Configurable AI models: Low-code, high impact

Creating AI solutions used to require experienced data scientists and IT infrastructure engineers; today, however, low-code and no-code AI software is widely available, allowing for faster and cheaper deployment without sacrificing effectiveness. Low-code AI software uses drag-and-drop interfaces, pre-built components, and other tools to make implementation easier for non-data scientists, effectively enabling AI without coding.

While pre-trained models are usable out of the box, configurable AI models require training for the specific task they will perform. Gathering and tagging this training data (in essence, creating an answer key for the model to learn from) is immensely important, as a model will only be as good as its training.

While the marketplace for low-code and no-code AI offerings is not yet fully mature, several vendors offer viable options for a wide range of use cases. Given the lack of experience in the public sector with these tools, we decided to put two products to the test to determine whether the tools were truly mission ready. We benchmarked the tool performance on quantitative factors and captured developer feedback on qualitative factors such as ease of setup, user friendliness, and other intangible factors.

We chose two separate low-code AI products for testing, representing different vendor approaches to low-code AI. Our test set included Microsoft’s AI Builder, a low-code application development suite’s offering, and UiPath’s AI Center, a leading RPA vendor’s offering.

For our test use case, we imagined an agency’s internal travel support help desk inbox, where all employee questions related to travel are directed. The machine learning (ML) model would take incoming emails and categorize, or “tag,” them depending on their content. These tags could then be used in multiple ways through additional technologies. For example, data analytics and dashboarding could be used to monitor the types of requests received most frequently, allowing automation to reply to common requests for policy information or route certain types of messages to a live agent trained to deal with urgent requests for assistance.

We deployed a group of developers with experience in intelligent automation and low-code AI tools within the public sector to set up the products, train the models, conduct testing, and provide feedback on qualitative factors. Test outputs were consolidated and analyzed by a separate team. Similar studies frequently gloss over the setup and installation of low-code products; however, these steps can be unintuitive or complicated and have the potential to significantly lengthen the project duration. Our infrastructure engineers, solution architects, and developers supporting this study took note of the challenges, or lack thereof, setting up each product.

Our two chosen products performed similarly in terms of their average accuracy; however, Microsoft was less likely to provide an incorrect tag and instead detect no tags. UiPath’s predictive confidence was a more meaningful metric than Microsoft’s, as it was more highly correlated with correct and incorrect outputs. We noticed an inverse relationship between ease of (and speed of) deployment and the flexibility of the vendor solution.

In fulfilling the promise of low-code AI, each tool has its place. If the reason for using low-code AI is lack of access to technical resources, Microsoft AI Builder offers pre-selected ML models and guided walk-throughs. For an organization that wants something less time-consuming than a custom model but still wants the flexibility to choose from a variety of open-source models, UiPath offers a wide range of choices.

Configurable low-code AI models offer a balanced solution between pre-trained and custom models. They provide more flexibility compared to pre-trained models, allowing customization to specific needs. While some effort is required to set up and train configurable models, they offer a happy midpoint for organizations seeking tailored AI solutions for common use cases without investing in fully custom models.

Get in touch