Smart Suggestions: Smart suggestions while encoding invoices
For several years, Spend Cloud has been offering Autosuggest booking proposals when coding invoices. Due to recent developments in Artificial Intelligence (AI), we have further developed this technology and are introducing: Smart Suggestions.
Smart Suggestions will be rolled out in phases, you will be informed beforehand as soon as this will be enabled in your Spend Cloud environment.
What is Smart Suggestions?
Smart Suggestions uses Artificial Intelligence (AI) and historical invoice data to provide suggestions and automatically pre-fill fields for invoice coding. This feature supports the invoice processing workflow by pre-filling coding fields or offering suggestions. Smart Suggestions utilizes historical data from exported invoices to generate suggestions when filling in coding fields. Additionally, it incorporates the Autosuggest technique to pre-fill these fields. By means of AI, Smart Suggestions automatically fills coding fields. This technique has been used in Spend Cloud for several years and has recently been improved by adjusting how we retrieve and train AI information.
How Does It Work?
When an invoice is added, a process starts to fill in as many fields as possible. This includes Text recognition by Smartscan
to read, among other things, the creditor and invoice date, but also default values, such as a template or default general ledger account for a creditor, if applicable. For fields not filled during this process, Smart Suggestions will attempt to fill them for you or provide suggestions. You can recognize fields filled by Smart Suggestions by yellow highlighting. If the field is empty and you open the dropdown list, you will see yellow-shaded suggestions based on historical data. A percentage is shown indicating how often this value has been previously coded, for example, in combination with the creditor. Naturally, you are free to deviate from these suggestions and fill in something else. Smart Suggestions will only complete the initial coding. If you adjust the coding afterward, no new suggestions will be made.
to read, among other things, the creditor and invoice date, but also default values, such as a template or default general ledger account for a creditor, if applicable. For fields not filled during this process, Smart Suggestions will attempt to fill them for you or provide suggestions. You can recognize fields filled by Smart Suggestions by yellow highlighting. If the field is empty and you open the dropdown list, you will see yellow-shaded suggestions based on historical data. A percentage is shown indicating how often this value has been previously coded, for example, in combination with the creditor. Naturally, you are free to deviate from these suggestions and fill in something else. Smart Suggestions will only complete the initial coding. If you adjust the coding afterward, no new suggestions will be made.
Would you like a more detailed explanation of why Smart Suggestions fills in certain fields or makes suggestions? Read the final section of this article.
Which conditions are in place?
Certain conditions apply before coding fields are filled or suggestions are made by Smart Suggestions. These conditions are explained below. An important detail is that automatic field filling has different conditions than giving suggestions. Suggestions are solely based on you historical data. Automatic filling is performed by AI, which also looks at your historical data but also discovers patterns in your data to determine the optimal value. Again the tip: Read more about when and why Smart Suggestions fills in values in the final section of this article.
Conditions for Automatic Filling by AI
- Fields are only filled if they are empty. If a field is already filled (e.g., due to a default creditor value), AI will never overwrite it.
- The confidence level for automatic filling must be at least 95%. This means that AI will only fill a field automatically when it is at least 95% certain that the value is correct. If the certainty is lower, nothing will be filled automatically.
- Automatically filled fields are highlighted in yellow.
- The best results can be achieved when there are roughly hundred unique invoice lines encoded.
Conditions for Suggestions Based on Historical Data
- Suggestions are displayed in yellow in the field's dropdown menu. Next to the suggestion, a percentage is shown indicating how often this value has been previously used.
- For suggestions we always look at certain combinations. For example how often this value is used combined with the recognized creditor. At the bottom of this article we go deeper into these possible combinations.
- For historical data, Smart Suggestions looks at exported invoices from the past six months. Suggestions will also be shown if only one invoice has been exported in this period.
Explanation of how Smart Suggestions works
In this section of the article, we will delve deeper into the technology behind Smart Suggestions. Please note that in some cases, the explanation may be highly technical. The automatic filling of fields uses AI. The technology behind AI is vast and highly complex. Essentially, AI—primarily through machine learning techniques—creates patterns by learning from the structure of data and capturing relationships through problem-solving processes in a way that can be applied to similar datasets. As a result, AI can recognize, predict, or classify new cases based on the learned patterns.
Data Input
The AI system is fed a large amount of data representing the fields for which we want to receive suggestions. For Smart Suggestions, these fields include:
- General ledger account
- Cost center
- Organizational unit
- Employee
- Contract (if the Contract Management module is used)
- Commitment (if the Commitments module is used)
- Cost unit
- Extra (custom) fields
Learning Process
During training, the AI system uses algorithms to process input data across all administrations from one Spend Cloud environment and exclusively for invoices with the status exported or paid. To fill in values in the fields mentioned above the AI-system must recognize patterns. The following data is used for this.
- Creditor’s name
- City where the creditor is located
- Invoice amount
- Invoice line amount
- Invoice description
- Administration
- Currency
- Payment reference
- Our payment reference
- Notes
- Payment method
- Payment condition (if applicable)
- Linked commitment (if the Commitments module is used)
- Linked contract to invoice line (if the Contract Management module is used)
The model then evaluates these patterns, also called predictions against known outcomes. As soon as the AI-system recognizes a pattern, like a certain ledger always being used for a specific creditor, then this value will be filled.
Pattern Recognition: The learning algorithm adjusts the model’s parameters to minimize the difference between predictions and actual outcomes. Through repeated adjustments, the model learns to recognize patterns by identifying relationships between input data and the outcome.
Model Representation: Once the model is trained, it represents these patterns so they can be applied to invoices. The findings are “stored” as a large neural network, allowing AI to use these patterns to validate new data and create new connections within this neural network.
Generalization: The ultimate goal is for the model to generalize from the training data, enabling it to predict or recognize patterns in new, unseen data accurately.
If the AI does not recognize a pattern because it is new and different, the system will not assign a high level of ‘confidence’ (for Smart Suggestions, the baseline is a certainty of 95%). This means that when the AI system provides a value, we do not use it unless the model indicates that it was confident at least 19 out of 20 times. A scenario in which an automatic entry might be incorrect is when an invoice that is normally coded in a certain way needs to be processed differently due to a specific detail. Over time, however, the AI system can learn from these deviations and still present the correct (deviating) values.
We use AI because it is capable of predicting new values based on data it has never seen before. This is an extremely innovative way to use data. However, this also means that we do not always know exactly how the AI arrives at a suggested value.
Suggestions Based on Historical Data
Spend Cloud analyzes static data from invoices that have been exported or have the status "paid" over a period of six months. Suggestions may still appear even if only one invoice has been exported or paid within this period. The following combinations may be used for displaying historical data suggestions:
- Organizational unit: Combination of relation, cost center, and general ledger account
- General ledger account: Combination of relation and administration (if multiple administrations exist)
- Cost center: Combination of relation and general ledger account
- Cost unit: Combination of relation, general ledger account, cost center, and organizational unit
- Custom fields: Combination of relation, general ledger account, cost center, and organizational unit
- Employee: Combination of relation, general ledger account, cost center, organizational unit, and action
Frequently Asked Questions
What Is the Difference Between AI Suggestions and Historical Data Suggestions?
AI trains by looking multiple times at a large dataset for conclusions such as: "if this, then that." It then checks whether the outcome is correct. This is the moment when "the patterns" are found: the rules are clear and tested against various parts of all the data. This machine learning approach is very powerful compared to the statistical data that we use for suggestions based on historical data. For the historical data, we have written a set of rules ourselves. This is not only maintenance-intensive but also does not learn from new patterns. This way of suggesting provides quick suggestions but is not sustainable over time.
Setup of historical data
- We look at static data based on invoices that have been exported or have the status "paid."
- The given suggestions are based on a period of six months.
- Historical data is used to provide all possible suggestions for the specific field. These suggestions are organized from most accurate to least accurate.
- When opening a dropdown, you will see these suggestions in color, along with a percentage that indicates the confidence level.
- Suggestions based on historical data can be used immediately.
Setup – Filled values by AI
- AI is used for pre-filling fields (this is a suggestion with a high probability of at least 95%).
- Based on previously recognized patterns.
- Can propose new values based on these earlier patterns.
- Provides values after a period of three months to a year (depending on the number of unique processed invoice lines).
- Processes data per administration and across all administrations within an environment.
How do I know whether I see a suggestion by AI or historical data in a field?
Historical data is only used to provide suggestions for specific fields. When the AI shows a value with a confidence level of above 95%, we automatically pre-fill this value.
When does AI start making suggestions?
The pre-filling by AI works at an absolute minimum of 50 uniquely coded invoice lines. However, more than 100 invoice lines yield significantly better results. Depending on the number of invoices being processed, it may take three months to a year before AI provides valuable suggestions.
Related Articles
Encoding invoices with Autosuggest
Instead of manually filling in various fields, you can save time with Autosuggest for invoice coding. This advanced technology uses artificial intelligence to reduce repetitive tasks. Curious about how Autosuggest can make your digital invoice ...Statistics invoice processing
On the Statistics page of Invoice processing, you will find various statistics regarding the invoices that have been processed via Spend Cloud. For instance, you can view the top 5 creditors with the most invoices. Use the selection fields on the ...An invoice based on a contract (automatic approving)
If your organization uses both the Contract Management and Invoice Processing modules, Spend Cloud provides the ability to link invoices to contracts. This allows you to set up automatic approval for invoice lines based on the associated contract. In ...Configuration settings for Invoice Processing
In the menu section Configuration Settings / Invoice Processing, you will find settings that apply to the Invoice Processing module. Also, read this article for more information on configuration settings. Make sure you understand what each setting ...Encoding an invoice
In the Coding menu, you'll find all the invoices that can be processed (coded). By clicking on a row or the pencil icon from the overview, you can access the page where you can code the invoice details. Encoding Invoices When you start encoding an ...