PoC as a Technology Benchmark
A PoC provides a unique opportunity to test AI’s capabilities in a controlled environment using your customers’ historical data. Unlike hypothetical or theoretical models, the PoC is built on real data from your customers, ensuring the outcomes are highly relevant and specific to your and your customers’ business needs. This allows you to see firsthand how the AI will perform in a live environment, eliminating guesswork and providing concrete evidence of its benefits.
The strength of a PoC lies in its ability to simulate the operational phase of AI on multiple customers without requiring full-scale implementation or months of data collection before obtaining meaningful insights. It allows for a risk-free process, offering insights into how AI handles complex coding at the line-item level, including General Ledger (GL) accounts, cost centers, and any other custom dimensions relevant to your organization. By providing transparency into the AI’s confidence levels and precision rates, the PoC gives your team the clarity needed to assess the value your customers will gain from embedding AI in your accounts payable system.
Comprehensive Insights into the AI’s Benefits
The primary goal of the PoC is to offer comprehensive insights into how integrating AI can improve your AP product offering and enhance the efficiency, precision, and scalability of your customers’ AP functions. The use of real customer data allows for a thorough evaluation of key performance metrics, such as:
- Prediction Precision: AI’s ability to accurately predict account codes and other dimensions, improving speed and reducing manual intervention.
- Confidence Levels: The PoC categorizes AI predictions by confidence levels (HIGH, MEDIUM, LOW), enabling your customers to focus on areas requiring attention while allowing the AI to handle the more straightforward tasks. Detailed metrics are provided for each confidence level.
- Operational Efficiency: The AI’s continuous learning from historical data helps improve long-term performance, reducing errors and streamlining the entire AP process.
By running this PoC on real customer data, you gain a comprehensive understanding of the AI's potential, building confidence in its ability to handle your and your customers’ specific requirements. Ultimately, the PoC provides a detailed analysis of the AI’s performance, offering an invaluable tool for making informed decisions about embedding AI into your invoice processing solutions.
Customer Fit for the PoC
The PoC phase is ideal for companies looking to transform their AP processes with AI-driven automation. Customers best suited for the PoC include:
- High Volume: Organizations processing large volumes of invoices where manual coding is resource-intensive, typically 15,000+ invoices yearly.
- Rich Data Quality: Kaunt utilizes the invoice and posting data of your customers. Ideally, the invoice data is extracted at the line level and includes the original PDF/image.
- Complex AP Workflows: Companies dealing with multi-dimensional coding (GL accounts, cost centers, projects, etc.) that require advanced automation and/or decentralized coding, meaning people outside the finance department do the invoice coding.
- Open to Innovation: Businesses ready to explore advanced AI technology to optimize AP functions and reduce manual efforts.
PoC Process and Timeline
The following guide outlines each step of the PoC process, giving you a clear roadmap from data preparation to performance evaluation, ensuring a smooth transition from testing to full implementation.
The Kaunt AI provides line-level account coding, including for e.g. GL accounts, cost centers, and any custom dimensions relevant to your organization. The AI uses your customers’ historical invoices and postings for training, identifying patterns in your invoice data and accounting information that can automatically predict account codes for new invoices.
AI Model Lifecycle Phases
- Training Phase:
- The AI analyzes your customer’s historical data to understand your customer’s accounting practices.
- Training data requirements:
- Historical Invoices: Ideally from the last 12 months. Formats can include e-invoices, images/PDFs, structured data, or a combination. Line-level invoice information is needed for splitting invoices into multiple postings.
- Posting Data: Includes GL accounts, dimensional coding (e.g., cost centers, tax codes), and other relevant dimensions. Each posting must reference the corresponding invoice.
- Data Volume: A year’s worth of data (15,000+ invoices) ensures the AI is trained with sufficient variety and depth.
- Test Phase:
- The AI is benchmarked and predicts postings on the test data set, simulating the operational phase to provide rapid insights into AI performance.
Process Overview
The entire PoC process is designed to be simple for you. Kaunt’s engineering team handles the majority of the technical work, allowing you to focus on the business case. Typically, the process takes about two weeks.
What Does It Take?
- Get acquainted with Kaunt’s API documentation (docs.kaunt.com).
- Identify 3–5 relevant PoC customers.
- Get relevant compliance documentation signed.
- Decide on the evaluation process and metrics.
- Map your data formats into the Kaunt API format.
- Create two data sets.
- Upload postings.
Kaunt will provide an evaluation report and present the results at an evaluation meeting.
Detailed Process Overview
- Planning Meeting: Discuss the PoC process, data requirements, and API introduction. Identify 3–5 test customers.
- Compliance: Ensure necessary compliance documents are signed (e.g., NDA, Data Processing Agreement). Kaunt is ISO/IEC27001 and ISO/IEC27701 certified.
- Data Mapping: Map data formats into the Kaunt API format (e.g., PDFs, e-invoices). Once mapped, initiate the PoC process.
- PoC Process:
- Data Splitting:
- Training Set: First 11 months of invoices and postings.
- Test Set: Last month of invoices and postings (hidden during training).
- Data Upload: Upload the training set to our API.
- AI Training: Train the AI model using the training set.
- Prediction: Send test set invoices to the API, and the AI makes predictions.
- Evaluation: Upload the test set postings, and the predictions are compared to actual postings.
- Results Presentation: Share a detailed report of the PoC results.
- Next Steps and Go-Live Planning: Discuss plans for transitioning to full implementation.
PoC Outcome
The final report will showcase the AI’s "out-of-the-box" performance, including key metrics per PoC customer:
- Confidence Levels: Percentage of invoices predicted with HIGH, MEDIUM, and LOW confidence for each dimension.
- Precision: Accuracy of predictions for each dimension, broken down by confidence level.
- OCR Extraction: Validation of OCR-extracted invoice fields.
Evaluation Process
There are two evaluation options:
- Kaunt-Run Evaluation:
- Provide 12 months of invoice PDFs and posting data. Kaunt evaluates performance on the last 2 months.
- Optionally, models can be exposed through an API for your testing.
- Own Evaluation:
- Provide 10 months of data for training, and 2 months of invoices (no postings) for predictions. Postings for the 2 months are sent later for evaluation.
Example Metrics in the Final Report:
- Automation Rate: Percentage of invoices with account coding proposals at a given confidence level.
- Precision: Accuracy of predictions matching actual postings.
What to Keep in Mind During the PoC
- Accuracy Enhancement: AI won’t achieve 100% accuracy but often surpasses human performance and improves over time.
- Data-Driven Performance: Quality and consistency of historical data are key to success.
- Continuous Learning: Feedback and fine-tuning improve accuracy over iterations.
These insights will guide the decision to proceed with full-scale implementation.