Six steps to unlock value in accounts payable and beyond

Artificial intelligence (AI) has ushered in a new era in business operations, automating vast swathes of routine operations, boosting productivity, and driving down costs. Yet as organizations push the boundaries of automation, they encounter an all-too-familiar barrier: the last mile. The last-mile concept originates in transportation to describe the final step before delivery. In the context of AI, it's the difference between impressive proofs of concept and systems that perform reliably in real-time operations.

 

Nowhere is this more apparent than in accounts payable (AP) within finance and accounting, a domain that, despite AI's advances, resists full end-to-end automation. Genpact's Autonomy by design research shows that:

 

• 78% of finance teams are using AI for process automation, yet only 40% are using AI with autonomous agents

• The most significant pain points with AI implementation are regulatory and compliance challenges (62%), skills gap (62%), technology limitations (42%), and difficulty integrating AI into workflows (36%)

   

These last-mile challenges, rooted in intricacies beyond the reach of traditional AI, highlight the vital role of domain expertise in solving them.

 

In this blog, I examine five major last-mile challenges and how domain-driven AI bridges this gap, using real-life examples from AP and other business functions.

1. Complex data variability

2. Contextual misinterpretation

3. Regulatory and compliance nuances

4. Human judgment-driven exceptions

5. Fragmented systems and inconsistent master data

    

What sets success apart from failure in executing the last mile in complex processes? It's the deep, specialized, and practical knowledge of the processes used to inform AI systems and models: domain expertise.

While foundational models are improving at a rapid pace, the complexity of the last mile requires domain-driven AI contextualized through years of operational expertise. Let's examine practical, counterintuitive insights into how domain-enriched AI is solving the five last-mile challenges in AP.

1. Data complexity: Unit normalization

 

Imagine receiving invoices where a simple unit like gallon appears as "1 gal," "128 oz," "3.785 liters," or "bulk drum." Across thousands of vendors, this generates millions of mismatches. While generic AI can extract these terms, it cannot confidently match them to the item master or normalize them for accurate payment and reconciliation.

 

Domain-enriched solution: Genpact AP Suite uses historical SKU-level matching, extensive unit libraries, and context to resolve convoluted variations, helping reduce manual intervention and drive first-pass yield.

 

2. Contextual interpretation: AI's hidden blind spot

 

Consider the phrase "gas refill." For a US vendor, this typically means gasoline or diesel for fleet vehicles and therefore demands urgent payment. For a European vendor, "gas" could refer to natural gas used in manufacturing, with different coding and urgency. Generic AI misses these subtleties, risking misclassification.

 

Domain-enriched solution: Genpact AP Suite uses vendor profiles, geography, and invoice history to disambiguate meaning and increase correct processing.

 

3. Regulatory nuances: Beyond public data models

 

In Brazil, invoices often include multiple taxes, each requiring separate ledger postings, whereas in India, Goods and Services Tax (GST) reverse charges depend on vendor type. Without a deep understanding of country-specific regulations, generic AI can mispost, creating compliance risks.

 

Domain-enriched solution: Genpact Record-to-Report Reconciliation is continually updated by our experts with evolving regional rules, increasing correct general ledger posting and audit readiness.

 

4. Exception handling: Where automation falls short

 

AP is rife with judgment-driven exceptions such as midcycle vendor bank detail changes or quarter-end accruals. Generic AI typically flags these as exceptions, leading to heavy manual workloads.

 

Domain-enriched solution: Genpact AP Suite is trained on past exception patterns, approval chains, and email trails, enabling it to resolve or escalate appropriately to significantly reduce manual touchpoints.

 

5. System fragmentation and master data complexity

 

Invoices are often raised by subsidiaries, referencing different parent entities or missing purchase order (PO) numbers. Generic automation often fails to reconcile these cases.

 

Domain-enriched solution: Genpact Record-to-Report Journal Entry applies proprietary vendor hierarchies and multi-ERP knowledge to match and validate invoices, boosting match rates and reducing cycle times.

These AP examples show that last-mile challenges are not technical alone – they are business and operational at heart. Here's how domain-specific AI is built to overcome them in AP:

 

1. Comprehensive data collection: It aggregates structured and unstructured AP data, including invoice details, PO records, email approvals, and regulatory artifacts

2. Detailed domain annotation: Subject matter experts annotate data to capture nuanced vendor relationships, ambiguous terms, and exception handling rationale

3. Model training and fine-tuning: Foundation models are fine-tuned on annotated data using advanced techniques, learning domain-specific terminology and logic

4. Predictive agent development: Agents are designed to autonomously manage exceptions, resolve data mismatches, and interpret regional regulations, mimicking expert decision-making

5. Rigorous validation and deployment: Models are tested in controlled scenarios, shadowed against live AP workflows, and continuously monitored post-deployment

6. Continuous learning loop: New data and feedback from AP teams fuel ongoing retraining, helping models adapt to evolving business realities

The breakthroughs enabled by domain-driven AI in AP provide a blueprint for automating last-mile challenges across other functions and industries.

 

Supply chains: Port congestion and ETA accuracy

 

Building on AP's lessons, supply chain automation benefits from domain-aware AI that understands not just average congestion but also priority rules for perishable goods.

• Generic AI: Predicts a three-day delay for all containers at a congested port

• Domain AI: Recognizes that frozen seafood receives expedited clearance, integrating customs cycles to revise ETA to hours, not days

   

Procurement: Contract clause variability

 

Procurement mirrors AP's complexity in contract interpretation.

• Generic AI: Treats "Net 30 from invoice" and "30 days from acceptance" as equivalent, causing payment misalignments

• Domain AI: Genpact Procurement Suite links contract language to PO and goods acceptance data, ensuring correct payment terms and working capital optimization

   

Insurance submission clearance: High-volume complexity

 

Domain intelligence is essential for making sense of messy, variable submission data.

• Generic AI: Struggles to classify fields consistently, misses anomalies, and still requires heavy manual review.

• Domain AI: Genpact Insurance Policy Suite applies deep domain intelligence to classify, extract, and summarize 200+ data fields, automatically flag anomalies, and learn continuously through a feedback UI.