Almost every company we talk to has a version of this process: an email arrives with an order — sometimes a PDF, sometimes loose text, sometimes an Excel — and a person reads it, interprets it and enters it into the system. Dozens of times a week. This note shows what changes when an intelligence layer enters that loop.
The starting point: email is not structured
The problem is not the ERP or the person — it is the format. An email saying "send us 200 units of that article from last week" doesn’t enter any system by itself. Someone has to translate human language into data: article, quantity, price, delivery terms.
That translation is exactly where language models are strong. They don’t replace the ERP; they feed it.
What the AI does — and what it doesn’t
The AI reads the email and its attachments, identifies the customer, extracts the order lines and checks them against the real catalogue (references, current prices, customer history). The result is not a chatbot answer — it is a structured record, ready for the ERP.
What the AI does not do: decide. If the margin is off, if the customer’s credit is suspended, if a reference is ambiguous — the case goes to a person, with the context already prepared. Human approval is not a limitation; it is a feature.
The complete loop
Email arrives → AI interprets and structures → data validated against the catalogue → entry prepared in the ERP → a person approves (when the rule requires it) → customer gets a confirmation. Every step is logged: what the AI read, what it proposed, who approved and when.
That log is what makes the system auditable — and it is what separates serious automation from magic nobody can explain when it fails.
When this pays off
The maths is simple: frequency × minutes per run × people involved. An occasional order justifies none of this. Thirty orders a week with ten minutes of manual work each — that is hours every week, every month, in the same place.
Before building, we measure. If the maths doesn’t work, we say so.
The essentials
| AI translates human language into structured data — the ERP remains the system of record | 01 | |
|---|---|---|
| Ambiguous cases go to a person, with the context already prepared | 02 | |
| Everything is logged: readable, auditable, explainable | 03 | |
| It only pays off when frequency justifies it — and that is measured before building | 04 |
Written by Pedro · Founder
Other notes
Why material loss never shows up where you look for it
In any operation that transforms raw material there is always a gap between what should have come out and what did.
Accounting without copy/paste: invoices that book themselves — with approval
The pile of supplier invoices is one of the most universal processes there is: they arrive by email, as PDFs, each supp…
AI in hospitality: the processes that no longer need to be manual
Hospitality runs on attention to the guest — which is exactly why it hurts to watch front-desk teams copy bookings from…
