DocCenter Design Considerations

Every DocCenter project involves balancing three factors: extraction accuracy, processing speed, and AI action consumption. Understanding how your DocCenter model configuration choices affect each factor helps you design a solution that meets your business requirements while staying within your AI action allocation.

Balance accuracy, speed, and consumption

Most teams prioritize accuracy first, then speed, then consumption. However, high consumption can become a blocker if it exhausts your AI action allocation before your DocCenter project delivers value. Discuss these tradeoffs with stakeholders early:

• Accuracy vs. consumption: Features like Advanced Layout can improve accuracy but increase consumption. Start with the simplest DocCenter configuration to assess your accuracy and enable additional features incrementally to hit your target.
• Speed vs. accuracy: Fast large language models (LLMs) such as Haiku and Nova 2 Lite process documents quickly and consume fewer actions per document, but may produce lower accuracy on complex documents. Reasoning LLMs such as Sonnet provide higher accuracy for complex extractions but take longer and may consume more actions.
• Volume vs. cost: High-volume processing magnifies per-document consumption. A configuration that adds a small cost per document can become significant at scale. Model the total monthly consumption before deploying DocCenter to production.

AI action consumption

AI actions are the unit of measurement for AI-powered feature usage in Appian. Each time DocCenter processes a document—whether classifying, extracting, or reviewing—it consumes AI actions. Your environment receives an allocation of AI actions based on your Appian license tier, and you can purchase additional bundles as needed.

Factors that affect consumption

Not all DocCenter model configurations consume AI actions equally. The following table summarizes the configurations that increase consumption and when to use them:

Configuration	Consumption impact	When to use
Advanced Layout	High—scales with page count, even blank pages.	As a last resort when Visual extraction alone does not achieve desired accuracy. See Choose between Visual extraction and Advanced Layout.
Mixed Vision + Text fields	High—makes multiple AI calls with the same document, doubling consumption.	Avoid this configuration. If your document has even one visual field, use Vision for all fields. If no fields need visual processing, use all text fields.
AI-Suggested Improvements	High—runs an AI agent that makes multiple calls per instance.	When you need AI-generated suggestions for improving model accuracy and accept the additional AI action cost.
AI Review	Moderate—runs a second model pass on every instance.	When straight-through processing (STP) is desired and you want automated verification before reconciliation.
Visual extraction	Moderate—processes page images in addition to text. Recommended configuration for documents with any visual elements.	For fields containing visual elements (checkboxes, signatures) or where spatial layout is important (complex tables, nonstandard layouts). Use Visual for all fields if any field requires it.
Bounding boxes	Low—makes an additional AI call to merge bounding box coordinates with extracted text.	When you need to highlight extracted values in the original document for reconciliation or audit purposes.
Field instructions	Low—adds context to the prompt for each field.	When the default extraction is not accurate enough and additional context improves results. Add only where they measurably help.
Describe Validations	Low—additional AI call per validation.	When you need AI-generated validation conditions rather than rule-based validations.
Text extraction	Low—processes only the text layer of each page.	For documents where all fields are standard text values (names, dates, amounts) and no visual elements need interpretation. This is the lowest-consumption configuration and the recommended starting point.
Number of fields	Cumulative—each field adds to the prompt size and response.	Extract only the fields your business process needs. Remove unused fields from the model.
Document page count	Cumulative—more pages means more content to process.	Use sections to limit what the model reads for long documents.

Manage consumption

To keep consumption predictable, follow the iterative model development strategy: start with text-only extraction, add features only when they demonstrably improve accuracy, and test at each stage before scaling. The key principles are:

• Start with the lowest-consumption configuration (text-only, fast LLM, minimal instructions) and add features incrementally.
• Extract only the fields your business process needs. Remove unused fields from the model.
• For classification models, set a page limit so the model reviews only the pages it needs to classify a document. For most documents, the first one or two pages contain enough context, so limiting pages on large documents reduces consumption without affecting accuracy. Configure this when creating a classification model.
• Monitor consumption over time using the DocCenter homepage. Watch for unexpected spikes after model changes or volume increases. You can also monitor AI actions at an instance level by comparing consumption across different configurations.

DocCenter usage groups for AI action attribution

A usage group is an Appian group that DocCenter attaches to its AI calls so that the resulting AI action consumption is recorded against that group. Usage groups change how AI action consumption is reported, not how much a model consumes. To reduce consumption rather than attribute it, see Factors that affect DocCenter AI action consumption.

Organizations that run DocCenter across several independent business units—for example, separate claims, underwriting, and compliance teams—can use usage groups for cost allocation, attributing AI action consumption to the team that generated it.

DocCenter attaches two kinds of usage groups:

• AIA All Users: DocCenter adds this group to every AI call automatically, which separates DocCenter consumption from other AI activity in your environment.
• Groups you specify: A calling process model passes these groups for a single execution, alongside AIA All Users rather than replacing it. Requires DocCenter 4.4 with Appian 26.6, and the groups must exist before you configure the process model.

Usage groups govern reporting only, and won't cap or throttle consumption. Usage groups record consumption after processing completes. Usage groups do not block processing when a business unit approaches a budget.

Extraction, classification, verification, and redaction models all support usage groups at runtime, so you can attribute their AI action consumption to a business unit. AI-Suggested Improvements does not currently support usage groups.

Where DocCenter records usage groups

DocCenter records usage groups in two locations:

Location	What it shows
The AI Usage tab of a model instance	Usage groups as tags beside the name of each AI call, such as Text Bounding Boxes or Extraction.
The artificial_intelligence.csv AI audit log	Usage groups for each AI call, for reporting outside DocCenter.

Choose between Visual extraction and Advanced Layout

Both Visual extraction and Advanced Layout improve accuracy for complex documents in DocCenter, but they work differently and have different cost profiles. Use the following guidance to choose the right option for your extraction model.

Start with Visual extraction

Enable Visual at the field level for any field that involves:

• Checkboxes or radio buttons: Fields that require interpreting graphical selection indicators.
• Signatures or stamps: Fields that require interpreting handwritten or image-based content.
• Complex tables: Tables with merged cells or irregular column structures.
• Nonstandard layouts: Fields where spatial position on the page determines meaning.

Visual extraction uses the multimodal capabilities of the large language model (LLM) to interpret page images alongside text. Visual extraction supports up to 500 pages (with reconciliation) or 1,000 pages (without reconciliation) with the latest models (Sonnet 4+, Nova 2 Lite) and provides strong accuracy for most visual elements at a moderate consumption cost.

Use Advanced Layout only as a last resort

Enable Advanced Layout at the model level only when:

• Visual extraction is insufficient: Visual extraction does not achieve your desired accuracy after prompt iteration.
• Prompt refinements are exhausted: You have tested prompt refinements and LLM selection changes without sufficient improvement.
• Higher consumption is acceptable: You accept the higher consumption as a necessary tradeoff for additional accuracy.

Advanced Layout combines Advanced IDP and generative AI processing, which significantly increases consumption—scaling directly with the number of pages in the document, including blank pages. Because of this cost profile, treat Advanced Layout as a last resort rather than a default configuration.

Plan for straight-through processing

If your goal is to minimize or eliminate manual review, design for straight-through processing (STP) from the start. STP means documents are processed end-to-end by AI with minimal human intervention, and exceptions are routed to reviewers only when needed.

The level of automation depends on your organization's risk tolerance—how comfortable stakeholders are with relying purely on AI versus requiring human verification. To implement STP, you combine confidence thresholds, AI Review, sampling, and validations as safety mechanisms.

Understand risk tolerance

Risk tolerance determines where your process sits on the automation spectrum:

Risk tolerance	Behavior	Recommended safety mechanisms
Low	All AI outputs are reviewed and reconciled by people.	Reconciliation task on every instance.
Moderate	Only uncertain or flagged outputs are reviewed by people.	Confidence threshold + AI Review + sampling.
High	AI outputs are never reviewed, enabling full STP.	Confidence threshold + validations + low sampling rate for anomaly detection.

Confidence threshold

AI models provide confidence scores at the field level to indicate how certain they are of their predictions. The confidence threshold defines the lower limit of what your organization is comfortable automating:

• If any field's confidence score falls below the threshold, the instance is routed for human review.
• If all scores meet or exceed the threshold, the instance is auto-approved.

From a process perspective, monitor the extraction status. If a document is flagged as Extracted with Errors, manual intervention is required.

AI Review

AI Review introduces a secondary validation layer that flags uncertain or incorrect extractions. Because this feature increases AI action consumption and generates items requiring manual oversight, use it strategically:

• Initial setup: Enable early to identify problematic fields and edge cases to create thorough test cases quickly.
• Strategic field selection: Prioritize AI Review for high-impact data such as total amounts. For predictable formats (like postal codes), use field validations instead.
• LLM selection: Iterate through available LLMs to find the model that achieves your goals for AI Review accuracy.

Sampling rate

The sampling rate selects a random percentage of documents for human validation, even when confidence scores pass the threshold. Use sampling as a quality control mechanism during STP:

• Start with higher sampling for new model iterations.
• Gradually reduce the rate as accuracy becomes consistent.
• Maintain a baseline level to detect anomalies over time.

To determine if a specific document has been selected for manual review, reference the isSamplingRequired field in the ExtractionInstance record.

Validations

Validations assess logic across document fields simultaneously. Unlike single-field verification, validations catch logical contradictions before downstream routing.

Balance two methods to maintain high precision while managing consumption:

• Expression rules: Use for programmatic checks that do not consume AI actions. Best for mathematical consistency (verifying that subtotals and tax equal the grand total), date sequencing, or detecting duplicates.
• Describe Validations: Use AI-powered validation only for scenarios that require reasoning or external context (such as geographic verification). Reserve this method for requirements that are difficult to express as rules.

If validations detect logical inconsistencies, DocCenter assigns statuses like Extracted with Errors or Instance Error alongside a validation message.

Handle unprocessable documents

DocCenter may encounter documents that it can't process due to encryption or data corruption. When this occurs, the extraction status changes to Instance Error, Extracted with Errors, or Classified with Errors, accompanied by a validation message stating the file contains unsupported content or may be corrupt.

Design your process model to handle these cases. Route the problematic file for manual review, or send a stakeholder notification and terminate the request.

Verify business requirements for handling unprocessable documents as part of your initial project scoping, since this applies to both classification and extraction workflows.

Feedback