How does AI order entry work for distributors?

What happens to a request between the inbox and the ERP?

One queue first. A catch-all address, orders@yourco.com, collects every channel: email, PDF attachments, WhatsApp messages, faxes, and voice notes in English, Spanish, Hebrew, or Yiddish. Nothing lives only on one salesperson's phone, and everything that follows is auditable.

Classification runs before extraction, because the first question is not what is in the message but what the message is: a new order, a quote request, a follow-up on something already placed, or noise. It is usually the first thing a desk asks about, and rightly, because a purchase order processed as a quote is worse than a slow answer. So when classification is uncertain, the message routes to a person for the call instead of into the wrong pipeline.

Then extraction: the request becomes structured line items with quantity, unit of measure, and description, whether it arrived typed, scanned, or spoken.

How does SKU matching handle a vague request?

Matching resolves each line against your item file, the inventory you upload, not a generic industry catalog. Your part numbers, your cross-references, your customers' shorthand and spelling variants map to your SKUs at your units of measure.

Ambiguity is the normal case, not the edge case. At one electrical distributor we talk to, a request for a 10 watt light maps to 50+ catalog variants. Customer context narrows the field; when it cannot narrow to one, the line is flagged with the candidates listed and a human picks. A flagged line costs a reviewer seconds. A wrong SKU costs a return, a re-ship, and a customer's confidence.

Pricing follows the same rule. Each customer's price level comes from your ERP and applies line by line. A price the system cannot resolve is flagged, not guessed.

Why does a human approve every order?

Because an order is not a document, it is a commitment: stock allocated, a truck loaded, an invoice cut. Every drafted order lands in an approval inbox where a reviewer confirms, edits, or rejects it. The job changes shape, not hands. Distributors we talk to average about 30 minutes of desk time per quote, most of it typing, and reviewing a clean draft takes a fraction of that. One inside salesperson at a distributor we talk to does about 100 quotes a day and works three hours past close; approval keeps the judgment in that day and removes the typing.

How much runs on its own is a dial, not a default. Assist drafts everything for full review, and Guarded and Autopilot let more through as the numbers earn it. The gate is your Human Edit Rate, the measured share of drafts a human had to correct: low and stable, loosen the dial; anything else, do not. In every mode, a flagged line waits for a person. That is deliberate. The system's job is to make ambiguity cheap to resolve, not to hide it.

How does the approved order get into the ERP?

On DDI Inform and Spruce, ERP write-back is live today: the approved order becomes a real order in your system with no retyping. Epicor, NetSuite, and SAP are in development. Every other system runs in ERP-light mode: the same queue, extraction, matching, and pricing in front, with clean ready-to-enter documents out the back while a native integration is scoped.

Getting started is two steps, connect your inbox and upload your inventory, followed by an included white-glove kickoff call where a human gets your desk live. ERP integration is scoped separately from those two steps.