Order entry is where many small sales and distribution teams lose their week, and not to hard decisions. It is a queue. Orders arrive by email, as PDFs, as spreadsheet attachments, and sometimes as a forwarded phone note. Someone opens each one, finds the customer, works out which of your products the customer means, checks the price and the stock, keys it into the system, and writes back to confirm. Most of that is looking things up and comparing them, which is the kind of work an agent does well, as long as the decisions about credit, price and substitutions stay with a person.
§ 01Where an agent takes the work
- Order intake. The agent reads the email body, the PDF or the spreadsheet and pulls out the customer, their purchase order number, the delivery address, the lines with quantities and the date they want it. Everything lands in one consistent record, whatever format it arrived in.
- Customer and product matching. It matches the sender to the right account, and the customer's own part numbers or descriptions to your product codes, using their past orders. When it is not sure, it lists the likely candidates instead of picking one.
- Price and terms check. It compares each line with your price list and any agreed customer pricing, and flags every line that differs, with both figures shown.
- Stock and lead time check. It reads availability and marks the lines that cannot be filled from stock, so a person can decide on a part shipment or an alternative.
- Order creation. Orders that pass every check are created in your system. The rest wait in a queue with the reason they stopped.
- Confirmation and chasing. The agent confirms receipt, asks for a missing purchase order number or delivery date, and logs the answer.
The pattern is the same in all six: read from your systems, prepare the order, and hand anything unusual to a person. That narrow definition is why these deployments hold up. The engineering behind it is written up in our Lab notes.
§ 02Where a person stays in charge
- Credit holds. Whether to ship to a customer who is behind on payments is a commercial decision, and it depends on the relationship. The agent flags the account and stops.
- Price overrides and discounts. If a customer asks for a price that is not on the list, the agent passes it to the salesperson with the order history. It never agrees a price itself.
- Substitutions and backorders. Offering a different product or splitting a delivery changes what the customer gets, so a person decides and the agent follows up.
- Large or unusual orders. Anything above a value or quantity you set gets a human look, however clean the checks. How that approval queue works.
- Changes to an order that is already being picked or shipped. The agent flags the request and a person decides what can still be changed.
§ 03What actually goes wrong
It is rarely the reading of the order. The trouble is everything around it.
- Duplicate orders. A customer emails an order and then sends the same one again as a reminder, or sends a PDF and also places the order in a portal. The agent has to check the customer and the purchase order number before it creates anything, and creating an order has to be safe to repeat.
- Units of measure. Ten might mean ten boxes, ten cases of twelve, or ten single items. The right reading depends on the customer, and the agent learns it from their history. Where it is unclear, a person is asked.
- Customer codes and nicknames. Customers call your products whatever they like. The mapping has to be built from past orders and corrected when it is wrong. Expect the first weeks to be mostly corrections.
- Prices kept in more than one place. The list is in the system, one customer's special deal is in an old email, another is in a spreadsheet. The agent can only check what it can read, so decide early where the price of record lives.
- Stock numbers nobody trusts. If the count is wrong, a confident answer is wrong too. That is why the first confirmation says the order has been received, and a delivery date is promised only after a person has checked.
§ 04Start read-only, then earn write access
The first version reads the inbox and your system, prepares the order and lists the checks it ran, and writes nothing. Someone on the order desk reviews a queue instead of building it. Then track one number: how often the person changes the prepared order. If it is high, the rules or the data are wrong, and you found that out for the price of a small build. If it is low and stable, let the agent create the routine orders that pass every check and keep the queue for exceptions. Credit holds and price overrides stay with people the whole way.
§ 05Cost and timing
A single agent covering one part of the process, for example order intake and matching against one system, costs 3,000 to 5,000 USD and takes 2 to 3 weeks. A coordinated set covering intake, matching, stock and price checks, order creation and confirmations costs 15,000 to 50,000 USD over 6 to 12 weeks. The price moves with the number of systems involved and how tidy your product and price data is, not with how many orders you receive. The full breakdown is in our cost guide.
§ 06Is this just EDI or RPA?
EDI works well when a customer sends structured orders, and rule-based automation copes with orders that look the same every time. Most small and mid-size customers do not send them that way. They send an email, a PDF or a spreadsheet that looks a little different each time, and that is where an agent helps. Most teams end up using both. The comparison is in AI agents vs RPA. The same approach applies to accounts payable on the supplier side, and to customer support for the questions that follow an order.
──
If your order desk spends its week keying emailed orders, a 30 minute discovery call will show which step is worth automating first and which should stay with a person. Engagement models and prices are on our Hire page.