AI Pipeline Accuracy Regression Goes Undetected After Deployment: Causes and Fixes
Accuracy Regression Undetected — A Bad Deploy Ships Silently
Document processing AI failures — extraction, classification, and verification mistakes
46 patterns
Document processing AI failures — extraction, classification, and verification mistakes. We've tracked 46 distinct Document Processing AI patterns across production AI agent deployments — most commonly in Document Processing systems. One recurring example: AI Pipeline Accuracy Regression Goes Undetected After Deployment: Causes and Fixes — Accuracy Regression Undetected — A Bad Deploy Ships Silently
Each pattern below documents the specific failure mode, why it happens, and how teams have mitigated it in practice — use this page as a reference when building guardrails, writing evals, or debugging an agent that's already misbehaving in this way.
Accuracy Regression Undetected — A Bad Deploy Ships Silently
Attribute Hallucination — Model Reads the Right Field but the Wrong Value
OCR mixes watermark and background pixels into the extracted text, producing garbled characters
Batch Processing Timing Failures — Arrival Order Isn't Logical Order
Agent Extracts Right-to-Left and Mixed-Direction Text Reversed or Jumbled
Agent Classifies Blank or Near-Blank Pages as Real Documents
A Single Extraction Error Cascades Across Downstream Systems
OCR swaps visually similar characters (0/O, 1/l/I, 5/S, 8/B), so extracted IDs and codes look right but are wrong
Agent Interleaves Text From Different Columns Into Nonsensical Output
Merged Cell and Nested Header Failures
Overconfident Wrong Answers — Model Confidence Doesn't Track Accuracy
The same field appears in multiple document locations with conflicting values, and the agent grabs the first or an arbitrary occurrence instead of the authoritative one
Long documents exceed the agent's context window, so naive chunking breaks cross-references and the agent loses track of earlier content
Agent Swallows Attachments and Embedded Documents Into the Parent Document's Classification
Extracted Data Lands in the Wrong ERP Fields
The agent's own error-recovery step corrects a value that wasn't actually wrong, corrupting good data to satisfy a validation check
Fabricated Content Not Grounded in Input
OCR fails on decorative fonts, stylized brand typography, and logos, consistently misreading the same sources
Agent Appends Footnotes as Inline Text Instead of Linking Them to Their Reference
Document Format Diversity Overwhelms Rule-Based Extraction
OCR/ICR returns garbled or nonsensical text on handwritten form fields and signatures
Agent Interleaves Repeating Headers/Footers Into Body Content
Input Quality Gap — Real Documents Are Messier Than Benchmark Documents
Preprocessing Strips Layout Structure Before Extraction Ever Runs
Low DPI scans, faxes, and compressed mobile photos produce fragmented, unreadable OCR output
Agent Flattens Spanning Headers and Merged Cells Into a Grid, Destroying Parent-Child Structure
Extraction Guesses Wrong Without Sender, Locale, or Historical Context
Object Hallucination — Model Populates Fields the Document Never Contained
Agent Groups Batch-Scanned Pages Into the Wrong Documents
Plausible but Wrong Outputs — Errors That Never Trip an Alarm
OCR misreads punctuation — decimals become commas, currency symbols drop, dashes get interchanged — silently corrupting parsed amounts
Quality Filter Rejects Readable Documents as "Unreadable"
The agent's reasoning chain is logically sound, but it operates on a misread input value, so the final answer is wrong despite correct reasoning
Relational Hallucination — Values Attached to the Wrong Header
Poorly Calibrated Confidence Thresholds Flood the Review Queue
Extraction Completes Successfully but the Data Relationships Are Wrong
Agent Misclassifies Documents That Share Near-Identical Layouts
Skewed, rotated, or angled document photos cause OCR to merge adjacent lines into scrambled text
Physical stamps and handwritten annotations overlap printed text, so OCR interleaves both into a corrupted field
Agent Extracts Table Data as Continuous Text Instead of Rows and Columns
Cell Omission and Merging — Empty Cells Cause Column Drift
A Vendor Changes Their Document Template and Extraction Silently Breaks
The agent mistranslates its own document understanding into tool call parameters, silently extracting from the wrong page or region
The agent struggles on free-form correspondence and narrative documents where key information isn't in a labeled field
Agent Extracts Fields Using a Schema That Doesn't Match the Document's Template Version
Performance Collapse Under Visual Noise