AI for wholesale distributors: orders, quotes and ERP.

02Which distributor workflows are worth automating first?

Four flows carry most of the typing, and they differ a lot in risk:

Start with order entry. It has the most volume, the clearest measure (lines that needed no human touch) and a safe landing place: a draft order that a person releases.

03How does AI turn a customer’s email or PDF into an ERP order?

A reliable pipeline has six steps, and only two of them involve a language model:

1. Ingest and classify. Separate new orders from change requests, order-status questions, invoices and spam. A change to an existing order must never become a second order.
2. Extract. Pull customer, PO number, ship-to, requested date and each line (customer part number, description, quantity, unit) from the body, the PDF or the spreadsheet.
3. Identify the customer. Match the sender and the sold-to against your customer master. An unknown sender stops here.
4. Match every line. Resolve each line to an item, with the unit of measure, using the rules in the next section.
5. Validate against the ERP. Check the price against the customer’s price list, stock, credit status and ship-to.
6. Write a draft and queue exceptions. Clean orders post as drafts; anything uncertain goes to a person with the source email beside the proposed order.

The write-back is where ERPs differ, and it is usually the largest piece of integration work. Microsoft’s Business Central API exposes a sales order with an external document number field for the customer’s PO and a status of Draft, In Review or Open. Oracle’s documentation for the NetSuite sales order REST record notes that taxation through REST needs SuiteTax enabled, and that quantity, rate and amount of item-group member lines can only be changed with a PATCH after creation. SAP’s course material on the Business One Service Layer shows an order created by a POST to the Orders endpoint, with a DocEntry returned that every later update depends on. Kits, tax setup and document numbering get tested early, on your configuration.

Large accounts often send EDI already. The common documents are the 850 purchase order, 855 acknowledgment, 856 ship notice, 860 change and 865 change acknowledgment, as IBM’s list of supported X12 document types shows. EDI needs mapping, not AI. The language model earns its place on everything else, and it should feed the same validation and write-back path your EDI orders use.

04How does it match customer part numbers and SKUs without guessing?

This is the step that decides whether people trust the system. Customers rarely use your item numbers. They use their own part numbers, a manufacturer’s number, an old catalog number or a description typed from memory. The design that works:

• Exact matches first. Look up the customer’s own part number in a customer-specific cross-reference before any model is involved. Most repeat orders should resolve here.
• Constrain the model to your catalog. For descriptions, retrieve candidates from the item master and let the model choose among them. It never invents an item number. Our note on retrieval that admits when it doesn’t know covers the abstention behavior this needs.
• Treat unit of measure as part of the match. “10 boxes” and “10 each” can be a hundredfold difference. A line with an ambiguous unit goes to a person.
• Handle supersessions by rule. A discontinued item produces a proposed replacement, flagged for review, never a silent swap.
• Learn through the table, not the model. When a person corrects a match, write the correction to the cross-reference. The next order resolves by exact match, and you have an audit trail of who approved it.

Set confidence thresholds per line, not per order: one uncertain line should not hold the other forty.

05Can AI draft quotes, and how do approval limits work?

Yes, with a division of labor: the model reads the request and assembles the quote, and the ERP computes the price. A language model should never calculate a price or discount from its own judgment. It calls the price engine for list price, customer contract price, quantity breaks and landed cost, and it reports what came back.

Approval limits then sit in code, outside the model, so no instruction in an email can talk past them:

• a price floor or minimum margin per item class and customer tier;
• a maximum discount from the customer’s contract price;
• a maximum quote value that can go out without review;
• a required review for non-stock items, special orders and long lead times;
• a block when the customer is on credit hold in the ERP;
• a record of who approved anything above a limit, and when.

This is the same principle OWASP describes as avoiding excessive agency: limit what the system can do, give it only the permissions it needs, require a person to approve high-impact actions, and enforce authorization in the downstream system rather than trusting the model to decide. Our post on prompt injection as input validation explains why the email itself is untrusted input.

06How should supplier confirmations and backorder updates work?

Supplier acknowledgments arrive as EDI 855s from some vendors and as emailed PDFs from the rest. Document AI reads the PDFs; both paths update the same fields: confirmed quantity, promise date and any substituted item against the open purchase order.

The useful part is the link from purchase order to sales order. When a promise date moves, the system finds the sales order lines that depend on that receipt and drafts the customer notice. Small slips inside a tolerance you define can send automatically from approved templates. A slip that breaks a customer’s committed date, or touches a key account, goes to the account manager with the new date, the alternatives from other warehouses, and a draft message.

Two guardrails matter here. First, a supplier email that says its bank details have changed is not a confirmation. The FBI’s guidance on business email compromise is to verify changes in account information through a secondary channel, so route those messages to a person every time. Second, once an invoice has been issued, an order change must produce a credit note and not an edit. In the UAE, the Ministry of Finance says an electronic credit note is required where a transaction is cancelled or the agreed consideration is reduced. Saudi Arabia’s ZATCA detailed guideline applies clearance to tax invoices and their credit and debit notes, and says an issued invoice is not modified or deleted: it is cancelled by issuing a credit note and a new invoice. The agent proposes; your ERP’s invoicing path issues.

07Where is AI the wrong tool, and where does a person stay in charge?

Several problems that distributors bring to “AI” are better solved without a language model:

A person stays in charge of these, whatever the model’s confidence:

• Orders from a new customer or a new sender. A joint advisory from the FBI, FDA and USDA describes criminals impersonating real companies, using look-alike email domains, to order full truckloads of food products on credit. It advises independently verifying new customers and confirming credit requests that were not required before. Your agent should hold these.
• A changed ship-to address or changed payment instructions. The same advisory warns about last-minute changes in shipping destinations and account information.
• Anything above an approval limit, including special pricing, returns and credit memos.
• Credit decisions about individuals. Under the UK GDPR, as changed by the Data (Use and Access) Act 2025, decisions based solely on automated processing with a legal or similarly significant effect face specific restrictions. The ICO’s guidance on automated decision-making was still in draft at its 31 March 2026 update. If some customers are sole traders, have a qualified professional check how this applies to credit and account decisions.

The system’s job is not to take people out of order entry. It is to make sure people only see the lines that need judgment.

08Which rules apply, by market?

As of October 2026, and without legal advice:

• United States. Robinson-Patman (above) for price differences between competing buyers, and the FBI’s fraud guidance for new-customer and payment-change controls.
• United Kingdom. The UK GDPR rules on automated decisions about individuals (see above). For VAT invoices, HMRC’s electronic invoicing notice says they must contain the same information as paper ones, and the supplier must keep authenticity, integrity and legibility.
• European Union. Under the AI Act, Article 4 asks providers and deployers to take measures to support the AI literacy of their staff. The EU’s AI Act Service Desk timeline shows the Article 50 transparency rules applying from 2 August 2026. They matter if a customer-facing assistant chats with buyers: systems that interact directly with people must be designed so those people are told they are dealing with an AI, unless that is obvious to a reasonably well-informed person.
• Gulf. UAE e-invoicing moves through accredited providers on the Peppol standard, and the Ministry of Finance’s timeline requires businesses at or above AED 50 million in revenue to implement by 1 January 2027, and smaller ones by 1 July 2027. Saudi Arabia’s ZATCA integration phase has been rolling out in waves since 2023. Your ERP’s invoicing path must be ready first, and the AI sits upstream of it.

09What should you measure, and what is an exception rate?

Define the numbers before you build, and count lines, not just orders:

• Touchless rate: the share of order lines that reach a draft order with no human edit.
• Exception rate, by cause: the share of lines sent to a person, split into unmatched part number, unit of measure, price mismatch, credit hold, ambiguous quantity and unknown sender. The split tells you which table or rule to fix next.
• Escaped error rate: mistakes found after release, such as a wrong item shipped or a credit memo raised. This is the number that matters most. A low exception rate with a rising escaped error rate means the thresholds are too loose.
• Time from email received to draft order, and quote turnaround from request to approved quote.
• Override rate on quotes: how often a person changes the drafted price or terms.

Score a labeled set of your own past orders, worst ones included, before launch, and keep scoring a sample after. Our evals guide explains how, and fallback chains covers what the system does when the model, the ERP or an inbox is unavailable.

10How do you start?

1. Choose one inbox and one customer segment. Pick the order stream with the most volume and the most consistent formats.
2. Build the item cross-reference first. Export existing customer part numbers from the ERP and clean the worst gaps before any model runs.
3. Run in shadow mode. The system proposes orders next to your staff’s own entry for a few weeks. Compare line by line and tune the thresholds.
4. Go live with drafts and review. Clean orders post as drafts. Exceptions go to a queue with the source email beside the proposed order.
5. Release more only on evidence. Let a customer or line type skip review once its escaped error rate has stayed low, then add quotes and supplier updates on the same layers.

The same layers carry to adjacent work such as logistics documents and shipment updates. If you want a partner to build it, see our services.

11Questions distributors ask

Do we need to replace our ERP to use AI for order entry?

No. The system reads your inboxes and writes to the ERP through its API, and it posts orders as drafts so your existing release and credit checks still apply. Older systems with weak APIs usually take the longest to integrate.

What exception rate is realistic?

It depends on how clean your customer part-number cross-reference is and how varied the order formats are, so any figure quoted before testing is a guess. Measure it on a labeled sample of your own past orders, and track escaped errors alongside it.

Can the AI approve discounts on its own?

Only within limits you set in code, such as a maximum discount from the contract price or a minimum margin. Anything outside those limits goes to a named approver, and the approval is recorded.

What happens when it cannot match a part number?

It holds that line, lists the closest candidates from your item master, and asks a person to choose. The choice is written back to the cross-reference, so the same part number resolves automatically next time.