Every Emailed PO Gets Typed Twice

It’s 7:40 on a Tuesday. A mechanical contractor’s purchasing clerk emails a PDF purchase order: 38 lines, the job name, a ship-to address at the site, and a note that says “per your quote” with a quote number. Your inside salesperson opens it, opens the quote in the ERP in a second window, and starts keying the order line by line.

The quote has 41 lines. The PO has 38. Three were dropped, two quantities changed, and on line 22 the contractor wrote “10 BX” for an item your system sells by the each. The rep catches the missing lines. The unit of measure slips through.

Thursday, the foreman on site opens a carton and finds 10 wire connectors where he expected 10 boxes. He calls. Your driver makes a second trip, and somebody writes a credit memo for the freight.

I should be clear about my angle here. I’ve never worked an order desk. I ran a digital agency, Infographic World, for almost 18 years, and now I use AI to fix the work that gets stuck inside companies. I don’t know your item master. What I’ve seen, over and over, is that the costly part of a process is usually the handoff where somebody copies information from one place to another.

The order desk is one of those handoffs. The quoting article covered the race to win the order. This one is about what happens after you’ve won it.

Where the order desk time actually goes

Follow one emailed PO from inbox to released order. The keying is only the visible part.

• Finding it. POs land in a shared inbox, a rep’s personal inbox, a fax-to-email address, a text from a foreman, or a photo of a handwritten list. Somebody has to notice each one and decide who owns it.
• The header. PO number, job name, ship-to, requested date, tax status, the buyer’s name.
• The lines. The customer writes their own descriptions and part numbers. The rep translates each one into your item number, often from memory.
• Units of measure. Boxes, cases, rolls, feet, each. This is where a lot of quiet errors live.
• Price. The PO shows a price. Is it the quote price, the customer’s level, last year’s job price, or a number the clerk made up? The rep has to check.
• The quote match. When the PO references a quote, every line should tie back to it. Doing that by eye on a 40 line order is slow and easy to get wrong.

Here’s the part that bugs me. When the PO references a quote, those lines were already typed once, at quote time. The order desk types them a second time. That’s the double entry in the title.

What an AI-assisted order desk looks like

It reads every PO, in whatever form it arrives

The system watches the order inboxes and reads PDFs, email bodies, spreadsheets and phone photos, pulling out the header and every line, including any quote reference. Handwriting it can’t read with confidence gets flagged for a person.

It matches the PO to the quote and the item master

If the PO cites a quote, every line is set against that quote. Matching lines are marked matched. Then it lists the differences in plain words: three quote lines missing from the PO, two quantities changed, one item on the PO that wasn’t quoted.

Lines without a quote get matched against the customer’s own part numbers, your cross references and the item master. Each match shows how sure it is.

It checks units and prices before anything ships

“10 BX” against an item sold by the each raises a flag with the conversion spelled out. A PO price that differs from the quote or the customer’s agreement raises a flag with both numbers side by side. Nothing gets resolved silently.

It builds the order and holds it

The order goes into the ERP in a held status, with the flags attached. A rep opens it, reviews only the exceptions, fixes what needs fixing and releases it. Every correction the rep makes, like “this customer’s ‘BX’ means a box of 100,” gets saved so the same question doesn’t come up next week.

Run the numbers

There are two costs here: the keying time, and the mistakes that slip through.

The keying formula: POs per day × lines per PO × seconds per line ÷ 3,600 = hours per day.

Say, for the sake of the math, a supply house gets 60 emailed POs a day, averaging 12 lines each, and each line takes 45 seconds to find, translate, key and check. That’s 60 × 12 × 45 = 32,400 seconds, or 9 hours a day.

Over 250 working days that’s 2,250 hours. At a made-up loaded cost of $35 an hour, about $78,750 a year.

The error formula: orders per year × share with an entry error × cost of one mis-shipment.

Keeping the same made-up example: 60 a day × 250 days is 15,000 orders. Say 2 percent go out with an entry error. That’s 300 mis-shipments. Say each one costs $150 once you add the second delivery, restocking labor, the credit memo, the rebill and the occasional cut item you can’t return. That’s another $45,000 a year.

The AI version doesn’t take that to zero. Reviewing flags still takes time. But if review takes a third of what keying used to, two thirds of those hours go back to selling.

To run it for real, pull three numbers:

1. POs per day and lines per PO, from the ERP’s order history, filtered to orders entered by hand.
2. Seconds per line, timed. Sit next to a rep for an hour and count.
3. Credit memos coded to entry errors over the last 12 months, with what each one cost. If nobody codes them, that’s a finding too.

Picture a Monday, before and after

This is a made-up scene. Picture a two-branch plumbing supply house on a Monday after a busy Friday.

Before: 70 POs are spread across four inboxes. The two inside reps split them by guessing. One rep spends the morning on a 60 line PO for a school job, checking each line against a quote from three weeks ago. A PO texted to a counter person’s phone on Saturday doesn’t get entered until Tuesday. Two orders go out with the wrong ship-to.

After: the same 70 POs sit in one queue, already read and matched. Fifty-one have no flags; a rep checks each header and releases them. Nineteen have flags, and the flags say exactly what’s off: a missing line, a price that doesn’t match the quote, a new ship-to address nobody has seen before. The school job takes ten minutes, because the system already lined it up against the quote. The Saturday text is in the queue with everything else.

Same people, same ERP, minus the retyping.

What has to be in place first

• Quotes in the ERP with numbers customers actually cite. If quotes live in spreadsheets on a rep’s desktop, the PO has nothing to match against.
• Customer part numbers recorded. Every large customer has their own numbering. Capturing it once saves thousands of lookups.
• Clean units of measure. Each item needs its selling unit and its conversions written down: box of 100, case of 25, roll of 500 feet.
• One place POs land. A single order inbox, or a few with clear owners, so nothing gets missed.
• A way into the ERP. An import, an order API, or at worst a held order a person can complete in one screen.

Guardrails worth keeping

• A person releases every order. The system builds and holds. People release.
• Price differences are never settled automatically. Both prices are shown and a person decides.
• New ship-to addresses get confirmed. A first-time address can be a new job, or it can be a fraud attempt or a typo. Either way a person looks.
• Customer data stays inside the business. Use business-grade tools set so POs, pricing and customer lists aren’t used to train public models.
• Every change gets logged. Who changed what and why, so a disputed invoice can be traced back to the order.

How you’ll know it’s working

• Median minutes from PO received to order released.
• Share of orders released with no edits after review.
• Lines entered per rep per day.
• Credit memos coded to entry errors per month.
• Invoice disputes on price, which usually trace back to the order.

Where I’d start

Pick your ten biggest customers who send PDF POs that reference your quotes. Build only the PO-to-quote match for them: read the PO, tie each line to the quote, flag the differences. That’s narrow, it touches real money every day, and it tells you quickly how clean your quotes and units are.

If this sounds like your week, send me a note through the contact page.