Flexibility. That’s what you need from your enterprise resource planning vendor’s artificial intelligence features. The AI modules need to be flexible in the data they use, the data quality consumed, where the AI model’s predictions are used, when the AI model is run, and when new information is evaluated. These are just some of the ways an ERP vendor needs to make its AI offerings flexible.

The flexibility takes its shape from the different ways the AI module can be used.

Unfortunately, many ERP vendors don’t have the required flexibility in their current offerings. To be fair, AI needs a human touch to work with the data, understand the modeling process, and then deploy the AI model into operational processes. Those are customization procedures more than they are hard and fast software functions.

Still, many companies will get into AI through their ERP vendor’s offering. It will be introduced to the company as part of an upgrade or sold as an additional module. The bar to entry will be low, but the company’s ability to effectively use the AI modules will be limited by the flexibility the ERP vendor builds into that module.

Here are five things you need from your ERP vendor’s AI module to make it work effectively for your company. Spoiler alert: They are all related to flexibility.



1. Data Fields Used

Too many times, generic models make assumptions about which fields companies use. In reality, every company uses an ERP system a little differently. Different fields are used in different ways. This is true not just between companies but within a single company across organizations and through time. Therefore, the ERP offering must allow the analyst setting up the AI analysis to modify the fields included in the modeling process.

2. Data Filtering

Just as companies may need to add or drop fields to get to the same AI model for predictions, they may need to exclude values from individual fields. For example, a company may only have good data about one type of customer. The AI modules must let the analyst filter the field.

Most modeling processes require the analyst to filter by date. Some use cases of predictive AI models may need data for years, while others work best if they use data only from the past quarter.

3. Algorithm Choices

Choosing which algorithm are used gets deep into the technical weeds, but the choice is a necessary one in many instances. Some algorithms work better with particular types of data. Most people think of text fields versus numeric fields, but you don’t need to go far into AI to understand that different ranges of values within a field also affect how an AI algorithm performs.

Likewise, a common practice among data scientists is to use several algorithms, combine their individual predictions, and present the combined prediction for action. Being able to swap algorithms into the combined AI prediction model is useful.

4. Model Parameters

Getting really deep into technical issues is the concept of tweaking model parameters. Think of these parameters as akin to tuning an old-fashioned dial radio. An analyst can change an input parameter just slightly and come up with a much more accurate model.

Altering parameters isn’t always about fine-tuning a model, however. Sometimes, it’s about making the AI model simpler. If the parameters can be adjusted to make the model runs faster without sacrificing accuracy, that’s likely a better model.

5. Reporting Options

Delivering AI predictions isn’t straightforward. Sometimes, a company simply wants to prompt an action, so an isolated prediction on an ERP screen is adequate. Other times, AI predictions are used for future planning, so those predictions need to be made available for the reporting and business intelligence systems.

ERP vendors’ AI modules must allow analysts to direct the output toward different destinations as needed.