As scientists compete to build the first human-level AI, enterprises find themselves in a less dramatic, but no less important, race. The race to deploy game-changing AI use cases in their organizations.
However, while the business impact of AI can be transformational, time to deployment and value are critical. For businesses lacking extensive AI and data science talent, developing or even just deploying an AI model can be long and painful. Indeed, 93% of executives see AI as essential to future success, but 73% cite severe AI skill shortages that keep getting in the way. As a result, many AI adopters continue to struggle with ROI, and transformational change remains elusive.
Quick results are possible using generic foundational models, but the outputs often suffer from inaccuracy and need considerable manual review. Increasingly, businesses need to conduct an orchestra of different, highly customized AI models to achieve high accuracy and perform crucial business processes confidently and efficiently. But pulling this off requires extensive platform capabilities, built-in governance and controls, and the ability to train or fine-tune these models on business-specific data. That’s no small task when talent and time are in short supply.
In this article, I’ll offer two approaches to custom AI model development that shorten time to value while maximizing performance and accuracy.
Few enterprises have the resources or technical talent to build an AI model from scratch. Instead, many follow a ‘build your own’ approach by assembling the necessary components from various third-party providers. However, these patchwork systems can still be costly, difficult, and risky to build. And that’s before model training, prompt engineering, performance and compliance monitoring even begin.
Fortunately, you don’t need to rely on complex, self-built systems to create a custom AI model. Leading providers give their customers access to prebuilt, state-of-the-art AI models and then provide comprehensive tools to tailor it to their exact needs. This usually takes the form of a powerful but generic foundational model, which users can then train to understand the specifics of their business through data labeling.
Imagine you wanted to automate collections in the order-to-cash (O2C) process. As a communications-centric process, you’d need an AI model able to accurately understand customer messages, extract the right data, and even recognize sentiment across various communication types. Assembling such a system, even with existing components, will be costly, risky, and time consuming.
Instead, consider a capability like UiPath Communications Mining™, which provides access to the UiPath CommPath AI model as a service. Once you’ve uploaded your data, the model immediately begins learning and training from it. Features like sentiment recognition and multi-language support are built-in, and comprehensive tools are provided for fine-tuning, performance monitoring, governance, and continuous improvement. Users also benefit from an easy-to-use, guided user interface and its integration with the wider UiPath Business Automation Platform for driving end-to-end automated workflows.
Of course, an AI as a service approach still requires time to customize a prebuilt model to your requirements. However, the impact of recent GenAI innovations has been transformative.
GenAI models—like UiPath DocPath and CommPath—can immediately start processing and extracting valuable data from documents and messages with little-to-no training. Finetuning for industry or company-specific lingo or concepts is then done through a process called active learning. This is where the AI model and your team work collaboratively to customize the model quickly. The bulk of training is unsupervised and done by the model. The model only asks users to label examples it isn’t sure of.
An AI as a service approach can deliver custom, accurate AI models at a fraction of the time and cost as going it alone. Access to a foundational model, and the simple, intuitive tools needed to finetune it, removes much of the pain of model development and greatly accelerates time to value.
However, what if you wanted a custom AI model that requires no training from your employees or subject matter experts? Even with all the enhancements AI as a service provides, you might not have the people or the time for even a little data labeling or prompt engineering.
The leading vendors are starting to build custom AI services teams or ‘model factories.’ These are specialist teams who work with you to create the best-possible, best-performing, custom AI models. They’ll perform all the required data labeling, prompt engineering, and model training. Depending on your data policies, they can do this using your own data or synthetic data (lookalike data based on your processes). And, given the team’s experience and expertise, they can finish model training much faster than your average team of employees.
While AI services teams work fast, the process is still deeply involved and collaborative to ensure the reliability and specificity of the model. Typically, it’s comprised of:
Consultation and use case analysis: the team gathers requirements and analyzes processes. They will also decide on the technical approach to the project, which is often a choice between prompt engineering for an existing model or data labeling for a custom one. In general, the more specific and complex the use case, the more likely data labeling will be needed.
Data collection and preparation: a secure data transfer takes place, sample data is collected, and data quality assessed. All this is done within customer’s compliance and privacy policies.
Model training and development: the team designs the model architecture and develops the required features. Training data is annotated, or prompt engineering begins, depending on the technical approach agreed. During this step, the model is trained and iterations performed.
Testing, validation, and deployment: user acceptance is tested, and model performance evaluated as the team prepares the model for customer deployment.
Post-deployment support: some providers will offer support services once the model has been deployed. This includes continued performance monitoring and retraining as business dynamics and priorities change.
