Skip to content
All insights

Agentic AI

Extraction is the easy part of document processing.

Every document automation tool quotes an extraction accuracy figure. Almost none of them quote the number that decides the business case: how many documents complete without anyone touching them.

Published October 14, 20257 min read

What this covers

  • Field extraction from PDFs and scans is largely a solved problem. It is no longer where document projects fail.
  • The failure point is validation: reconciling what the document says against what your master data holds.
  • Straight-through rate, not extraction accuracy, is the number that determines whether headcount is freed.
  • Confidence has to be per field and act on the workflow, not a single score printed next to the document.

A purchase order arrives as a scan. It has a header, a table of eight lines, a delivery address and a set of terms at the bottom. A document processing tool reads it and returns a structured record: customer, PO number, eight line items with codes, quantities and prices, a requested date. It reports ninety-something percent field accuracy and the demonstration ends there, which is exactly where the interesting part starts.

Because the record is not the goal. The goal is a sales order in the system, correct, without anyone keying it. Between those two states sit questions the extraction never touched. Is that customer code ours or theirs? Is the price on line four the price this customer is entitled to? Does 'ea' on line six mean the same thing our catalog means by it? Is the delivery address one we have on file?

Three generations, and what changed

It helps to be clear about what actually improved, because the market talks about intelligent document processing as if the whole problem were new.

ApproachHow it readsWhere it fails
Template OCRFixed coordinates on a known layoutEvery new sender, and every layout change from an existing one
Trained extractionA model trained on labeled examples of each document typeDocument types with too few examples; drift as formats change
Agentic readingA general model reads the document the way a person does, in contextNot at reading. At deciding what the extracted values mean in your data

The third row is the real change. A model that reads a document it has never seen, in a layout nobody labeled, and understands that the column headed 'Menge' is a quantity and the handwritten note in the margin is a delivery instruction, removes the setup cost that made document automation a project rather than a feature. That is genuine progress and it is worth the attention.

Reading the document stopped being the bottleneck. Knowing what it means in your systems never was one, and now it is the only one left.

The gap between a field and a fact

Extraction gives you what the document says. Operations need to know what is true. The distance between those is where the work is.

  • The document says part 'AX440-SS'. Your catalog has AX-440S. Same item, different string, and no amount of extraction accuracy closes that.
  • The document says $98.40. The contract for this customer says $100.80. One of those is wrong and the agent has to decide which before anything is committed.
  • The document says 'qty 24'. Your catalog sells that item in boxes of six. Twenty-four pieces or twenty-four boxes changes the order by a factor of six.
  • The document has no customer code at all, only a letterhead, and the sender's email domain matches two accounts in the system.
  • The invoice has a freight line the purchase order never carried, within a tolerance nobody has written down.

None of these are reading errors. The extraction was perfect in every case. They are reconciliation problems, and they are solved with access to master data, customer history and business rules rather than with a better model.

What a working pipeline actually does

  1. 01Capture

    Take the document from wherever it arrives: a shared mailbox, an attachment, a document library, a scan. Including the email body, which frequently carries terms the attachment does not.

  2. 02Read

    Extract the content in context, including tables that break across pages, handwritten annotations and the difference between a header value and a line value.

  3. 03Resolve

    Match every extracted entity against master data: the sender to an account, each line to a catalog item, each unit to your unit of measure, the address to one on file.

  4. 04Validate

    Check the resolved record against the rules: pricing against contract, quantities against what this account orders, totals against the sum of lines, invoice lines against receipts and tolerances.

  5. 05Score

    Attach a confidence to each resolution, not to the document as a whole, so a single uncertain line can stop one line rather than the entire transaction.

  6. 06Route

    Complete what cleared. Hold what did not, presented to a person as a specific question with the candidates the agent was choosing between.

  7. 07Commit and log

    Write the result into the system of record and keep the full trail: the source document, what was read, what it was matched to, which rule fired.

The number to ask for

Extraction accuracy is a component metric. It is useful to an engineer and misleading to a buyer, because a document can be read perfectly and still require a person for every one of the reasons above.

The number that maps to the business case is the straight-through rate: the share of documents that complete, correctly, with no human touch. It is the only figure that converts into hours. And it is a property of your data as much as of the software, which is why any vendor quoting it without seeing your documents is quoting somebody else's operation.

Two supporting numbers make it interpretable. What share of held cases were held for a good reason, and what share of completed cases were later corrected. A high straight-through rate with a meaningful correction rate is worse than a lower one with none.

What improves the rate over time

A deployment should get better after it goes live, and the mechanism should be legible rather than mysterious. Three things move the number.

  1. 01Aliases. Every time a person resolves a customer part number to a catalog item, that mapping should be captured, so the same code from the same account resolves automatically next time.
  2. 02Rules. Cases held repeatedly for the same reason are usually a rule nobody had written down. Writing it down converts a recurring exception into a routine case.
  3. 03Master data. Held cases point at the specific records that are wrong or missing, which is a far more targeted cleanup than a general data quality project.

Notice what is not on that list: retraining a model on your documents. With agentic reading, the model does not need to learn your layouts. What the system learns is your data, and it learns it from the decisions your own team makes while working the held queue.

Where this shows up

Document processing is not a workflow. It is a capability inside several, and it is worth evaluating in the context of the transaction it feeds rather than on its own: inbound RFQs in quoting, purchase orders in order entry, supplier invoices in finance, and supplier confirmations in procurement.

Each of those ends in a different system with different rules, which is the argument against buying document processing as a standalone tool. Extraction that stops at a structured record hands you a new file to process. The value is in what happens after.

Questions

Questions we get about this.

Does it handle scans and photographs, not just digital PDFs?

Yes. Scanned documents, photographed pages and images embedded in email bodies are read the same way, including handwritten annotations in margins. Poor scan quality lowers confidence on the affected fields, which routes those lines to a person rather than guessing.

How many sample documents do you need before it works?

It is not a training exercise, so there is no minimum sample. What a deployment needs is access to the master data the documents have to be resolved against, and a decision on the rules and tolerances that define a correct outcome.

What accuracy should we expect?

Extraction accuracy is the wrong target. Ask for the straight-through rate, meaning the share of documents completed correctly with no human touch, and expect it to depend on how clean your master data is. Any figure quoted before someone has looked at your documents is a figure from another company's operation.

What happens to a document it cannot resolve?

It stops, on the specific line or field that failed, and goes to a person with the reason and the candidates it was choosing between. Nothing is committed on a guess, and the person's decision is captured so the same case resolves automatically next time.

Can it handle documents in other languages?

Reading a document in another language is not a separate capability for a general model, so mixed-language inbound mail is normal rather than a special case. What stays language-specific is your master data and rules, which is what the extracted values are resolved against.

See what this would look like in your operation.

A Workflow Review traces one of your real workflows end to end and names the first step an agent could take over. No platform to evaluate first.

Start with one workflow

Fixed scope for the first workflow, agreed before we start. Expansion is your decision once it proves value.

How it works

Security, stated plainly

Where the agents run, what they can and cannot do, and exactly what we do and do not claim.

Security