Industry
Estimators rekeying spec sheets, CSRs chasing proof approvals through email threads, purchasing retyping mill quotes into the ERP — that is the layer we automate. AI drafts; your people review and approve before anything touches a live system. The same architecture is running in live operations today at a North American manufacturing and packaging company, on their own servers.
The grind
Customers send specs as emailed PDFs, marked-up dielines, and spreadsheets — every one formatted differently. Your estimator burns the first hour of every quote rekeying dimensions, stock, colors, and quantities before the actual pricing work even starts.
The buyer approved artwork version 3; version 4 is sitting unanswered in someone's inbox; nobody is certain which one is at prepress. CSRs spend hours chasing sign-offs, and the cost of the wrong version going to plate is a reprint you eat.
Board, film, and corrugate prices move with every mill announcement, and vendor quotes arrive in different units — per hundredweight, per ton, per thousand sheets. Someone normalizes them by hand, and your standard costs lag reality.
Vendor invoices against POs and receipts, freight bills against shipments, commissions against invoiced jobs. It is days of careful three-way matching that a sharp person does on autopilot — which is exactly the problem.
The split
We automate
RFQ and spec-sheet intake — extracts dimensions, stock, colors, and quantities from emailed specs into a draft your estimator prices
Proof-approval tracking — logs who approved which artwork version, drafts follow-ups on stale approvals, flags version mismatches before plating
Substrate quote processing — reads mill and merchant quotes and normalizes them to your cost basis for side-by-side comparison — built on the vendor quote processing that runs live today
UPC/GTIN and spec validation — deterministic check-digit math and spec cross-checks across every SKU, no model guessing
Invoice and freight reconciliation — three-way match of vendor invoices, POs, and receipts, with exceptions queued for a person — the invoice match runs live today
We protect
The press OK and color judgment — whether the sheet matches the target is a trained eye's call, and it stays that way
Estimating and pricing decisions — margin, press fit, and whether to take the job at all belong to your estimator
Press scheduling — sequencing, gang runs, and makeready trade-offs are judgment we deliberately do not touch
Customer and vendor conversations — a late proof or a price increase gets a phone call from a person, not an automated reply
How it runs
AI drafts. A person reviews and approves. Nothing touches a live system without a human gate — that's the architecture, not a phase.
The full build architecture →A worked example
A merchant emails a quote for SBS board — price per hundredweight, basis weight, sheet size, lead time, all buried in a PDF. The automation extracts those terms (the only step that uses an LLM), converts them to your standard cost-per-thousand-sheets with plain arithmetic you can audit, and lays the result next to that vendor's last quote and your current standard cost. Your purchasing lead reviews the comparison, questions anything that looks off, and approves before a single number changes in the ERP. Which vendor gets the order — and whether a price increase deserves a pushback call — stays with the person who knows the vendor, because that relationship and that negotiation are exactly the work we refuse to automate. The vendor quote processing underneath this runs live today at a North American manufacturing and packaging company, on their own servers.
That is the one place we use a language model: pulling structured data out of unstructured emails, PDFs, and spec sheets. Everything downstream — unit conversion, check digits, matching, pricing math — is deterministic code you can audit line by line. And a person confirms every extraction before it enters your estimating or ERP system, so a misread never silently becomes a mispriced job.
It is in live operation at a North American manufacturing and packaging company: vendor quote processing, shipment tracking, invoice reconciliation, and commission calculation, all running on their own servers with a human approval gate on every action. We do not name them publicly. Those four workflows are deployed and running today; the rest of this page applies the same proven architecture to the specific paperwork a packaging plant generates.
No, and that is by design. Every automation runs in parallel with your manual process until it is proven, your team keeps the ability to operate without it, and nothing reaches your estimating system, the ERP, or the plate room without a person's sign-off — on the first job and on every job after. Recent research on automation and skill erosion is exactly why we work this way: we automate the rekeying and the chasing, never the pricing judgment or the approval itself.
We sit with your team, map how the work actually happens, and draw the split: what's worth automating, and what automation should never touch. You keep the map either way.
Map your workflows