Email me

How do you stop keying data in from PDFs by hand?

Have software read the documents, and have a person review only the ones it isn’t sure about. AI can now pull names, dates, amounts and line items out of PDFs and scans. Most of the effort goes into checking what it read against records you already hold, so that mistakes get caught before they’re saved.

What can AI read reliably?

It reads typed documents with a consistent layout well. That covers most invoices, statements, forms and purchase orders, including PDFs whose tables fall apart when you try to copy them.

The documents differ by trade and the job is the same. A field service company keys in supplier invoices and field tickets, a builder or developer keys in draw requests and lien waivers, and a mineral company keys in revenue statements.

Scans, handwriting and unusual layouts are less reliable. AI might misread a smudged digit, lose track of a column that shifts partway down a page, or pick up a total from the wrong section, and the wrong answer looks as confident as the right one.

For that reason I spend more time on the software around the model than on the model. It records how certain the reading is, compares the result with other information, and shows a person the original page beside the value when something looks off. If a vendor demonstrates this on clean, typed documents, ask to see it run on your worst scans.

How do I know the numbers are right?

By checking every document against something else before it’s saved. Having a person look it over doesn’t do it. In a study of 195 people typing in data, researchers Barchard and Pace found that “visual checking resulted in 2958% more errors than double entry, and was not significantly better than single entry.”

I use three kinds of check. Arithmetic: the line items should sum to the subtotal, and subtotal plus tax should equal the total. Matching: a customer, well or account named in the document should correspond to one already in your records, and a close-but-different name gets flagged for a person to decide. History: an amount far out of line with previous months gets flagged too.

A document that fails any check goes to a person, with the page open and the problem marked. One that passes is saved without anyone touching it. The people who used to type now review the flagged ones, and you get a running list of every document that failed a check and why.

Where do you start?

I start by sitting with whoever does the keying. That person knows things no document shows: which senders change their layout every few months, which fields the business acts on, and what they check in their head before they trust a number. Those checks are the ones the software has to make.

I watch a normal batch go from the mail or the inbox to the system it ends up in, at your office in Weatherford or elsewhere in Parker County and west Fort Worth. I’m looking for where the pile waits and what happens to a document that doesn’t look right.

Then I build for the most common format first and add others once that one is working. The person who used to key reviews what the software flags, so the judgment they’ve built up stays in the job. If the figures get typed into a second system afterward, I connect the two so that step goes away as well.

Is this worth building for a small company?

It depends on how many hours go into it now. If someone spends two days a week typing from documents, that’s 40% of a salary each year. The work also piles up when they’re away, and month-end close can’t finish until they catch up.

Delay has a cost too. If figures are entered two weeks after the documents arrive, decisions in those two weeks are based on old numbers, and a bill that sat in the pile is harder to dispute.

If it only takes an hour a week, I’d leave it. I’ll tell you when I think something isn’t worth building.

The number of documents matters less than what they’re like. Twenty a month that take an hour each are worth automating. Two thousand that take ten seconds each may not be. The best candidates arrive on a regular schedule, come in a small number of formats, and feed a figure that someone acts on.

What does this look like in a mineral company?

In a mineral company the biggest pile is royalty revenue. Enverus notes that the check stub for a single low-volume well can exceed 300 pages, which can cost more to process than the well pays. The same article lists what goes unnoticed when nobody has time to check the statements: underreported volumes, wells missing from a statement, wrong decimals and deductions that shouldn’t have been taken. All of that is revenue the owner is owed.

The volume is why nobody has time. Enverus has estimated that interests in 1,000 wells across 40 operators produce 3,000 to 10,000 lines of revenue data. Some operators deliver it electronically, and many still send PDFs or paper.

Title is the second pile. Enverus says deals close in days while manual title takes weeks. Working through deeds, leases and probate records is slow and has to be right, and the buyer who can establish ownership sooner is in a better position to win the deal.

NEXT STEP

I have room for three more companies.

$2,000 to $6,000 a month, month to month. It starts with an hour at your office, at no charge. Show me how the work gets done, and I’ll tell you what I’d do next.

Email me to pick a day. gw@garrettwinder.com

CALL ME WHEN