It'll either admit it doesn't know or worse, make something up. It's never seen your books and it doesn't know. A language model learns from text. Your invoices aren't in that text, so if you want AI to improve how your company runs, it has to connect to where the numbers live. Your ERP. ERPNext is a good example of how this can be done. It's open source, built on a framework called Frappe.
On screen: Three DocType boxes drop in: Customer, Sales Invoice, Stock Entry. Each gets an API path: /api/resource/Customer, /api/resource/Sales Invoice, /api/resource/Stock Entry.
In Frappe, every kind of record is a DocType. Customers, invoices, stock entries. And every DocType gets an API automatically. So an AI can read the real invoice instead of just guessing. Problem solved? Well, not quite. Now you have two new problems. One, memory. If the AI remembers last Tuesday's stock number and repeats it today, it's confidently wrong. So it should re-read every single time so as to not introduce stale data. A stale financial number is worse than no number. Number two, permissions. An AI that can read your books is useful. One that can edit them is suddenly now a risk. Start in read-only mode, and then let it earn the rest.
On screen: Card: What this doesn't prove. That connected AI gives right answers. Bad data gets wrong faster.
Connecting AI to your ERP doesn't make its answers right. If your stock count is wrong, the AI will be just as wrong, but only faster and more polite. The underlying data still has to be correct, and it still has to be a correct process.
On screen: End card: the Highland Willow willow-tree logo grows from its trunk and roots into its canopy, then the words Highland Willow, highlandwillow.com/journal, and We publish our working.
Sources
- Frappe Framework documentation: REST API.
- ERPNext source repository (licence: GPL-3.0). Frappe Framework (licence: MIT).
Narrated and produced by Highland Willow. The audio-description voice is synthetic (macOS text-to-speech). The animation is drawn in code. None of the film's data comes from any client.