I85PREP
⌘ K
AA Advanced RPA study materials
AA Advanced RPA study resources

Invoice Processing Bot: Handling Exceptions and Scale

AA Advanced RPA · Put your knowledge to work in context.

All case studies

Invoice Processing Bot: Handling Exceptions and Scale

A shared services center processes 12,000 invoices per month. An attended bot handles invoice data entry into the ERP. The bot currently succeeds on 85% of invoices without human intervention. The remaining 15% are exceptions: 60% are due to missing PO numbers, 25% due to vendor name mismatches, and 15% due to illegible scans. The bot takes 45 seconds per invoice on average. Human agents take 4 minutes per invoice. The center operates 8 hours per day, 22 days per month. Management wants to increase straight-through processing to 95% and reduce handling time by 50%. The RPA team proposes adding an AI-based document understanding model to extract PO numbers and match vendors. The model has 92% accuracy on PO extraction and 88% on vendor matching. The cost of the AI model is $50,000 per year plus $0.02 per document. The RPA team also suggests implementing a queue-based dispatcher to distribute exceptions to human agents efficiently.

1. Calculate the current monthly bot processing time and human exception handling time. Then estimate the new exception rate and time savings if the AI model is implemented. Assume the AI model correctly handles all exceptions it can process, and that the remaining exceptions still require human intervention. Show your reasoning.

Read the worked answer

Current state:

  • Total invoices: 12,000/month
  • Bot success: 85% → 10,200 invoices processed by bot
  • Exceptions: 15% → 1,800 invoices handled by humans
  • Bot time: 10,200 × 45 sec = 459,000 sec = 127.5 hours
  • Human time: 1,800 × 4 min = 7,200 min = 120 hours

After AI model:

  • AI model addresses PO extraction and vendor matching, which together account for 60% + 25% = 85% of exceptions.
  • Exceptions due to PO: 1,800 × 0.60 = 1,080; vendor: 1,800 × 0.25 = 450; illegible: 1,800 × 0.15 = 270.
  • AI accuracy: PO extraction 92% → 1,080 × 0.92 = 993.6 ≈ 994 resolved; vendor matching 88% → 450 × 0.88 = 396 resolved.
  • Total resolved by AI: 994 + 396 = 1,390 invoices.
  • Remaining exceptions: 1,800 - 1,390 = 410 invoices (including illegible 270 + unresolved PO 86 + unresolved vendor 54).
  • New exception rate: 410 / 12,000 = 3.42% (so STP = 96.58%, exceeding 95% target).
  • New human time: 410 × 4 min = 1,640 min = 27.33 hours.
  • Bot time remains same for successful invoices? Actually bot still processes all invoices initially, then AI handles some exceptions. The bot time for 12,000 invoices at 45 sec = 540,000 sec = 150 hours. But previously bot only processed 10,200. Now bot processes all 12,000? The scenario says bot handles data entry; exceptions are those bot cannot handle. So bot processes all invoices but fails on some. The time per invoice for bot is 45 sec regardless? Typically bot attempts all, so time = 12,000 × 45 sec = 150 hours. But previously bot time was 127.5 hours because it only processed 10,200? That implies bot only processes successful ones. To be consistent, assume bot attempts all invoices and takes 45 sec each, but on exceptions it may take less? Better to assume bot time is for successful invoices only. However, the AI model is applied to exceptions, so bot still processes all invoices? Let's clarify: The bot does data entry; if it encounters an exception, it stops and routes to human. So bot time is for all invoices it attempts, but for exceptions it may spend some time before failing. We don't have that data. For simplicity, assume bot time remains 45 sec per invoice for all invoices it processes, but after AI, the bot might still process all invoices? Actually the AI model is an additional step: the bot extracts data, and if it fails, the AI model tries to extract PO and vendor. So the bot still processes all invoices, but for exceptions, it hands off to AI. The bot time for exceptions might be less because it fails early. To avoid complexity, we'll assume bot time per invoice is constant 45 sec for all invoices, so total bot time = 12,000 × 45 sec = 150 hours. But previously, bot time was only for successful invoices? That would mean bot doesn't attempt exceptions? That doesn't make sense. Let's re-read: 'The bot currently succeeds on 85% of invoices without human intervention.' That implies bot attempts all, but on 15% it requires human intervention. So bot time should be for all invoices. But then the 45 sec per invoice is average across all? It says 'The bot takes 45 seconds per invoice on average.' That likely means for all invoices it processes, including those that become exceptions? But if it fails, it might take less time. We'll assume 45 sec is the average time for all invoices the bot processes, so total bot time = 12,000 × 45 sec = 150 hours. Then human time is for exceptions only. So current total time = 150 + 120 = 270 hours. After AI, bot time remains 150 hours (since bot still processes all), AI processing time is negligible, and human time reduces to 27.33 hours. Total time = 177.33 hours. Savings = 270 - 177.33 = 92.67 hours, which is 34.3% reduction. But management wanted 50% reduction in handling time. So not met. Alternatively, if bot time only for successful invoices, then after AI, bot still only processes 85%? That would be inconsistent. I'll go with bot processes all invoices, time 150 hours. Then handling time reduction is 34.3%, not 50%. But the question asks to estimate new exception rate and time savings. So I'll present that.

Cost-benefit: AI cost = $50,000 + $0.02 × 12,000 = $50,240/year. Savings in human time: 120 - 27.33 = 92.67 hours/month × 12 = 1,112 hours/year. At an assumed labor rate of $30/hour, savings = $33,360/year, which is less than cost. So not financially viable unless labor rate is higher or other benefits. This is a judgment call.

2. The RPA team proposes a queue-based dispatcher to distribute exceptions to human agents. What are the key design considerations for this dispatcher to ensure efficient handling and avoid bottlenecks? How would you measure its effectiveness?

Read the worked answer

Key design considerations:

  • Priority rules: Exceptions should be prioritized by business impact (e.g., high-value invoices, aging).
  • Skills-based routing: Route exceptions to agents with appropriate expertise (e.g., PO issues to procurement specialists, vendor mismatches to master data team).
  • Load balancing: Distribute work evenly based on agent availability and current workload, avoiding overloading top performers.
  • SLA monitoring: Track time in queue and escalate if exceeding thresholds.
  • Exception categorization: Automatically tag exceptions by type to enable routing and analytics.
  • Feedback loop: Allow agents to provide resolution data to improve bot/AI models.
  • Integration with bot: Ensure seamless handoff with context (e.g., invoice image, extracted data).

Effectiveness metrics:

  • Average handling time per exception.
  • Queue wait time and backlog.
  • First-time resolution rate.
  • Agent utilization.
  • Exception recurrence rate (to measure learning).
  • Cost per exception.
  • SLA compliance.

Judgment: The dispatcher alone won't reduce handling time by 50%; it improves efficiency but the AI model is the main lever. The combination may achieve the target if AI reduces exceptions significantly and dispatcher reduces idle time.