Accounting automation software shouldn't rely solely on product hunches or marketing trends. At Viridem, every technology decision rests on two validation pillars: peer-reviewed literature in operational finance, and field feedback from more than 300 companies across Canada since 2013.
Our AI models and workflows are calibrated on studies published by recognized institutions in finance, accounting and document automation. No figure cited on this site is invented: every percentage, every cost, every productivity gain comes from an identifiable, verifiable source.
Our AI department and product leads work in a continuous loop with our clients. Every quarter we analyze real usage patterns, new vendor formats, exceptions raised by accountants, and update our models accordingly.
Our AI research department, founded in 2023, does one simple thing: carefully select the AI capabilities that bring real value to our Canadian clients. Continuous technology watch, internal testing on your real document formats, cautious rollout. No marketing gimmicks, no hype, just validated models.
The goal: every AI evolution we ship in your environment must concretely improve extraction accuracy, processing speed or document traceability. Everything else, we wait on.
Before signing, ask yourself four simple questions. They cover the real points that separate a good solution from an average one.
Ask whether capture really combines OCR and artificial intelligence. A hybrid approach is more accurate and more robust than OCR alone, especially on non-standard invoices.
A good solution comes with workflows preconfigured for your role (accounting, HR, construction, manufacturing, distribution). Building processes from scratch costs time and money.
Check that the tool connects to your existing ERP (capture, validation, approval, archiving, accounting write-back) to enrich it rather than replace it. A well-integrated specialized software beats a generic module.
Demand concrete figures: how many months to payback, what cost per invoice, what automation rate is targeted. A serious vendor puts those numbers on the table from the first meeting.
Six peer-reviewed works supporting Viridem's three founding technical choices: OCR + AI capture, preconfigured workflows, and integration with your ERP. Every link has been verified and points directly to the original publication on its open archive.
LayoutLM: Pre-training of Text and Layout for Document Image Understanding
Demonstrates that OCR + pre-trained language model outperforms OCR alone by +6 to +12 F1 points on structured extraction from invoices and admin forms. Academic validation of the hybrid approach.
Used at Viridem: Calibrates Viridem's OCR + AI architecture (official arXiv).
LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking
Evolution of LayoutLM with unified text + image masking. Reaches state of the art on understanding forms, receipts and invoices, as well as on document image classification and layout analysis. Confirms the superiority of pre-trained multimodal transformers for structured document extraction.
Used at Viridem: Calibrates the new generation of models used by our AI department for non-standardized vendor formats (official arXiv).
DocFormer: End-to-End Transformer for Document Understanding
End-to-end multimodal transformer combining text, vision and layout for document understanding. Achieves state of the art on document classification and extraction.
Used at Viridem: Confirms the value of text + image fusion in our OCR + AI engine (official arXiv).
AI-Augmented Business Process Management Systems: A Research Manifesto
Manifesto co-signed by the global BPM leaders (van der Aalst, Dumas, Mendling) on integrating AI models into business-process software. Establishes the « orchestrated workflows + hybrid AI » architecture as state of the art, exactly Viridem's preconfigured-workflow approach.
Used at Viridem: Consolidated academic backing for our preconfigured workflows: orchestration enriched with specialized AI modules, rather than standalone AI (official arXiv).
MarkupLM: Pre-training of Text and Markup Language for Visually-rich Document Understanding
Text + HTML/XML markup pre-training: shows that a « markup-aware » model outperforms purely visual approaches on natively digital documents (email invoices, web forms, EDI exchanges).
Used at Viridem: Foundation for our differentiated handling of natively digital documents exchanged with your ERP (tagged PDF, XML, EDI) vs scanned documents (official arXiv).
A Scoping Review of the Application of the Task-Technology Fit Theory
Review of Task-Technology Fit theory: technology only improves performance when its capabilities genuinely match the user's tasks. Scientific basis for choosing specialized software over a generic ERP module.
Used at Viridem: Theoretical basis for our choice of specialized software that integrates with your ERP to enrich it, rather than a generic module (PubMed Central).
Each link has been verified and points to the official publication. If a publisher moves a URL, write to info@viridem.ca and we'll restore the source.
The rest of our bibliography: publications that calibrate our document-extraction models, our BPM workflows and the OCR + AI architecture. Each reference is tied to a concrete use at Viridem.
Business Process Management: A Comprehensive Survey
The reference paper on business process management. Formally demonstrates that preconfigured workflows (process modelling) outperform ad-hoc approaches in admin environments exactly what Viridem delivers for accounting.
Used at Viridem: Source of our "preconfigured workflows" architectural choice over pure ML engines (read the article).
OCR-free Document Understanding Transformer (Donut)
Shows that an encoder-decoder transformer can extract a document's data without an explicit OCR step, avoiding the propagation of OCR errors on ambiguous characters. Performance comparable to classic OCR+AI pipelines with a reduced inference cost.
Used at Viridem: Active technology watch: the reason Viridem keeps an OCR quality-control layer rather than moving to all-AI, despite the growing maturity of end-to-end models (official arXiv).
Robotic Process Automation
Peer-reviewed editorial in the BISE journal dedicated to RPA. Draws the line between "lightweight" RPA (surface scripting) and BPM-orchestrated RPA. Recommends the BPM + RPA + hybrid AI approach for financial processes, exactly the Viridem architecture.
Used at Viridem: Academic source for our positioning as "orchestrated software" rather than an "isolated RPA bot" Springer BISE.
FUNSD: A Dataset for Form Understanding in Noisy Scanned Documents
Benchmark dataset for understanding noisy scanned forms: documents annotated with entities, links and key-value pairs. Used to train and evaluate field extraction on real administrative documents.
Used at Viridem: Evaluation reference for the robustness of our extraction on non-standardized forms (official arXiv).
ICDAR2019 Competition on Scanned Receipt OCR and Information Extraction (SROIE)
Reference competition on receipt OCR and key information extraction (number, date, total, supplier). Establishes the standard evaluation protocols for extraction on receipts and invoices.
Used at Viridem: Evaluation framework for our extraction of financial fields, total, taxes and supplier (official arXiv).
BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents
Pre-trained model focusing on text AND 2D spatial position for key information extraction. Shows that encoding relative positions improves field extraction, without even using the image.
Used at Viridem: Source for our spatial encoding of fields on invoices and purchase orders (official arXiv).
LiLT: A Simple yet Effective Language-Independent Layout Transformer for Structured Document Understanding
Language-independent layout transformer: a single pre-trained model transfers across multiple languages. Especially relevant for processing French and English documents with one architecture.
Used at Viridem: Justifies our bilingual FR/EN document processing with a single model (official arXiv).
DocILE Benchmark for Document Information Localization and Extraction
Largest public dataset for information localization and extraction on business documents (invoices, purchase orders): thousands of annotated documents and hundreds of thousands unannotated. The current reference for extraction on financial documents.
Used at Viridem: Reference benchmark to calibrate extraction on invoices and purchase orders (official arXiv).
Best Of Breed IT Strategy: An Alternative To Enterprise Resource Planning Systems
Analyzes the best-of-breed strategy (choosing the best specialized software per function) versus the all-in-one ERP: the specialized solution wins on functionality, domain expertise and deployment speed.
Used at Viridem: argument for specialized accounts payable software rather than a generalist ERP module (ResearchGate).
The Impact of Workflow Automation on Financial Efficiency and Operational Cost Reduction in SMEs
Study on automating financial workflows (accounts payable, receivable, bank reconciliation) in SMEs: average operational cost reduction of 30 to 70%, up to 85% productivity gains on repetitive tasks.
Used at Viridem: validates the ROI of accounts payable automation for SMBs, at the heart of Viridem Lite (ResearchGate).
Unifying Vision, Text, and Layout for Universal Document Processing (UDOP)
Unified architecture combining vision, text and layout in a single backbone: one model for classification, field extraction, question answering and summary generation on structured documents.
Used at Viridem: Calibrates our vision of a single multi-task engine (extraction, classification, search) instead of a stack of specialized models (official arXiv).
DocVQA: A Dataset for VQA on Document Images
Dataset and benchmark for question answering on document images: 50,000 questions over 12,767 images of real administrative documents. Reference benchmark for measuring a system's ability to « understand » a document.
Used at Viridem: Evaluation framework for our « conversational search » layer over archived administrative documents (official arXiv).
Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding
Pre-training a multimodal model by parsing web screenshots into structured HTML. Achieves state of the art on 6 document and interface understanding tasks without relying on a separate OCR stage.
Used at Viridem: Active watch on OCR-free models: helps us size our OCR quality-control layer against the evolution of multimodal transformers (official arXiv).
ERNIE-Layout: Layout Knowledge Enhanced Pre-training for Visually-rich Document Understanding
Pre-training enriched with reading order, layout-aware text masking and text-image masking. Beats LayoutLMv3 on FUNSD, CORD, SROIE and DocVQA for structured extraction.
Used at Viridem: Calibrates the marginal gains to expect from the most recent document transformers vs LayoutLM/LayoutLMv3 already deployed (official arXiv).
Workflow Control-Flow Patterns: A Revised View
Formal catalogue of the 43 workflow patterns used in BPM. The de-facto standard for modelling finance processes: cascading approvals, escalations, matchings, conditional routing.
Used at Viridem: Direct inspiration for our sector-preconfigured workflows (official PDF).
An Overview of the Tesseract OCR Engine
The seminal paper on the Tesseract OCR engine, foundation of most modern commercial OCR systems. Establishes pure-OCR limits on non-standardized documents motivating the need for an additional AI layer.
Used at Viridem: Calibrates our OCR quality-control layer against the standard open-source engine (Google Research archive).
Handwritten Optical Character Recognition (OCR): A Comprehensive Systematic Literature Review
Systematic review of 142 scientific publications on OCR from 2000 to 2019. Concludes that hybrid OCR + deep-learning approaches outperform single-technique methods on accuracy and robustness against document variability.
Used at Viridem: A meta-reference to calibrate the hybrid-OCR architectural choice on non-standardized invoices (IEEE Xplore).
Fundamentals of Business Process Management
The reference textbook on modern BPM. Extends the work of van der Aalst and Russell: BPMN 2.0 modelling, process mining, value-centric design, orchestrated automation. A university teaching standard in more than 200 process-management programs.
Used at Viridem: A consolidated academic reference to calibrate the BPMN modelling of our sector-specific accounts payable workflows (Springer Link).
Turning Robotic Process Automation into Commercial Success Case OpusCapita
Academic case study (LSE / U. Missouri-St. Louis) on the RPA transformation of accounts payable and financial back offices. Shows that sector-specific RPA deployments with preconfigured workflows generate a faster ROI than monolithic approaches.
Used at Viridem: Calibrates our sector-based deployment methodology (construction, manufacturing, distribution) and the RPA → BPM → AI sequencing (Springer Link).
Artificial Intelligence for the Real World
A synthesis of 152 AI projects deployed in enterprises. Establishes that AI projects targeting administrative-process automation (document capture, validation, matching) deliver the fastest and most measurable ROI vs fuzzier decision-support AI.
Used at Viridem: A strategic framework to calibrate AI deployment priority for our clients: high, measurable-ROI cases first (official HBR).
Automation of a Business Process Using Robotic Process Automation (RPA): A Case Study
Peer-reviewed case study on automating a complete business process with RPA. Measures the actual gains (processing time, error rate, cost per transaction) and identifies the critical success factors: task selection, upfront modelling, ERP integration.
Used at Viridem: Calibrates our pre-deployment assessment grid for clients (selecting which tasks to automate first) (official Springer LNBIP).
Process Mining: Data Science in Action
The reference work on process mining: automatic process discovery from event logs, conformance checking, continuous optimization. A university teaching standard and the scientific foundation for objective business-process measurement.
Used at Viridem: Calibrates our continuous-improvement approach: every Viridem deployment generates audit logs that let the client objectively measure gains and optimize workflows (Springer Link).
We publicly document what we learn. Client case studies, industry analyses, post-mortems on complex deployments: everything is accessible to your team before the purchase decision.
Detailed accounts of successful accounting transformations: Sani Marc, Chocolat Lamontagne, APCHQ, Vaillancourt, Métis-Neigette, Groupe Bellemare. Before/after figures, encountered roadblocks, deployment methodology.
Read the blog ›Industry-specific automation breakdowns: construction, distribution, manufacturing, financial services. Best practices, pitfalls, sector-specific ROI.
Read the blog ›Our AI department shares model evolutions, newly supported formats, and benchmark comparisons (OCR alone vs AI alone vs hybrid). Located in the Technology tab.
See the Technology section ›Complete calculation methodology, cited sources, explicit assumptions. Reproduce our numbers, challenge them, adapt them to your reality before calling us.
Open the calculator ›Our AI department answers technical, methodological or statistical questions directly. No question is too pointed.
Because our claims about accounts payable automation must be verifiable. We rely on recognized industry research (Ardent Partners, IOFM, CEPS) rather than on unverifiable figures.
Yes. Since 2023, Viridem has a dedicated AI research department tasked with integrating the most robust advances (document extraction, hybrid AI) while filtering out hype. That ongoing watch is what calibrates our models.
Our approach draws on leading document-AI research (work from Microsoft Research, Amazon and Naver Clova on document understanding) and on reference process-mining work (notably Wil van der Aalst). These publications validate our technical choices.
Through dual validation: empirical research (peer-reviewed scientific literature) and field feedback (hundreds of SMB clients). A model is deployed only if it holds up on both.
Real AI relies on documented, measurable methods (extraction accuracy, straight-through processing rate), not promises. Viridem publishes its sources and quantifies its results rather than overselling the word "AI".
According to Ardent Partners, best-in-class teams process a supplier invoice in about 3.1 days on average, versus more than two weeks for the least automated, a gap that automation closes.
