AI order processing, step by step
AI order processing uses an AI model to read incoming orders and prepare them for your ERP, while your ERP's own rules still decide prices, credit and stock. The AI reads and proposes. A person approves.
The steps
Eight steps take an order from the inbox to an approved sales order. The middle column is the agent's work. The right column is what stays with your people, and it never shrinks to nothing.
| Step | What the AI does | What stays with people |
|---|---|---|
| Intake | Watches the shared orders inbox and picks up each new order email and its attachments. | Who can send to the inbox, and which senders are customers. |
| Reading | Reads the PO: customer, PO number, dates, addresses and every line, flagging what it is unsure of. | Checking any flagged field against the original. |
| Customer match | Matches the order to a customer account from the name, email domain and ship-to. | Choosing the account when a new site or a new buyer appears. |
| Line match | Maps each customer part number to your SKU from past orders and the part-number table. | Picking the item when a code is new or ambiguous. |
| Checks | Prices each line from your ERP and checks stock for the requested dates. | Deciding what to do about a price difference or a shortfall. |
| Draft | Writes a draft sales order in the ERP, or an import file, with the original attached. | Nothing yet. The draft is not a live order. |
| Review | Shows the draft beside the original with flagged lines highlighted. | Approving, editing or rejecting the order. |
| Confirmation | Sends the customer an acknowledgement after approval, if you switch it on. | Anything unusual the customer needs to hear from a person. |
Where rules beat AI
An AI model is good at reading messy documents and at matching things that look different but mean the same. It is the wrong tool for decisions your business already makes with rules. Those stay in the ERP:
- Pricing. Price levels, contract prices and quantity breaks are rules. The agent asks the ERP for the price and compares it with the PO. It does not decide what to charge.
- Credit holds. Whether an account can take more orders is your credit policy, applied by your ERP or your accounts team.
- Stock allocation. Which order gets the last ten units is a business decision, often made by a person who knows the customer.
Keeping those with the ERP is what makes AI order processing safe to switch on. The model never has the authority to give a discount or ship to an account on stop.
What uncertain means, and why it matters
A reader that is right most of the time but never says when it might be wrong forces you to check every field of every order. That is most of the work you were trying to remove.
So every value the agent reads carries a flag. When it guessed, when the scan was hard to read, when two values on the PO disagree, or when it inferred something the PO did not state, the field is marked uncertain and the reason is written beside it. People check the flagged fields and trust the rest, and the shadow test measures whether that trust is earned.
| Part number | Description | Qty | Unit price |
|---|---|---|---|
| OR-110 | O-ring nitrile 25mm | 50 | 0.85 |
| Quantity: the scan shows 50 or 30. Read as 50. | |||
This kind of check is the gate between reading and drafting. Lines with nothing flagged go into the draft as read. Flagged lines still go in, highlighted, so the person approving looks there first.
Order processing automation without replacing your ERP
Order processing automation used to mean a new ERP module or a migration. An agent does not need either. It sits beside the ERP you have and writes to it the way your own staff do: through a supported API where one exists, or through the sales order import that most ERPs already have.
Where there is no connector yet, approved orders leave as an import file. Price and stock checks then use your daily exports, so stock can be a day old, which the review screen says plainly. The connection can come later without changing anything for your team.
The detail of each step: sales order automation covers matching, purchase order OCR covers reading, and AI for distributors covers where the order desk sits among other uses of AI. Or try the reading step in the PO reader.
AI order processing questions
Is AI order processing safe to use on real orders?
It is when a person approves every order before it reaches the ERP, and when the system says which fields it is unsure of. Then the AI does the reading and the matching, and the decisions that matter stay with your team and your ERP's own rules.
What happens with orders that contain mistakes?
Mistakes in the order itself, such as a wrong code, a price that does not match your list or a total that does not add up, are flagged on the draft with the reason. The agent does not correct the customer. A person decides whether to fix the line or call them.
Which document types can it handle?
Order emails, PDF purchase orders, scans and faxes, phone photos of printed POs, and spreadsheets or CSV exports. Multi-page PDFs and email threads are fine; quoted earlier messages are ignored.
Does it learn from corrections?
Yes, in the parts that should be learned. When a person maps a customer code to one of your items, the mapping is kept, so the next order with that code is matched. It does not learn to change prices or skip approval.
Do we need to change ERP?
No. The agent sits beside your ERP and writes draft sales orders through its connection, or an import file where there is no connection yet. Your ERP keeps pricing, credit and stock.
See it on your own orders
A free shadow test runs a month of your emailed orders through the agent. Read-only, about a week.