Beyond Automation: One Company’s Journey to Autonomous Accounts Payable with Generative AI

Peeriosity Peercast | August 2026

Overview

The presenting company’s Global Service Management team presented its journey from traditional accounts-payable (AP) automation toward an autonomous, Generative-AI-powered invoice processing model. The session walked through the business case, the agentic AI architecture built on the company’s internal agentic AI platform, pilot results from a live market rollout in Thailand, and lessons learned, closing with an open Q&A on invoice-number derivation, paper-invoice intake, and how AI compares to traditional RPA.

Note that this document is a summary. Peeriosity-100 members have full visibility to company names, presenters, Peercast attendees, and a recording of the Peercast discussion.

The Business Problem

The company processes approximately 4.4 million invoices annually across more than 70 markets, with roughly 2.4 million of those invoices still requiring significant manual effort. Traditional rules-based automation (RPA and legacy OCR/tax engines) was falling short of the company’s efficiency targets because of:

• High volumes of unstructured data: 30% of invoices still arrive as paper or PDFs sent via email
• Operations spanning 70+ diverse markets with country-specific regulations and evolving compliance mandates
• Elevated manual intervention and troubleshooting costs, concentrated in the “posting” and “troubleshooting” stages of the process (30% and 35% of total effort, respectively)
• A need to make processes “AI-ready,” integrate with existing and upcoming systems, and sustain accuracy against KPIs

Strategic Approach

The company scoped the initiative deliberately rather than tackling the hardest markets first, following an 80/20 approach:

• Defining the strategic scope: targeted manually posted invoices in high-volume markets with the largest automation gaps.
• Why this scope: high manual effort in key markets offered a clear opportunity for cost reduction and efficiency gains.
• Initial developments: proved out GenAI capabilities and built foundational models through data collection and model training.
• From pilot to product: the GenAI solution is now live and being prepared for broader scale.

Pilot markets were chosen for being low-complexity (high-impact suppliers under the 80/20 principle, single-PO-line invoices, simple tax coding), for high transaction volume to prove scalability, and to validate multi-language GenAI capabilities against local-language complexity.

Solution Architecture

End-to-End GenAI Workflow

The production workflow blends human review, SAP, an AI agentic layer, and RPA (BluePrism):

• Scanned invoices enter the SAP system as images
• An AI agentic workflow extracts invoice data and performs tool calls, checks assignment/payment controls (PO line, receipt, vendor, quantity, bank details, price), runs a compliance check, and validates tax coding
• Invoices that meet all criteria are posted automatically via RPA; exceptions route back to manual posting and human quality control

Agentic AI Design

The agentic layer is a multi-step workflow (built with LangGraph) using multiple right-sized Gemini models, optimized for cost and performance while meeting accuracy thresholds:

• P0: PDF checks (size, validity, password protection, encoding, rotation, name)
• S1 Supervisor triggers a validate-PO-line tool; T1 performs bounding-box detection for the invoice number
• T2: if OCR does not match the invoice number, an evaluator reviews a magnified image, OCR output, and the supplier’s last invoice numbers
• T3 retrieves the SAP purchase order, GR/IR history, and supplier details; T4 validates invoice compliance
• T5 validates tax codes; S2 Supervisor combines all details and generates the final response, which is parsed and verified before posting

An initial “toe in the water” pilot applied this agentic flow to PDF invoice pre-coding in Europe and North America, achieving:

Pilot Results: Thailand

Thailand was selected as the first full end-to-end pilot market and went live in July 2026:

2,200 successful automations were recorded during the pilot window.

Roadmap

The company’s stated next step is global expansion of the GenAI-powered end-to-end invoice process beyond the Thailand pilot, with candidate markets and hubs including North America, Latin America (Mexico City, São Paulo), Europe (Paris, Berlin, London), the Middle East/Africa (Cairo, Dubai, Lagos), and Asia-Pacific (Tokyo, Singapore, Sydney).

Lessons Learned

• Knowledge source gaps: existing documentation was often outdated, incomplete, or not practically actionable.
• Inconsistent process execution: discrepancies existed between documented process and actual practice.
• Current vs. future state mapping: required careful identification of gaps, challenges, and next steps.
• 80/20 approach: don’t attempt to solve the most complex problem first.
• Data integration: connecting to enterprise data sources takes real time.
• Governance, support, and business continuity (BCP): the solution can break and needs a support model.
• Operationalization: requires designing the future support organization.

The team framed this as an ongoing experiment, learn, adapt cycle rather than a one-time deployment.

Audience Q&A Highlights

• Paper invoice intake: A member asked whether shared services still receives paper invoices, or whether local teams scan mail to electronic format first, reflecting that paper/PDF intake remains a meaningful share of volume (about 30% per the process breakdown shown).
• Missing invoice numbers: Asked whether the solution can derive an invoice number when one is missing (e.g., monthly utility or leasing statements with only a statement date and a non-unique account number), the presenters said such documents are identified as “not an invoice” and are routed out for manual review rather than being forced through automation.
• AI vs. traditional RPA: A participant observed that AI appears more nimble and adaptable than structured RPA scripts, which can break easily (particularly when accessed URLs or GUIs change), underscoring a key advantage of the agentic AI approach over legacy rules-based bots.

Attendees responded positively overall, with multiple participants thanking the presenters for a strong presentation.

Key Takeaways

• The company is moving beyond rules-based RPA/OCR toward an agentic, GenAI-driven AP model built on a proprietary multi-model platform, combining Gemini models, LangGraph orchestration, and human-in-the-loop exception handling.
• A disciplined, staged rollout (pilot on low-complexity, high-volume, high-impact markets before scaling) produced strong early results: 99.7% AI accuracy and 30x cost efficiency in the initial PDF pre-coding pilot, and a 90.3% auto-post rate with 99.9% accuracy in the Thailand production pilot.
• Non-invoice documents (e.g., statements without unique invoice numbers) are deliberately excluded from automation and routed to manual review, preserving control and compliance.
• Success depended as much on process discipline (scope definition, data integration, governance/BCP, and future-state operating model design) as on the underlying AI technology.
• Global expansion beyond Thailand is the next phase, extending the model across the company’s broader market footprint.

Source: Peeriosity Peercast recording and chat transcript, “Leveraging GenAI and Agentic AI for Auto-Posting in Accounts Payable,” August 2026.