Gokan's Studio

1 – Introduction

Organizations, from small to very large, no matter which area they’re focused on, from military to B2B, today face increasing pressure to automate document processing and classification, and all of that in an intelligent way. Earlier generations of Microsoft document processing, commonly referred to as Syntex machine teaching models, were based on supervised learning. These solutions required users to label documents, train extraction or classification models, and maintain them as document layouts changed over time.

While effective for highly structured and stable document types, this approach introduced notable limitations: long setup times, continuous retraining, sensitivity to layout changes, and limited ability to reason beyond simple field extraction.

Microsoft’s modern document processing capabilities—SharePoint Autofill, AI Builder AI Prompts, and the Copilot Studio Document Processor Agent—are fundamentally different. They are LLM-first and prompt-driven, removing the traditional training phase entirely and replacing it with natural language instructions, contextual understanding, and reasoning across documents.

Moving from legacy Syntex models to these newer capabilities is not a one-to-one migration. In practice, organizations should retire trained extractors for semi-structured or frequently changing documents, preserve classic models only where strict determinism is required, and introduce Copilot Studio Agents when document processing must drive validation, decisions, or downstream actions.

By 2026, document processing is no longer about teaching a model what a document looks like. It is about enabling systems that understand what the document means and can act on that understanding.:

• Autofill with SharePoint
• AI Prompts with AI Builder
• Document Processor Agent with Copilot Studio

Each solution leverages generative AI and LLMs but differs in architecture, use cases, licensing, and capabilities. Here are the key findings of each solution as part of the introduction:

1. SharePoint Autofill excels at simple, self-service metadata extraction from SharePoint libraries, with minimal or no setup. Having the Copilot License as part of the Autofill in the Knowledge Agent or PAYG for SharePoint AI is enough.
2. AI Builder AI Prompts offers the most flexibility for custom AI workflows across the Power Platform ecosystem, with a wide range of connectors you can use, including SharePoint and Dataverse for storing your results.
3. Copilot Studio Document Processor Agent provides end-to-end orchestration for complex, multi-step document processing workflows.

2 – Technical Architecture Comparison

2.1 – SharePoint Autofill Columns

Powered by Microsoft Syntex (formerly part of SharePoint Premium, now called SharePoint Document Processing) and leveraging GPT-4 Turbo models via the Azure OpenAI Service, this feature is deeply integrated into the SharePoint document library experience.

Key Architectural Components

• LLM Processing: GPT-4 Turbo model for content analysis and extraction (Today, but can be outdated tomorrow!)
• Storage Layer: SharePoint list & library schema for metadata persistence.
• Processing Queue: Asynchronous document processing pipeline.
• Monitoring Interface: Autofill Activity panel for processing status tracking within the SharePoint Online interface.

Supported File Formats

• PDF (.pdf) + Microsoft Word (.doc, .docx) + Microsoft Excel (.xlsx, .csv) + Images (.png, .jpg)

Current Limitations

• Only library columns can use autofill (not site columns or content types as of November 2025 – no lists either).
• Maximum file size: 65 pages per document (only first 65 pages are processed and billed).
• Recommended limit: Up to 10 autofill columns per library.
• No reusable Autofill configuration

2.2 – AI Prompts with ai builder

Part of the Power Platform AI capabilities, AI Builder Prompts enables custom GPT-powered solutions consumable across Power Apps, Power Automate, Copilot Studio, and Dataverse low-code plug-ins.

Model Options and Efficiency

Model	Use Case	Token Limit	Cost Efficiency
GPT-4o mini	High-volume, simple tasks	16K tokens	Highest (Basic tier)
GPT-4o	Complex reasoning, document analysis	32K tokens	Medium (Standard tier)
GPT-o1	Reasoning-focused tasks	Variable	Lower (Premium tier)

Grounding Capabilities: AI Builder Prompts can be grounded with Dataverse tables, Connector data (SharePoint, SQL, …), and Document inputs.

Current Limitations

• Token and context limits: Large or highly unstructured documents must be chunked manually or via flows. This increases complexity.
• No native document lifecycle orchestration: AI Builder Prompts handle reasoning and extraction but do not provide built-in end-to-end document processing, validation stations, or autonomous decision-making without additional Power Automate or Copilot Studio components.
• Manual human-in-the-loop design: Human validation, approvals, and exception handling must be explicitly designed using Power Automate or Power Apps. There is no out-of-the-box review experience.
• Operational complexity at scale: High-volume document processing requires careful flow design, concurrency control, error handling, and monitoring. Without governance, solutions can become difficult to maintain.
• Cost predictability challenges: Token-based or credit-based consumption varies by model and prompt design, making cost forecasting more complex than page-based pricing.
• Prompt maintenance overhead: Changes in document structure or business rules require prompt updates, testing, and redeployment, often by makers or developers.
• Not SharePoint-native: While deeply integrated with the Power Platform, AI Builder Prompts are not a native SharePoint library feature and require flows or apps for integration.

2.3- Copilot Studio Document Processor Agent

This is a managed-agent solution—a pre-packaged, end-to-end document processing system that orchestrates the entire lifecycle. It is an autonomous agent operating with minimal human intervention.

Key Architectural Components

• Agent Core: Autonomous agent engine^.
• Document Extraction Flow: Processes documents with GPT-4o. GPT 5 chat, 5 auto, and 5 reasoning are available. You can also use Claude. (Today–this can change anytime!)
• Validation Station App: Canvas app (Power Apps) for human document review and editing.
• Dataverse Tables: For document tracking and metadata storage.

Key Differentiators

• Autonomous Operation.
• Human-in-the-Loop: Built-in validation station for quality assurance.
• End-to-End Orchestration: Handles the entire workflow from intake to export.
• Interactive Monitoring: Users can query the agent about document status via natural language.
• Scalability: Designed to handle over 1,000 documents with proper source storage configuration.
• Application Lifecycle Management (ALM): Supports versioning, deployment, and update management for document processing workflows.
• Agent365: Leverages the new Agent365 framework to unify AI agents across Microsoft 365 apps for cross-platform orchestration and enhanced collaboration.

Current Limitations

• Requires setup and configuration; not fully “plug-and-play” for small deployments.
• Higher cost for low-volume use due to Copilot Credits and message-based pricing.
• Dependent on proper Dataverse and source system configuration for scaling.
• Less flexible for one-off or very simple document extraction tasks.
• Requires some user training to manage agent monitoring and human-in-the-loop validation.

3 – Licensing and Cost Analysis

3.1 – SharePoint Autofill Columns

• Licensing Model: Pay-as-you-go (consumption-based) via Azure subscription or Part of the Knowledge Agent part of the Copilot offering.
• Pricing: $0.005 per page (as of March 2025).
• Key Points: Charged per page processed (up to 65 pages per document if PAYG).

3.2 – AI Prompts with AI Builder

• Licensing Model: Credit-based consumption (transitioning to Copilot Credits).
• Changes (as of November 2025): AI Builder features can now consume Copilot Credits. New customers must purchase Copilot Credits to run AI Builder features.
• Power Apps Premium ($20/user/month) and Power Automate Premium ($15/user/month) currently include a small monthly allocation of AI Builder service credits (around 500 for Power Apps Premium and 5,000 for Power Automate Premium). Historically, additional AI capacity was purchased via the AI Builder add-on (~1 million credits for ~$500/month). However, Microsoft has made a significant shift:
  • New customers are now required to purchase Copilot Credits for all new AI Builder capacity. The sale of the legacy AI Builder add-on ended for new customers on November 1, 2025.
  • Existing customers with active AI Builder licenses can continue to renew and purchase additional capacity using the old add-on until November 1, 2026.
• Crucially, the bundled, seeded AI Builder credits in Power Apps/Automate licenses will be removed entirely on November 1, 2026, for all customers.
• For AI Builder and document processing workloads in Power Platform, Copilot Credits are the intended, unified currency for consumption and billing starting in late 2026. This simplifies the licensing model by consolidating AI capacity across Power Apps, Power Automate, and Copilot Studio. Organizations should plan to transition fully to Copilot Credits before the November 2026 deadline to avoid a capacity gap.
• Cost: Variable based on model and token consumption

Model Tier	AI Builder Credits (Per 1,000 Tokens)	Copilot Credits (Per 1,000 Tokens)
Basic (GPT-4o mini, GPT-3.5)	Input: 1, Output: 3	Input: 0.1, Output: 0.1
Standard (GPT-4o)	Input: 20, Output: 60	Input: 1.5, Output: 1.5
Premium (GPT-o1)	Input: 140, Output: 560	Input: 10, Output: 10

3.3 – Copilot Studio Document Processor Agent

• Licensing Model: Copilot Credits consumption (message-based).
• Pricing: $200/month per pack (25,000 Copilot Credits) or pay-as-you-go via Azure.
• Microsoft 365 Copilot ($30/user – limited use rights)
• Cost: Message-based consumption

4 – Use Cases and When to Use Each Solution

4.1 – SharePoint Autofill Columns

Use Case	Why Autofill?
Simple Metadata Extraction in SharePoint (e.g., resumes)	Native SharePoint integration, no development required, self-service setup.
Invoice Processing (Single Library)	Pay-as-you-go pricing, automatic processing on upload.
Contract Management (extracting dates, terms)	Natural language prompts, no coding required.

When not to use: Multi-system integration required, complex approval workflows, or real-time processing needs. Bulk processing on existing documents.

4.2 – AI Prompts with AI Builder

Use Case	Why AI Builder?
Multi-Step Document Processing Workflows	Power Automate integration, connector ecosystem, flexible routing logic.
Custom Document Types with Complex Extraction	Custom prompt engineering, model selection (GPT-4o for quality), grounding.
Cross-Application AI Integration (e.g., Power Apps/Dataverse)	Power Apps integration via Power Fx, Dataverse grounding, real-time processing.
Customer Service Automation (sentiment, severity)	Sentiment analysis prebuilt functions, integration with ticketing systems.

When not to use: Simple SharePoint-only scenarios (Autofill is better), or when end-to-end orchestration is required (Copilot Studio Agent is better).

4.3 – Document Processor Agent with copilot studio

Use Case	Why Document Processor Agent?
End-to-End Invoice Processing with Validation	Pre-built orchestration, Validation Station for human review, export actions to ERP. If you want to know more about the Validation Station: Document Processing Agent for SharePoint AI people. – Gokan’s Studio
Loan Application Processing	Multi-document handling, complex validation workflows, and human-in-the-loop for compliance.
Insurance Claims Processing	Autonomous agent monitors new claims, extracts data from multiple document types, routes to adjusters.
HR Onboarding Document Processing	Structured workflow, validation station for HR review, export to HRIS.

When not to use: simple, one-step extraction (Autofill or AI Builder is cheaper), or for limited volume (< 50 documents/month).

5 – Decision Framework

5.1 – Selection Criteria Matrix

Criteria	Autofill	AI Prompts	Doc Processor
Ease of Setup	5/5	3/5	1/5
Flexibility	2/5	5/5	2/5
Cost Efficiency (Low Volume)	5/5	4/5	2/5
Multi-System Integration	1/5	5/5	5/5
Human Validation	1/5	3/5	5/5
Orchestration Capabilities	1/5	4/5	5/5

5.2 – Recommendation Guidelines

5.3 – Key Takeaways

To maximize results, organizations must adopt a hybrid document processing strategy that leverages the combined power of SharePoint Autofill, AI Builder AI Prompts, and the Copilot Studio Document Processor Agent. No single tool is a silver bullet; success lies in strategic integration.

• SharePoint Autofill: The Quick Win Specialist
  • Focus: Simple, SharePoint-native tasks and immediate ROI.
  • Strength: Ideal for extracting metadata from a document library or automating repetitive updates. It’s zero-code, fast to deploy, and provides instant value for teams already embedded in SharePoint. Use it for straightforward, library-specific automation.
• AI Builder AI Prompts: The Flexible Workflow Engine
  • Focus: Complex workflows, multi-source ingestion, and custom logic.
  • Strength: This is the low-code maker’s choice when documents originate from diverse sources, require multi-step reasoning, or need tight integration across the Power Platform. AI Builder orchestrates sophisticated logic, processes unstructured content, and connects with hundreds of data sources without requiring custom development. Use it to build flexible, integrated, and custom processing flows.
• Copilot Studio Document Processor Agent: The Mission-Critical Orchestrator
  • Focus: End-to-end governance, validation, and high-value processes.
  • Strength: Unmatched for mission-critical processes. The Agent autonomously manages document intake, triggers validated downstream actions, provides human-in-the-loop review, and monitors progress via natural language. It guarantees that high-risk processes—like invoice approvals, loan applications, or insurance claims—are handled with unwavering accuracy, auditing, and efficiency. Use it for crucial, governed, and autonomous process management.

Conclusion: This layered approach ensures comprehensive coverage: Autofill for rapid, simple automation; AI Builder for custom, complex workflows; and Copilot Studio Agents for governed, end-to-end orchestration. The future of document processing is not about selecting a single tool—it’s about building an integrated ecosystem where these three solutions work in concert to meet every level of organizational demand.

5.4 – Comparison Matrix Reference

Capability	Autofill	AI Prompts	Doc Processor
Primary use case	Metadata extraction in SharePoint	Custom AI workflows	End-to-end orchestration
Best For	Simple SharePoint scenarios	Multi-system integration	Complex validation workflows
Setup complexity	Low (No-code)	Medium (Low-code)	Medium (Low-code)
Development time	Minutes	Hours to days	Days to weeks
AI Model	GPT-4 Turbo (Fixed)	Multiple (GPT-3.5 to GPT-o1)	GPT-4o (Configurable)
Cost Model	Pay-as-you-go or Copilot	AI Builder Credits or Copilot Credits	Copilot Credits (message-based)
Typical monthly cost (1K docs)	$50 (10 pages/doc)	$9-150 (model dependent)	$18-50
Multi-System Integration	No	Yes (1000+ connectors)	Yes (1000+ connectors)
Human Validation	Manual post-processing	Custom approval flows	Built-in Validation Station
Batch Processing	Yes (Backfill)	Yes (Power Automate)	Yes (Designed for scale)
Real-time Processing	Async (Automatic)	Yes (Sync or Async)	Async (Agent-driven)
Max Document Size	65 pages	25 pages per prompt	Configurable (chunking)
Self-Service Setup	Yes	No (requires Power Platform)	No (requires Copilot Studio)
SharePoint Native	Yes	Via Power Automate	Via Dataverse/flows
Power Platform Integration	Limited	Full (Native)	Full (Agent flows)
Conversational Interface	No	No	Yes (Chat interface)
Orchestration Capability	Low	High (via flows)	Very High (Autonomous)
Ideal Volume Range	100-5,000 docs/month	Any volume (scalable)	500+ docs/month

5.4 – Decision Framework

Choosing the right AI document processing solution starts with identifying where your documents live and how complex the data extraction needs to be. For documents stored exclusively in SharePoint with simple requirements—specifically fewer than ten fields—SharePoint Autofill is the ideal choice, offering a highly affordable, no-code setup for rapid metadata extraction. However, if your SharePoint extraction is complex or requires human oversight, the decision splits: use Copilot Studio Agent when end-to-end validation is critical, or opt for AI Builder Prompts if you need custom workflows without mandatory human checks. When moving beyond SharePoint to handle multiple sources like emails or APIs, your operational goals dictate the tool; AI Builder Prompts are best for flexible integration across systems using Power Platform connectors, while Copilot Studio Agent shines in high-volume scenarios requiring autonomous orchestration and built-in monitoring.

What do our MVPs think about Intelligent Document Processing?

I reached out to a few friends in the SharePoint AI community — Mike, Joanne, Drew, and Sari — to gather their thoughts, ideas, and recommendations for IDP in 2026. They generously took the time to share their experience and perspectives, and I want to thank each of them wholeheartedly for their input. With this technical introduction and the insights they contributed, I genuinely hope this will be helpful and inspiring for all of you

Ready?

Let’s goooo!

1 – Which tool do you think fits best for document processing for 2026 ?

Joanne – MVP

SharePoint Autofill. I think most organizations have struggled for so long to get business users to tag metadata that this capability alone will be seen as a “quick win” to reduce/eliminate that struggle. However, the scalability of the autofill columns created will be a quick follow-up request from organizations looking for consistency across their sites. i.e., 1 autofill column should be able to be named and behave consistently across many sites.

I expect we’ll see significant adoption of the new Autofill column capabilities in SharePoint, largely because the feature is incredibly easy for end-users to configure and comes with a pricing model that’s far more flexible than traditional AI-powered document processing tools. Since Autofill is still in its early stages, we can anticipate rapid iterations and enhancements, especially considering that SharePoint remains the most widely used enterprise content and document management platform. Over the past few years, the broader content-management approach in Microsoft 365 has shifted away from centralized, IT-controlled repositories and more toward site-level and document-library–based governance. Autofill aligns perfectly with this evolution by empowering users to classify, extract, and structure information directly where the content lives, without needing to depend heavily on IT teams, custom solutions, or advanced technical expertise to get accurate and scalable results.

Copilot Studio Document Processor. Autofill is only part of document processing, which is content extraction.  While SharePoint Autofill is fantastic for quick wins, the industry is moving from “extracting data” to “acting on data.” A Copilot Studio Agent is the future because it doesn’t just read the document; it can orchestrate the lifecycle and perform tasks such as metadata and request validation. In 2026, we won’t just want metadata in a SharePoint column for accurate document processing; we will want an agent to say, “I processed this invoice, but the vendor date looks wrong. Should I email them?”

I think Copilot Studio Document Processor is the most versatile tool of them and can serve in many different cases.  

2 – What are your predictions for 2026 for Document Processing?

Joanne – MVP

Elevate the information architecture metadata components being created via document processing from a list level to a site level, or better yet, to a tenant level. Although having them at the list level does allow for better searchability, improved workflow processes, and provides additional context to Copilot, it still leaves some site- and tenant-wide capabilities, such as compliance controls, grappling to be leveraged in a scalable way.

I’m expecting to see fewer big updates to AI Builder going forward, and a lot more focus on Autofill and the Copilot Studio Document Processor. Autofill is still pretty new, but it’s already clear that Microsoft plans to grow it into a much more powerful, enterprise-ready feature. Over time, I can see it scaling across the whole tenant—similar to how the Content Type Hub works—so organizations can standardize and deploy it much more easily.

Copilot Studio’s Document Processor is also still in preview, and I think we’ll see it mature quickly. I’m expecting support for more metadata types, better extraction accuracy, and stronger, more flexible filtering across knowledge sources. As these improvements roll out, the Document Processor is likely to become a key tool for document understanding and automation in Microsoft 365.

It will become increasingly important as the amount of unstructured data grows exponentially, and it is poisoning the LLMs and discovery tools we want to use with too much information. This also means that the ability to process unstructured data has become easier with the tools discussed in this whitepaper, so makers can do it themselves rather than buying big tools and spending months to implement.

SharePoint Autofill and Knowledge Agent now support content types and metadata filling in existing site/library columns, so that organizations with established, unified metadata models can also benefit from them. 

3 – What are your recommendations for document processing, given the rise of Generative AI across all platforms?

Joanne – MVP

Having well-structured sites is still essential; although Gen AI can help sift through the clutter and extract valuable information, it does not replace the need for some level of consistency and structure across an organization, such as how sites and libraries are laid out. The more regulated an organization is, the more important this still is. Also, training staff on the different document processing options available so they understand “when to use what” will become very important.

With Generative AI becoming standard across Microsoft 365 and other platforms, the best way to approach document processing is to modernize your tools, clean up your content, and start small with pilots. First, move away from old-school, model-based OCR and explore newer options, such as SharePoint Autofill columns or Copilot Studio’s Document Processor. These tools use AI to understand documents, extract details, and automatically apply metadata—saving a ton of manual effort. Second, make sure your content foundation is solid. Clean up old libraries, set consistent metadata, and establish good governance. AI works best when it has organized, high-quality content to learn from. And finally, test things through small pilots before rolling them out broadly. Try the tools on a few document types, measure accuracy, and see where AI actually saves time. This helps you scale confidently.

Clean your house (data cleanup) before you invite the guests (AI doc processing). Audit your SharePoint permissions. If you turn on an AI Agent that can read everything, you might inadvertently surface sensitive data to the whole company. Start with SharePoint Autofill in a single library and build internal education on how it works. It’s the lowest barrier to entry. If the AI hallucinates there, it’s just a bad metadata column. If an Agent hallucinates, it might pay a fraudulent invoice. Focus on “Knowledge,” not just “Data.” Don’t just extract “Date: 12/01/2025.” Extract the context: “This contract expires soon and has unfavorable renewal terms.”

Organizations need to know what documents they have and ensure they delete outdated documents before expecting good results from Gen AI. Document processing tools can also help with that task. 

4 – Having Autofill part of the Copilot License was a huge deal – we clearly see that PAYG is left behind for core AI features in SharePoint – do you feel?

Joanne – MVP

I agree with this. The volume of unstructured content across organizations today is staggering, and being able to gain insights from documents so easily is extremely helpful. Organizations don’t have time for data cleanup, and business users don’t have time or see the value in manually tagging metadata, yet the benefits of doing so are evident. Tying these two capabilities together is excellent news.

I think this licensing model is best for Autofill. We are seeing higher adoption of the M365 Copilot, and this fits naturally within this licensing model. I am glad we are not complicating licensing more and adding a separate SKU for it (or making it an add-on). Making it available for users with the Copilot licensing model is a way to get the organization to adopt it quickly.

This is a double hit for cost. One, we need the Copilot license, but it still has PAYG costs with Autofill. This adds unnecessary mental friction, and people are terrified of a runaway bill if someone uploads 10,000 files.

I think it makes it easier for huge organizations to predict the costs of document management and stay within the budget. PAYG was always an issue in that sense. 

5 – How will the rise of Generative AI change the approach to handling unstructured documents by 2026?

Joanne – MVP

This will allow organizations to be less intentional in how they manage unstructured documents today for classification and metadata extraction. The wide variety of unstructured content that coexists in most SharePoint libraries today makes it extremely difficult to classify disparate content within one library correctly. Generative AI will allow this to happen relatively easily. Although the legacy Syntex models can classify documents with great precision, they require significant up-front work. They may require many models to address the variety of content that can coexist in a library. Also, not everything should require a Syntex model for it to be classified.

I see less focus on the legacy Syntex’s content extraction for unstructured documents. We have seen what Gen AI can do to summarize documents, index metadata, and generate metadata. The ease of use of Gen AI, which uses natural language to process documents, will gradually replace legacy systems.

It eliminates the “Training” phase. In legacy Syntex, you had to tag 5-10 documents to teach the model. With GenAI/Prompts, you describe what you want in natural language. The “Setup Time” drops from day to minute. This will also lead to less trust and more time spent deciphering the response’s result.

6 – Beyond data extraction, what is “one important capability” (reasoning, fraud detection, end-to-end workflow) that Intelligent Document Processing (IDP) tools should offer by 2026

Joanne – MVP

Compliance controls. Understanding the contents of the document, classifying it, and based on that, knowing the appropriate data protection and retention controls to apply in the moment. E.g., automatically apply a sensitivity label and retention label right then and there. Ideally, IDP tools should have insights into both the sensitivity label taxonomy and the retention label file plan and understand how to tie both back to a document’s classification. Although sensitivity labels are reasonably good at auto-applying based on a document’s content on the client-side, retention labels do not currently have that capability. This would be a game changer for orgs trying to implement retention labels at scale based on a document’s classification.

One important capability that Intelligent Document Processing (IDP) tools should offer by 2026 is reasoning—the ability to not just extract data, but also understand, interpret, and make contextual decisions based on that data.

For example, reasoning allows an IDP system to:

1. Identify inconsistencies or anomalies in documents that could indicate errors or fraud.
2. Infer missing information based on patterns or business rules.
3. Automate complex decisions in end-to-end workflows, reducing human intervention.

This goes beyond basic extraction, making IDP systems truly intelligent, enabling them to handle nuanced, unstructured documents at scale and integrate more deeply into enterprise processes.

Hope that helps!

Renewed Revolution