# Intelligent Document Processing AI: From OCR to Judgment

> Intelligent document processing AI with agents: extracting invoices and delivery notes, validating against systems, and designing the confidence threshold.

- Canonical: https://arkatai.com/en/intelligent-document-processing/
- Site: Arkatai (https://arkatai.com) — agentic operations as a service
- Language: en
- Published: 2026-07-19

---


Intelligent document processing is the ability to read invoices, delivery notes, contracts and emails arriving in any format, extract their data, validate it against your systems, and decide what to do with each one. The word that matters is "decide". For twenty years the problem was framed as OCR: turning an image into text. That is solved today, and it is not what clogs an operation. What clogs it is the judgment: looking at an invoice, understanding what it is, checking whether it matches what is in the ERP, and knowing whether it gets recorded straight through or goes to review.

In the operations I have seen up close, the bottleneck was never reading the document. It was the judgment that comes after. A person picks up an invoice, looks at it, compares it against the order, spots that two lines are missing, decides whether it is a supplier error or a partial delivery, and acts. That work —not the scanning— is what eats hours and what an agent can start to execute, within limits you set.

## OCR is the beginning, not the job

It pays to separate two layers the market blurs. OCR extracts text from an image. Intelligent document processing does what a clerk does with that text: it understands which document it is, pulls the fields that matter even when they sit in different places for each supplier, cross-checks them against your systems, and decides.

The difference shows in heterogeneity. An old extraction model needs a template per document type: if the supplier changes the format, it breaks. An agent with a language model behind it needs no fixed template. It interprets a document it has never seen, just as a person who knows what an invoice is would, even with a new layout. That tolerance for variation is exactly what separates an agent from a rigid flow, a difference I develop in [AI agents in operations](/en/ai-agents-in-operations/).

## The four steps: extract, understand, validate, decide

A well-processed document goes through four steps, and only the first is OCR.

**Extract.** The agent reads the document —PDF, photo of a delivery note, body of an email, scanned attachment— and pulls the fields: issuer, number, dates, lines, amounts, taxes. Here heterogeneity is the norm, not the exception.

**Understand.** It classifies what it is: invoice, credit note, delivery note, order, contract, complaint. Processing a credit note is not the same as processing an invoice, and the same email can carry both. This step decides which circuit the document enters.

**Validate.** It cross-checks the data against your systems: does the order this invoice cites exist? Does the amount match what was agreed? Does the delivery note match what shipped? Is the supplier registered? Validation is where the document stops being text and becomes a verified fact or a discrepancy. This requires the agent to have read access to the right data, and I have written about what data an agent needs to work in [data requirements for AI agents](/en/data-requirements-ai-agents/).

**Decide.** With the document understood and validated, the agent decides: record it straight through, send it to an approver, flag it as a discrepancy, or request human review. This decision is the heart of the system, and where its usefulness or danger is settled.

When the document is a supplier invoice or an invoice issued to a customer, this circuit connects to the full revenue cycle, and I cover it from that angle in [order to cash automation with agents](/en/order-to-cash-ai-automation/), where the delivery-note-to-invoice match is one of the boundaries that leak the most margin.

## The confidence threshold: when it records straight through and when it asks for review

Here is the design decision that separates a serious system from a demo. Every extraction and every validation carries a confidence level. Designing the system means deciding, for each document type and each amount, above which confidence the agent records straight through and below which it asks for review.

This is not a technical decision. It is a business decision about how much risk you accept per document. A twelve-euro credit note with every field validated does not need a person to look at it. A fifty-thousand-euro invoice from a new supplier with a line that does not match the order goes to review even if the model is reasonably confident. The threshold is calibrated by the harm of being wrong, not by the model's average confidence.

In practice I set it in three bands:

| Band | Condition | What the agent does |
|---|---|---|
| Green | High confidence, full validation, low amount | Records straight through and leaves a trace |
| Amber | Medium confidence or minor discrepancy | Prepares the record and hands it to a reviewer with context ready |
| Red | Low confidence, new supplier, or high amount | Escalates to a person without acting |

The goal is not to maximize the share that goes green on day one. It is to start with a prudent threshold —more amber than you would like— and move it toward green as the data shows the agent is right on that document type. I develop that logic of human review before releasing control in [human in the loop](/en/human-in-the-loop/), and how to measure whether the agent has earned a higher threshold in [AI agent evaluation](/en/ai-agent-evaluation/).

## The controls that get this into production

An agent recording documents in your systems needs the same discipline as any operating agent: concrete permissions on what it may write and up to what amount, and traceability of every decision. The trace has extra value here: faced with a discrepancy found weeks later, you have to be able to reconstruct which document came in, which fields were extracted, what it was validated against, and why it was recorded. Without that trace no audit is possible. I cover permissions in [AI agent permissions and controls](/en/ai-agent-permissions-controls/).

## Why you cannot buy it off the shelf

Products exist that promise document processing out of the box, and for the standard case they work. The problem shows with your documents: the supplier who invoices in an impossible format, the rule that a certain customer's delivery note is not accepted without an order reference, the contract with a clause that changes how the invoice is validated. That set of rules and exceptions is yours, and encoding it is the real work. On why that specific part cannot be bought off the shelf, even when the engine can, I have written in [the custom phase manifesto](/en/the-custom-phase/). If you want the general frame of what an agent does inside a company, the starting point is [AI agents for business](/en/ai-agents-for-business/).

## Frequently Asked Questions

### How is intelligent document processing different from OCR?

OCR turns an image into text and stops there. Intelligent document processing understands which document it is, extracts the fields that matter even when each supplier places them differently, validates them against your systems, and decides what to do with each one. OCR is the first of four steps. The value is in the other three.

### What document types can an agent process?

Invoices, credit notes, delivery notes, orders, contracts, and emails with instructions or attachments. An agent's advantage over a template extractor is that it tolerates format variation between suppliers without breaking, because it interprets the document rather than looking for fields in fixed positions.

### How is the confidence threshold decided?

By the harm of being wrong for each document type and amount, not by the model's average confidence. A small, validated credit note records straight through. A high-value invoice from a new supplier goes to review even if the model is confident. You start with a prudent threshold and loosen it as the data proves accuracy.

### Does the agent record invoices without anyone looking at them?

Only the ones that fall into the high-confidence, low-risk band you have defined, and always leaving a trace. The rest are prepared for a reviewer or escalated. The system does not remove human review. It concentrates it where it genuinely adds value and withdraws it where it was only a formality.