For years, accounts payable automation started with the same problem: invoices arrived as PDFs, scans, images, and email attachments, but the ERP needed structured data.

OCR helped close that gap by converting document images into machine-readable text. But recognizing the characters on an invoice is not the same as understanding what those characters mean.

An AP system still needs to determine which number is the invoice number, which date is the due date, which company is the supplier, where the line items begin, and whether the information on the document agrees with the purchase order, contract, supplier record, and other data already in the business.

Traditionally, much of that interpretation depended on templates, mappings, and rules. Modern invoice AI changes that model by using document context, layout, language, and other signals to extract structured invoice information without requiring a supplier-specific template for every format.

That difference matters most when invoice formats vary.

Table of contents

• What OCR does, precisely
• What invoice AI does differently
• Side by side comparison
• Why templates stall around half your invoice volume
• How template free extraction works
• What “AI-powered invoice processing” can mean
• Where OCR is still useful
• What to expect when moving from templates to invoice AI
• How Relish Invoice AI approaches invoice extraction
• Frequently asked questions

What OCR does, precisely

Optical character recognition, or OCR, converts text contained in an image or scanned document into machine-readable characters.

Give an OCR engine a scanned invoice and it can identify the words, numbers, and other characters printed on the page. Depending on the technology, it may also return information about where those characters appeared.

That is valuable, but it is only the first part of invoice processing.

The system still needs to determine what the information represents.

For example:

• Is 125487 an invoice number, purchase order number, supplier number, or customer account?
• Is $48,321.10 the subtotal, tax amount, invoice total, or outstanding balance?
• Does “March 15” represent the invoice date, delivery date, or payment due date?

Historically, invoice-capture systems solved that interpretation problem in a few different ways.

Template-based or zonal OCR

With template-based capture, a document format is mapped in advance.

A system might be told that the invoice number normally appears in the upper-right portion of a particular supplier’s invoice, while the invoice total appears near the bottom. When another invoice matching that format arrives, OCR reads those predefined areas.

For stable, repetitive documents, this approach can work very well.

The challenge comes when document variability increases.

A supplier may redesign its invoice. A field may move. An invoice that usually has one page may suddenly have three. Line-item tables may contain different numbers of rows. A scan may be rotated or distorted. A new supplier may use a completely different format.

The more an extraction process depends on a fixed layout, the more those variations can create maintenance work or exceptions.

Full-text OCR with rules

Another approach is to recognize all of the text on a document and use rules or patterns to locate important values.

For example, a system might look for a date near the words “Invoice Date” or search for a currency value near “Total Due.”

That can reduce dependence on fixed zones, but the rules themselves still need to accommodate the many ways suppliers format and describe the same information.

The result is an important distinction:

OCR recognizes the text. Something else still has to interpret the invoice.

What invoice AI does differently

Modern invoice AI focuses on the interpretation layer.

Instead of relying exclusively on a fixed supplier template, AI-based document processing can evaluate multiple signals together, including the document’s text, layout, structure, context, and, in some systems, the visual representation of the document itself.

Consider how a person reads an unfamiliar invoice.

They do not need to know that a specific supplier always places the invoice number 1.5 inches from the upper-right corner. They understand what an invoice number looks like, recognize labels and surrounding context, and distinguish it from other numbers on the page.

AI-driven extraction is designed to make a similar type of contextual determination at scale.

That creates several practical differences.

New invoice formats require less supplier-specific setup

Template-free systems are designed to process variable invoice structures without requiring a new zonal template for each supplier.

This is particularly important for enterprises with a large supplier base or a significant long tail of suppliers that submit relatively few invoices each.

Layout changes become easier to absorb

Moving a field on the page should not automatically require rebuilding an extraction template if the system can interpret the information from its context.

That does not mean every layout change will process perfectly. It means extraction is less dependent on a particular field remaining in a particular coordinate.

Complex documents become more addressable

Images, scanned documents, unusual page layouts, handwriting, variable line-item tables, multiple attachments, and other invoice variations have historically created challenges for rigid capture systems.

AI-based extraction gives AP teams another way to process those documents without creating a unique template for every variation.

Feedback can improve future processing

Some AI-driven systems can use reviewer corrections and exception resolution as feedback for future invoices.

That creates a fundamentally different operating model from maintaining a library of static templates.

The important evaluation question is not simply whether a vendor says its system “learns.” Ask exactly how corrections affect future processing and whether additional configuration is still required.

Side by side comparison

Capability	Traditional template-based capture	AI-based invoice extraction
New supplier	May require supplier-specific setup	Can process without a supplier-specific template
Layout changes	May require template adjustment	Designed to adapt to layout variation
Fixed-format documents	Strong fit	Strong fit
Variable invoice formats	More configuration-dependent	Designed for higher variability
Complex line-item tables	Can require additional configuration	Can interpret variable table structures
Images and scans	OCR-dependent	OCR/document vision plus contextual interpretation
Handwritten content	More challenging	More addressable, depending on quality and model
Corrections	Often require rules/template updates	May feed future processing or learning
Supplier setup effort	Can grow with supplier count	Less dependent on supplier-specific configuration
Ongoing effort	Template and rule maintenance	Exception review, business rules, and model oversight

Why templates stall around half your invoice volume

The pattern is consistent enough across enterprises to be predictable, and it comes from supplier distribution.

Invoice volume follows a long tail. A minority of suppliers send the large majority of invoices, and a very large number of suppliers send a handful each. Templates are economical for the first group. Building and maintaining a template for a supplier that sends four invoices a year costs more than typing those four invoices in.

So the rational implementation builds templates for the high volume suppliers, captures a substantial share of documents, and stops. The remaining invoices, spread across hundreds or thousands of low volume suppliers, stay manual forever. They are also disproportionately the messy ones, since large suppliers tend to have clean standardized invoice output and small ones do not.

This is why so many AP teams describe their automation as working and still have people doing data entry every day. Both things are true. The template covered slice is genuinely automated. The tail was never in scope, and the tail is where the labor is.

Template free extraction changes the economics because the cost of adding a supplier is close to zero. There is nothing to build. The tail becomes addressable for the first time, and the tail is where the remaining savings live.

How template free extraction works

The exact architecture varies by platform, but modern invoice automation generally follows four stages.

1. Ingest the invoice

Invoices may arrive through email, PDF attachments, scans, images, procurement networks, e-invoicing systems, or other channels.

Structured electronic invoices may already contain machine-readable data and require little or no document extraction.

Unstructured documents need to be interpreted.

2. Extract and structure the information

The extraction system identifies fields such as:

• Supplier
• Invoice number
• Invoice date
• Purchase order number
• Currency
• Tax
• Subtotal
• Total
• Line-item descriptions
• Quantity
• Unit price

Rather than relying entirely on the coordinates of those fields, template-free systems can use contextual information to determine what each value represents.

3. Validate the extracted data

Extraction accuracy alone is not enough.

A value can be read correctly from an invoice and still be wrong from a business perspective.

The supplier should correspond with supplier data. The purchase order should exist. Invoice lines may need to match PO or contract information. Totals need to reconcile. Duplicate invoices need to be identified. Other business or compliance rules may also apply.

This is where invoice automation begins moving beyond document capture.

Extraction determines what the invoice says. Validation determines whether the information can be trusted and processed.

4. Route exceptions appropriately

Invoices that satisfy configured extraction, validation, and matching requirements can continue through the process.

Those that do not can be routed for review.

A useful exception experience should tell the reviewer what needs attention and give them enough context to resolve it efficiently.

The objective is not eliminating humans from AP entirely. It is reducing the number of invoices that require human intervention simply to read, rekey, or reconcile routine information.

What “AI-powered invoice processing” can mean

Nearly every invoice-automation vendor now uses AI terminology.

That makes the label itself less useful than the behavior of the product.

When evaluating invoice AI, ask:

Can it process an invoice from a new supplier without supplier-specific setup?

If each new supplier requires a template, mapping, or significant configuration before invoices can be extracted, understand how that affects the effort required to automate your full supplier population.

What happens when the layout changes?

Ask the vendor to show how the system behaves when fields move, line counts change, or an invoice arrives in a different format.

What happens after a user corrects an error?

Does the correction influence future processing? Does someone need to create a rule? Is model training involved? How does the system prevent the same issue from appearing repeatedly?

How are unfamiliar documents handled?

The answer should include both extraction and exception handling. No system should simply assume every value is correct.

How much of our invoice population can actually be automated?

Ask for results from organizations with similar invoice volumes, supplier diversity, ERP environments, and document complexity.

And ask how the vendor defines “touchless.”

A percentage without a denominator tells you very little.

Where OCR is still useful

OCR is not obsolete.

For highly standardized documents, template-based approaches can still be efficient, predictable, and cost-effective.

If an organization processes enormous volumes of the same document format, there may be little reason to replace a workflow that already performs reliably.

OCR is also frequently part of modern AI document processing. A scanned invoice still needs to be converted into information that software can interpret.

The better question is therefore not:

OCR or AI?

It is:

How much does the system depend on fixed templates and manual configuration to understand the document after the text has been recognized?

What to expect when moving from templates to invoice AI

The largest improvement may not come from invoices your existing system already handles well.

It often comes from expanding automation to documents that previously required manual processing because of supplier or format variability.

Exception handling also does not disappear.

Instead, the objective is to shift human effort away from basic transcription and toward situations that actually require judgment: missing purchase orders, pricing discrepancies, receiving issues, supplier mismatches, unusual transactions, or other business exceptions.

That distinction matters.

A successful AP automation program should not be measured by whether a human ever touches an invoice.

It should be measured by whether people are spending time on work that actually requires a person.

How Relish Invoice AI approaches invoice extraction

Relish Invoice AI is designed to extract and validate invoice data within existing AP workflows before that data enters downstream systems.

Rather than requiring an OCR template or model training for each supplier format, Invoice AI extracts and structures data from variable invoice formats and then validates that information against purchase orders, contracts, supplier data, and other enterprise records.

Template-free extraction

Relish Invoice AI does not require supplier-specific OCR templates or model training. It is designed to handle complex and changing invoice structures, including PDFs, email-based documents, scans, and other invoice formats.

Structured validation and matching

Extraction is followed by validation against the business data that determines whether an invoice is ready to move forward.

Invoice AI can validate invoice information against purchase orders, contracts, and vendor reference data to identify discrepancies before incorrect data reaches downstream processing.

Adaptive intelligence

Invoice Assistant provides conversational AI capabilities for working with complex invoices, while corrections and rules can be used to improve the handling of future invoices and recurring scenarios. Relish’s current product materials describe this as “adjust once” functionality that allows Invoice Assistant to apply rules to future invoices.

Rule-based automation and exception handling

Organizations can incorporate business rules and conditional logic into processing so invoices that meet requirements can continue while exceptions receive additional attention.

This brings extraction, validation, and exception handling into the same broader process rather than treating invoice capture as a standalone data-entry step.

Global invoice processing

Relish Invoice AI supports more than 300 languages, provides Peppol access-point capabilities, and supports government-mandated invoicing processes or validations across more than 30 countries.

For Denny’s, Relish reports an increase from approximately 50-60% electronic invoicing to 75-80% touchless processing.

The larger point is that Invoice AI does more than convert an invoice into fields.

It combines:

Invoice intake → AI extraction → structured validation → exception handling → clean ERP entry.

That sequence aligns with Relish’s broader approach: validate data before it becomes a downstream problem.

Frequently asked questions

What is the difference between OCR and invoice AI?

OCR recognizes characters contained in a scanned or image-based document. Invoice AI goes further by interpreting the document and identifying what the information represents.

Traditional invoice-capture systems often combine OCR with templates or rules that determine where fields such as invoice number, date, total, and line items should appear. Template-free invoice AI can instead use text, layout, context, and document structure to identify those fields across different invoice formats.

OCR may still be one component of an AI invoice-processing system.

How can invoice data be extracted without OCR templates?

Template-free document AI interprets invoice content based on the document as a whole rather than relying on fixed zones configured for each supplier.

The system identifies fields such as invoice number, date, supplier, tax, totals, and line items and turns them into structured data. That information can then be validated against supplier, PO, contract, and other enterprise data.

Is invoice AI more accurate than OCR?

It depends on what is being measured.

A well-configured template can perform very accurately on the document format it was designed to process.

The advantage of AI-based extraction is adaptability and coverage. It can reduce dependence on maintaining a template for every supplier and can better accommodate unfamiliar or changing invoice formats.

For AP teams, the more useful metric is often not accuracy on one controlled document type but the percentage of total invoice volume that can be processed reliably without unnecessary manual intervention.

Why do template-based invoice automation projects leave manual work behind?

Template-based approaches work best when document structures are predictable.

As supplier counts and invoice variability grow, organizations may need to build and maintain additional templates and rules. Low-volume suppliers can be especially difficult to justify because the effort required to configure a template may outweigh the labor saved.

Template-free extraction is designed to reduce that supplier-by-supplier setup requirement.

Can AI process handwritten invoice information?

Modern document AI can process some handwritten and annotated content, although performance depends on handwriting quality, document quality, language, and the specific technology.

Relish Invoice AI specifically includes handwriting among the invoice formats it is designed to process.

Lower-confidence or ambiguous information should still be routed for review rather than assumed to be correct.

Does invoice AI replace accounts payable teams?

Invoice AI is primarily designed to automate repetitive invoice-processing work.

Routine invoices can require less manual transcription and reconciliation, while unusual invoices and business exceptions can still require AP judgment.

That changes where people spend their time: away from rekeying information and toward resolving exceptions, managing suppliers, maintaining controls, and improving the overall AP process.

What is template-free invoice extraction?

Template-free invoice extraction means the system does not require a fixed supplier-specific document layout to identify invoice fields.

Instead, the system interprets the invoice using contextual information such as text, layout, labels, relationships between fields, and document structure.

This allows organizations to process a wider variety of invoice formats without creating and maintaining a separate template for every supplier.

How long does AI invoice extraction take to implement?

There is no universal implementation timeline.

Template-free extraction can remove the work of building a large supplier-specific template library, but an enterprise implementation may still require ERP integration, invoice-channel configuration, business rules, validation logic, testing, security review, workflow decisions, and change management.

When evaluating vendors, ask for implementation timelines from customers with a similar ERP environment, invoice volume, number of business entities, and geographic footprint.