GenAI-Powered Invoice Document Processing and Automation

uber · Tech · 2025 · source Agentic: no · promotional bias: low · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: global invoice processing and accounts payable automation

Uber needed to automate high-volume supplier invoice processing because manual entry, RPA, Excel uploads, and rule-based systems were too slow, costly, error-prone, and brittle for global invoices with diverse templates, languages, handwriting, scans, multiple pages, and many required attributes.

Constraints:

• Uber processes a high volume of global supplier invoices daily
• invoices come from thousands of suppliers with varying templates and formats
• invoice documents arrive in over 25 languages and may include handwritten text or scanned copies
• each invoice contains 15-20 attributes plus line information that must be captured accurately
• manual handling caused high average handling time, operational cost, and reconciliation risk
• the system needed human review for critical fields while reducing manual processing

Architecture

Uber built TextSense, a reusable document-processing platform for invoice automation. Invoice PDFs arrive through manual upload or ticket ingestion, then a common backend sends documents to TextSense. TextSense performs ingestion, pre-processing, OCR through Uber's Vision Gateway, AI/ML extraction with trained or pre-trained LLMs, post-processing with business rules, data validation, HITL review, and metrics monitoring. The invoice workflow validates and enriches extracted data, presents side-by-side PDF and model outputs for human review, and sends approved invoices to the ERP system for payment. Cadence orchestrates nonlinear document-processing workflows, and model selection favored GPT-4-style GenAI for line-item extraction while leaving business logic to post-processing.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  supplier-submission-channels["Supplier submission channels"] --> ticketing-system["Ticketing system"]
  supplier-submission-channels --> front-end-web-app["Front-end web app"]
  ticketing-system --> ingestion-service["Ingestion service"]
  front-end-web-app --> common-backend-endpoint["Common backend endpoint"]
  ingestion-service --> common-backend-endpoint
  common-backend-endpoint --> textsense["TextSense"]
  cadence-workflow-platform["Cadence workflow platform"] --> textsense
  textsense --> object-storage["Object storage"]
  textsense --> pre-processing["Pre-processing"]
  pre-processing --> vision-gateway-ocr["Vision Gateway OCR"]
  vision-gateway-ocr --> ai-and-ml-models["AI and ML models"]
  ai-and-ml-models --> post-processing-layer["Post-processing layer"]
  post-processing-layer --> hitl-review-ui["HITL review UI"]
  hitl-review-ui --> erp-system["ERP system"]
  textsense --> metrics-monitoring["Metrics monitoring"]

Components

Data flow

Supplier invoices enter through either manual PDF upload or email/ticket ingestion. The front-end or ingestion service sends PDFs and parsed email details to a common backend endpoint, which calls TextSense. TextSense ingests and standardizes files, applies image augmentation and OCR, extracts invoice fields and line items with AI/ML models, validates and enriches the output with business rules and database checks, sends the result to a side-by-side HITL review UI, then forwards approved invoices to ERP approval and vendor-payment flows while capturing speed, accuracy, and cost metrics.

Why this architecture

Uber chose a modular platform instead of a one-off invoice bot because document processing needs recur across entity extraction, summarization, classification, invoices, and future document types. TextSense centralizes common OCR/LLM/post-processing/HITL components, Cadence handles nonlinear workflows, configuration-driven integration reduces country-template rollout time, and post-processing keeps business rules out of the model's raw extraction step.

Models used: TextSense · Cadence · Vision Gateway · OpenAI GPT-4 · seq2seq · Meta Llama 2 · Google Flan T5 · NLP · OCR · RPA · ERP

Scale

• Uber works with thousands of suppliers.
  “Uber works with thousands of suppliers, each using varying invoice templates and formats.”article, stated
• Invoices arrive in more than 25 languages.
  “Beyond diverse formats, invoices arrive in over 25 languages, complicating data extraction.”article, stated
• Each invoice contains 15-20 attributes plus line information.
  “Each invoice contains 15-20 attributes along with line information that needs to be captured accurately.”article, stated
• Uber used the last year of invoice data for model training and testing.
  “We considered the last year of invoice data for training and took 90% for training data and 10% for test data.”article, stated
• The solution produced a 2x reduction in manual invoicing.
  “We’ve seen a 2x reduction in manual invoicing.”article, stated
• The solution achieved 90% overall accuracy.
  “It has an impressive overall accuracy rate of 90%.”article, stated
• Thirty-five percent of submitted invoices reached 99.5% accuracy.
  “35% of submitted invoices achieved a near-perfect accuracy of 99.5%, and 65% achieved more than 80% accuracy.”article, stated
• Average handling time dropped by 70%.
  “The solution reduced the average handling time of invoice processing by 70%.”article, stated
• Uber reports 25-30% cost savings compared with the manual process.
  “They have also resulted in a 25-30% cost saving compared to the manual process.”article, stated

Alternatives considered (roads not taken)

Manual data entry and manual handling

Rejected because: Manual processing was slow, costly, delayed invoices through high average handling time, and increased error/reconciliation risk.

RPA, rule-based systems, Excel uploads, and self-service supplier platforms

Rejected because: These systems lacked adaptability for changing formats and required manual rule-setting, maintenance, and intervention at high volume.

Fine-tuned open-source LLMs such as T5, seq2seq, and Llama 2

Rejected because: T5 exceeded 90% on header fields but line-item accuracy dropped after the first line and fine-tuning led to hallucinations for line information.

Encode business rules directly in the model output

Rejected because: Uber found GPT-4 better at extracting what was available in documents than detecting existing invoice business patterns, so business logic moved to post-processing before HITL review.

Patterns

human-in-the-loop-gate (financial operations)

Uber routes extracted invoice data through a HITL review UI with side-by-side PDF comparison, alerts, and warnings before invoices are approved and sent to ERP payment workflows.

structured-output-contract (document extraction)

TextSense extracts specific invoice fields such as invoice number, date, amount, and line items, then validates those fields by exact or fuzzy matching depending on field type.

durable-execution (workflow orchestration)

Uber uses Cadence to manage nonlinear and verbose document-processing workflows across ingestion, OCR, model extraction, post-processing, validation, and integration.

eval-driven-development (model selection)

Uber trained and evaluated multiple model families, tracked header and line-level accuracy, and uses accuracy trends to drive retraining.

cost-aware-routing (model selection)

The model choice was based on cost-benefit analysis comparing fine-tuned open-source models with proprietary GenAI, and the rollout reports handling-time and cost savings.

fallback-degradation (automation scope)

The future roadmap reserves full end-to-end automation for cases that historically reach 100% accuracy, while other cases continue through HITL review.

observability-tracing (operations)

TextSense captures processing speed, accuracy, cost efficiency, and accuracy trends over time to drive continuous improvements and retraining.

Tradeoffs

Platform scope

Chose a reusable TextSense document-processing platform over a one-off invoice-only automation.

Rationale: A modular platform lets Uber onboard new document processing use cases through configuration changes and reuse OCR/LLM/post-processing components.

Model family

Chose OpenAI GPT-4-style GenAI over fine-tuned open-source seq2seq/Llama 2/Flan T5 models.

Rationale: GPT-4 handled complex unstructured documents and line prediction better, while fine-tuned models showed line-item degradation and hallucinations.

Automation boundary

Chose HITL review for extracted invoices over fully automated processing for all invoices.

Rationale: Invoice accuracy and compliance require human oversight except for future cases where the system has historically achieved 100% accuracy.

Where to implement invoice business rules

Chose post-processing layer after model extraction over forcing the model to infer existing business patterns.

Rationale: The model was better at extracting document-visible details, while deterministic business logic could be applied before user review.

Gotchas (what bit them)

RPA and rule-based systems did not scale with invoice-format diversity.

Root cause: As Uber onboarded new formats, RPA required manual rule-setting and training for each new format.

Resolution: Move to a GenAI/TextSense architecture that adapts to diverse formats through models, configuration, and reusable components.

Fine-tuned models performed unevenly across invoice headers and line items.

Root cause: T5 learned header patterns well but line-item prediction degraded from the second line onward and hallucinated line information.

Resolution: Prefer GPT-4-style GenAI for extracting document-visible details, then apply deterministic business logic in post-processing.

Business-rule inference was a poor fit for raw GenAI extraction.

Root cause: The model was better at extracting visible document content than detecting historical invoice-data patterns and business rules.

Resolution: Extract all required details from documents and apply business logic afterward in the post-processing layer before HITL review.

Field accuracy needs field-specific scoring.

Root cause: Some invoice fields need exact matching while others, such as descriptions, allow fuzzy string matching.

Resolution: Calculate accuracy at header and line levels and define the match type per field.

High-volume suppliers deserve targeted profiling first.

Root cause: Many invoices came from a small subset of suppliers, so labeling all suppliers uniformly would waste effort.

Resolution: Prioritize high-volume suppliers whose field-level accuracy falls below a threshold for profiling and labeling.

Evaluation

Uber evaluated invoice extraction with one year of invoice data split 90/10 for train/test, compared fine-tuned open-source models against GPT-4-style GenAI, measured header and line-level accuracy with field-specific exact or fuzzy matching, tracked accuracy trends over time, and monitored processing speed, accuracy, and cost efficiency.

Metrics: 90% training data · 10% test data · header-level accuracy · line-level accuracy · exact field match · fuzzy string match · processing speed · accuracy rates · cost efficiency · average handling time · manual invoicing reduction

Outcomes

• Uber reports a 2x reduction in manual invoicing.
• The GenAI-based system achieved 90% overall accuracy.
• Thirty-five percent of submitted invoices reached 99.5% accuracy and 65% exceeded 80% accuracy.
• Average handling time dropped by 70%.
• The solution produced 25-30% cost savings compared with manual processing.
• TextSense is positioned to support additional document types and enterprise integrations.

The case is production-scale for Uber finance: global supplier invoices across thousands of suppliers, more than 25 languages, 15-20 attributes per invoice plus line items, and reported outcomes of 2x manual-invoicing reduction, 90% overall accuracy, 70% lower handling time, and 25-30% cost savings.