By Micah Smith, VP, Developer Relations, Community & Learning, Automation Anywhere

Every enterprise I’ve worked with has shared a version of the same frustration. Somewhere in their workflows sits a mountain of unstructured documents: invoices, enrollment forms, contracts, claims, applications, medical records, compliance reports. They are messy, inconsistent, and critical to business operations.

For years, enterprises tried to automate these processes with template-based extraction or custom-trained machine learning models. Those approaches delivered incremental gains, but they came at a cost. Models had to be retrained for every new format and form version. Complex layouts confused template engines. Edge cases piled up until humans had to step in. And don’t even get me started on the use cases involving unstructured forms.

The result was a cycle of partial automation followed by manual correction — a source of cost and delay that scaled with the business rather than shrinking.

The rise of vision models and large language models (LLMs) has completely shifted what is possible. Suddenly, the most challenging document types — free-form, unstructured documents such as contracts, handwritten notes, and complex financial statements — are no longer off-limits. At Automation Anywhere, we saw this shift early and built it directly into Document Automation, powered by the Process Reasoning Engine (PRE).

Why Traditional Approaches Fell Short

Template-based methods worked well for structured documents like invoices in fixed layouts. As soon as variability increased — different suppliers, new formats, multilingual content — accuracy collapsed.

Machine learning models improved flexibility, but they required large amounts of labeled data and document variations. Training was expensive and time-consuming. A single layout change could force a retraining cycle. For many organizations, this overhead outweighed the benefits.

These limitations created a ceiling. Enterprises automated what they could, but they never touched the hardest, most valuable document types. Those remained in the hands of employees, consuming hours of repetitive work.

Enter Vision and Language Models

Vision models changed the equation by enabling systems to interpret images and layouts in ways that go far beyond OCR and key-value pair data relationships. These vision models, combined with language models, enable the extraction of meaning from documents where the structure is irregular or unpredictable.

At Automation Anywhere, we recognized this early and introduced generative extraction into our document automation product. Today, nearly 80 percent of pages processed in our system use this capability. That number continues to grow, and for good reason.

Vision models can understand a handwritten note in the margin of a contract. They can interpret tables where columns don’t align perfectly. They can handle documents scanned at odd angles, or PDFs cobbled together from multiple sources. Paired with LLMs, they can even interpret context, such as identifying which date on a contract is the start date versus the renewal date.

This leap makes extraction possible and practical.

The PRE Advantage

Extraction alone is not enough. Enterprises need accuracy, context, and learning. That is where PRE amplifies the power of vision and language models.

PRE doesn’t simply pass raw, OCR’d text to a model. It preprocesses documents by classifying them, enhancing image quality, understanding layout structures, and applying a proprietary chunking algorithm. It adds context from past interactions and customer-specific rules. It integrates feedback so that corrections feed forward into the next run.

Imagine a claims form with a scanned attachment and a handwritten correction. PRE parses the layout, extracts the key fields, recognizes the handwritten note as a correction, and validates the final data against the enterprise system of record. If a human intervenes to adjust the outcome, that correction becomes part of the agent’s knowledge going forward.

This layered approach produces accuracy that generic prompts can’t achieve. It also scales — across thousands or millions of documents — without requiring the constant retraining that legacy ML approaches demanded.

A Real Example: Invoice Reconciliation

Consider the process of reconciling invoices received by email.

A customer sends an invoice as an attachment and includes clarifying notes in the body of the email. Historically, this would require manual review. A template engine could extract the invoice fields, but it would ignore the email context. A machine learning model might classify the document correctly, but it wouldn’t compare the two sources.

With PRE, vision models, and an agentic workflow, the process looks very different. The system extracts data from both the invoice and the email, compares them, and flags discrepancies. It validates supplier IDs against the enterprise system, applies any corrections from the email, and finalizes the record automatically. Only if ambiguity remains does it escalate to a human.

The enterprise gains speed, accuracy, and resilience. The employee gains freedom from repetitive review work.

The Learning Loop

One of the most powerful aspects of PRE is how it incorporates learning into document automation.

Every correction — whether from a human, an eval, or a downstream system — feeds back into PRE. Over time, the system adapts to the specific quirks of each enterprise. Supplier A always formats dates a certain way. Legal contracts from Region B often include an extra clause. PRE learns these nuances and integrates them into future executions.

This loop transforms automation from a static set of rules into a dynamic system that improves continuously.

Governance Still Matters

As with all agentic automation, governance plays a central role in document processing. Enterprises handle sensitive data: contracts with confidential terms, invoices with bank details, medical records with personal identifiers.

Automation Anywhere’s Enterprise Control Room ensures that role-based access controls define who and what can see this data. Every action — from extraction to validation to delivery — can logged for auditability. Accuracy matters, but trust matters more. Document automation must meet the same bar as any other enterprise system when it comes to security and compliance.

The Broader Impact

The significance of this shift goes beyond efficiency. Documents are the connective tissue of many important business transactions - especially when signatures are required. They carry contracts that govern relationships, invoices that move money, and records that underpin compliance. When documents remain manual, they slow down the entire business. When documents become part of intelligent automation, they accelerate it.

By reimagining document processing with vision models and PRE, enterprises unlock processes that were previously too costly or complex to automate. Claims processing, supplier onboarding, contract management — each becomes faster, more accurate, and more resilient.

Plus, employees will no longer spend hours correcting OCR errors or hunting through PDFs for a single clause. They focus on exceptions, strategic analysis, and higher-value work.

What’s Coming Next

The next wave of innovation in intelligent document processing will come from deeper integration of multimodal models, capable of reasoning across text, images, tables, and even audio or video. PRE will continue to provide the reasoning and orchestration that translate those capabilities into enterprise outcomes.

I expect to see more on-demand document agents, composed dynamically to handle new document types without weeks of setup. I also expect stronger evaluation frameworks, ensuring accuracy remains high even as document diversity grows.

Most importantly, I see document automation moving from a back-office function to a front-line enabler. As accuracy and reliability increase, enterprises will trust automated systems with their most critical records, from contracts and legal documents to compliance filings.

Unlocking What Was Out of Reach

For years, enterprises lived with the frustration of unstructured documents that resisted automation. Vision models, combined with the Process Reasoning Engine, have broken through that barrier.

Document automation now handles the messy, the complex, and the high-value — not by replacing human judgment, but by amplifying it. Enterprises move faster. Employees focus higher. Customers see better outcomes.

The ceiling has lifted. Documents are no longer a roadblock. They are fuel for intelligent automation.