EDI was supposed to end this twenty years ago. It did, for your three largest accounts. The other two hundred still email a PDF.
So somebody in customer service opens the attachment, reads it, and types it into the ERP. Forty lines, sizes and colours, a customer reference that has to go in the right field, a delivery date in the notes. Then the next one.
This is the highest-volume manual task in most wholesale businesses and it is almost never measured. Here is the whole path, and where AI actually fits into it.
Step 1: the order arrives, in whatever form it likes
A PDF generated by the customer's system. A spreadsheet. An order typed into the body of an email. A photograph of a handwritten sheet from a rep at a trade show. Occasionally all four from the same customer depending on who sent it.
Reading unstructured documents is the thing modern AI is unambiguously good at. Format variation, which is what defeated every previous attempt at this, is the part it handles best. This step is essentially solved.
Step 2: matching to your products
This is where it gets real.
Your customer orders using their code, or the style name, or last season's colour name, or a description. You need your SKU. That mapping is the hard part of the entire problem and it always has been.
AI handles the fuzzy cases well: "navy rain jacket size medium" resolving to the right SKU when there is one obvious candidate. What it must not do is guess when there are two. If the customer wrote something that could plausibly be either of two colours, the correct behaviour is to stop and flag it, not to pick the more likely one.
A system that is right 97 percent of the time and confidently wrong the other 3 percent is worse than no system, because nobody trusts it and everybody double-checks it. Confidence thresholds are the whole design.
Step 3: pricing
Customer-specific price lists, volume breaks, promotional pricing, contract rates, currency. Most of this already lives in your ERP, which means the right move is to let the ERP do it rather than have AI reproduce the logic.
Where AI helps is the exception: the customer wrote a price on the purchase order and it does not match what your system says. That is a conversation, and it should be flagged before the order is confirmed rather than discovered at invoicing.
Step 4: stock and allocation
Whether you can fill it, and what happens when you cannot.
Here is where a firm line belongs. AI can tell you what is available, what is short, and what is on the water. It should not decide who gets the last two hundred units when three accounts want them. That decision involves the relationship, the season, the account's history of over-ordering and then cancelling, and who you spoke to last week. None of that is in your data.
Step 5: exceptions
Discontinued lines. Minimum order quantities not met. Credit hold. Delivery dates that cannot be met. Substitutions the customer might accept.
Each of these needs a human decision and a message back to the customer. AI drafts the message and assembles the context. The person decides and sends.
What actually changes on the desk
The job stops being typing and becomes reviewing. An order arrives staged in the ERP with the confident lines matched and the uncertain ones highlighted. The customer service officer checks the exceptions, resolves them, and releases it.
A forty-line order goes from something like twenty minutes to something like three. More usefully, the person spends those three minutes on the two lines that were genuinely ambiguous instead of on the thirty-eight that were obvious.
Rolling it out without risking an order
Nobody should switch this on across all accounts on a Monday.
Run it in parallel first. AI stages the order, a person keys it the old way, and you compare. Two or three weeks of that tells you the real accuracy rate on your actual order mix rather than a vendor's claim. Then move one customer, then a segment, keeping the manual path available throughout.
Start with the accounts that send the most consistent format. Save the trade-show photographs for last.
The question everyone asks
Whether the customer service team loses their jobs. In every rollout we have seen, no. The order volume that was capped by keying capacity goes up, the same team handles it, and the work shifts to the exceptions and the customer conversations, which is the part they were hired for and the part that was being squeezed out.
Worth saying that out loud early, because a team that suspects otherwise will find reasons the system does not work.
Where to start
Count first. One week, every order that arrives by email, how many lines and how long each took. Most businesses are surprised, and the number is the business case.
After that it is a question of whether to train the team to do this faster or to build it out entirely. High volume and rules based means build. AI training for wholesale and distribution teams covers the training side, and the AI Ops Audit works out which of the two your order desk actually needs.