A commercial furniture manufacturer takes orders from three directions at once: distributors sending PDF purchase orders, its own online brands selling under different SKUs, and walk-in business keyed straight into the ERP by hand. Until recently, every one of those orders was typed, translated, and shepherded through production by people who knew the quirks. We built a pipeline that reads the POs, resolves every line against the price list, routes assemblies to manufacturing, and posts clean sales orders to the ERP. This is what that actually involved, including the parts that surprised us.
The Starting Point
The company sells restaurant seating (chairs, barstools, table tops and bases) through distributor channels and its own ecommerce brands. Orders arrived by email all day. Each one meant: identify the customer and channel, translate the SKUs (the online brands use their own part numbers, different from the factory's), check pricing against the right price list, figure out which lines are assembled-to-order versus shipped-from-stock, and key the whole thing into the ERP. Ten to twenty minutes per order, all day, with the tribal knowledge living in two people's heads.
What the Pipeline Does
Reads whatever arrives
PDFs, spreadsheets, order emails, even photos. AI extraction pulls the PO number, ship-to, and line items, and every extracted value keeps a link to the exact spot on the document it came from, so the review screen can prove its work. We covered why that evidence matters in Human-in-the-Loop AI.
Knows the channels apart
A distributor PO and an online-brand order look completely different and follow different rules. The pipeline classifies the channel first, because everything downstream (which price list applies, which SKU translation table, even what a valid PO number looks like) depends on it. One detail that earned its keep: the online brands' PO number formats (prefix plus digit count) turned out to be a more reliable brand signal than anything written on the document, because one brand's orders arrive with another brand's letterhead more often than you'd think.
Translates SKUs like the veterans do
The online brands sell the factory's products under their own part numbers, hundreds of them. Those live in a translation table now, built from the company's own mapping exports rather than hand-entered. Where a brand SKU maps to more than one factory part, the system refuses to guess and flags the line. The most valuable translation lesson came from the factory floor: the manufacturing system and the price list name the same chair differently (one keys by frame color, the other by seat construction), so "is this an assembly?" can't be answered by string matching. It takes a rule set built with the people who run production.
Validates before anything posts
Price on the PO versus the applicable price list, complete ship-to, duplicate PO detection, and one rule that paid for itself early: the customer's quoted price is the agreed price; the catalog price is a variance check, not a billing rate. An early version had that backwards on one path, which would have overbilled a channel systematically. A price disagreement now routes to review instead of silently resolving in either direction.
Routes production and posts the order
Lines that need assembly generate work for manufacturing; quick-ship lines head to the warehouse. Clean orders post to the ERP through its integration layer with the ERP's own validation intact. Orders with flags wait in a review queue where a person sees the original document beside the extracted data and fixes only what's flagged.
The Part We Didn't Plan For: Orders Born in the ERP
A few weeks in, a gap surfaced: walk-in and phone orders were keyed directly into the ERP by staff, exactly as they always had been, which meant the pipeline never saw them, and they were invisible to the new picking and shipping screens. No PO to ingest, nothing to extract.
The fix reversed the flow: the pipeline now watches the ERP for orders it doesn't recognize and adopts them into the same workflow, deriving what the floor needs (assembly routing, pick classification) from the ERP's own data. The lesson generalizes: map every door an order can enter through, including the manual ones, before deciding what "all orders" means. An automation that handles 80% of orders beautifully and loses the other 20% from every screen isn't 80% done; it's broken in a way nobody can see.
Where It Stands
The system runs the order flow end to end: intake, translation, validation, ERP posting, production routing, pick lists, shipping paperwork, and invoicing that bills exactly what shipped, partial shipments included. The team's role shifted from keying orders to working an exception queue, and the exceptions keep teaching the system: confirmed matches become rules, and the clean-pass rate keeps climbing months after go-live.
Two takeaways if you're considering something similar. First, the AI extraction was the most mature part of the stack; the project's hard work was encoding the business rules (channels, translations, assembly routing) that lived in people's heads. Second, staged rollout wasn't a compromise; it was the plan. Everything went through human review until the accuracy was proven order by order, and automation was earned category by category. That's the pattern we'd repeat, and the subject of an upcoming post.
Uptimize Solutions builds order automation for manufacturers and distributors: AI intake, SKU translation, ERP posting, and the production routing behind it. If orders enter your business through more doors than your systems can see, see our AI workflow automation or book a free workflow audit and we'll map the doors together.