Managing Artificial Intelligence Projects: Framework for Success. Part 6
Model deployment
In the deployment phase, sometimes referred to as model serving, model scoring, or model production, the evaluated model is put into operational use. This phase is less extensive compared to "operationalization," which will be discussed later. Deployment typically involves a smaller group of experts and users rather than providing organization-wide access.
Key considerations during deployment include the choice between real-time and batch processing of the AI model, the number of end-users, the types of applications, the expected output formats, turnaround time, and frequency of use. For instance, a predictive maintenance model in manufacturing might be deployed to monitor machinery in real-time, alerting a select group of maintenance engineers of potential failures. Alternatively, a retail company might use a recommendation engine in a batch mode to generate customer-specific product suggestions overnight for marketing teams.
At this stage, aligning with the preliminary risk assessment, a technical risk classification and analysis is essential, as the AI model now interacts with external systems and processes. Additionally, this phase involves extending considerations of AI ethics, governance, and regulation into the deployment. The risks posed by the AI deployment should be evaluated for all stakeholders, organizational functions, governmental regulations, social norms, and any other societal implications. For example, a financial institution deploying a credit scoring model must consider regulatory compliance and the ethical implications on different demographic groups.
Utilizing a risk register and a risk assessment matrix can help evaluate the criticality of each risk and document potential actions and mitigation strategies.
Post-deployment review
Depending on the industry and project scope, an expert panel, steering committee, or regulatory body will perform a thorough technical and ethical review of the project. This review covers everything from dataset methodologies and AI model validation to the evaluation of metrics and overall effectiveness. During this phase, compliance, standardization, post-implementation documentation, contracts, and service-level agreements are also managed.
For example, in healthcare, additional steps such as observational studies, small-scale clinical trials, training sessions, and user acceptance exercises are necessary. The team will also address intellectual property protection, considering options like patents, trade secrets (which provide ongoing protection after patents expire), or defensive publications in academic journals.
Examples include a tech company developing an AI tool for predictive maintenance in the aviation industry, where a regulatory body would closely scrutinize the model for safety and reliability. In healthcare, a startup working on an AI application for cancer diagnosis would undergo rigorous ethics reviews, clinical trials, and user training before the product reaches the market.
Operationalization
AIOps and MLOps are methods inspired by DevOps, which is a set of practices that aims to automate and improve the process of software development and deployment. DevOps is all about collaboration between different teams (development and operations) to continuously deliver new software versions that are reliable and work correctly.
AIOps and MLOps bring these automation and collaboration principles to the world of AI and machine learning. They involve creating systems to efficiently deploy and maintain AI models.
Unlike DevOps, AIOps deals with the complexity of handling dynamic and varying data that AI models process. This requires building special data and AI pipelines.
Data Pipeline: Involves tasks such as gathering, storing, and pre-processing data, ensuring its version control, and considering ethical aspects.
AI Pipeline: Involves aspects such as compressing models to run efficiently on different devices, defining the service, version control, auditing, retraining models, and maintaining, and monitoring them.
To manage these pipelines, technologies like microservices and containers are used:
Microservices: These are small, independent programs that perform a single function. For AI, a microservice could be an AI model that makes predictions.
Containers: These are packages that bundle the application code and all its dependencies together in a virtualized environment, making it easy to deploy across different platforms. Containers ensure that these applications run consistently and efficiently in various environments.
These practices help achieve automating as many tasks as possible in a tech-agnostic manner, allowing various technologies to work together seamlessly.
By Yesbol Gabdullin, Senior Data Scientist, MSc in Artificial Intelligence, MRes in Engineering, BEng in Electronics and Control Engineering, Microsoft Certified
The article was written for the Forecasting the Future of AI Applications in the Built Environment session of the Project Controls Expo 2024.