Tesla is doing it with its automated vehicles, Apple with facial recognition, and Telco with its first-responder walkie-talkie functionality. Business leaders are looking for ways to bring AI applications closer to users — to the edge — to improve reliability, security, and user experience.
Depending on the task at hand and the infrastructure in place, AI at the edge could mean on a user’s device or a local edge network like Content Delivery Networks (CDNs) or cloud-based edge computing platforms. Retailers, for example, might have a dedicated local hub that collects data from various sensors like thermal cameras, light levels, or pressure mats. This device would process the raw data (at the edge) and potentially extract and anonymize relevant insights, such as footfall, before sending it to a central server for further analysis.
It takes leveraging a multi-cloud infrastructure for geographically dispersed stores to benefit from edge applications and accelerate innovation. Still, the company’s orchestration that connects the computing on the back end, and the user on the front, is a different set of challenges.
Business leaders must weigh messaging complexity, bandwidth, ease of integration, latency, context window size, power consumption, scalability, data security, and governance. Following are some of the benefits of AI at the edge and what questions you should ask to help implement the right tech stack.
Analyze data on the spot for instant customer tailoring
Some of retailers' largest and most important datasets include customer transactions, behaviors, inventory, logistics, and store sensor data. But does all this information need power-hungry transportation to, and analysis in, a central hub? Or can it be processed at the point of creation?
While retailers might continue to analyze vast customer purchase data across national stores centrally for broader insights, decisions based on local behavior would be much more 'real-time' (and cost-effective) if they were processed at the edge.
AI at the edge reduces the time lag between data collection and analysis. It can analyze real-time customer online shopping data (like location, browsing behavior, and purchase history) in the edge network, which lies between the user's device and the website's main server. This network can offer functionalities like:
-
Caching static website content closer to the user, reducing latency and improving website loading times.
-
Performing basic data processing tasks like filtering, aggregation, or anonymization before sending data to the main server.
-
Offloading some website functionalities (such as user authentication) to the edge to reduce load on the main server.
-
Personalizing promotions in real-time, targeting specific users with relevant discounts or coupons based on their current shopping cart.
-
Analyzing real-time inventory data from nearby stores (if applicable) and displaying accurate stock availability on the user's screen.
Inventory data for a specific store or customer behavior data from a particular website section could be processed locally to provide targeted promotions or optimize product placement. Moreover, customers can benefit from a smoother user experience — offloading data analysis tasks from the main server to the edge, which improves overall system performance and scalability.
Personalize experiences without invading privacy
Data privacy is on the tip of everyone’s tongues today, including customers. They want personalized experiences, but you’ll lose them if your security is not on point: 94% of customers said they would not buy from stores if they did not protect data properly.
Edge devices enable on-device learning and adaptation. This proximity significantly reduces the risk of data interception or unauthorized access, as less data is transmitted over networks. Since the refinement process occurs directly on the user’s device, it respects privacy and creates almost the perfect scenario for secure personalization. However, this is only valid if the user device’s security is airtight. Otherwise, malicious apps or malware could potentially access and exploit the user's data.
Personalization, privacy, and scalability form a complex triangle with trade-offs to consider when analyzing real-time customer data at the edge. More data often leads to better personalization, but it also raises privacy concerns. Enhanced privacy measures like data anonymization might require additional processing power at the edge, potentially impacting scalability, and these solutions might not be able to offer the same level of hyper-personalization compared to highly customized edge deployments.
A hybrid approach, combining centralized, edge, and device processing, offers the best of all worlds. Retailers can provide users with granular control over their data and the level of personalization they desire, alleviating privacy concerns. They can build trust by notifying customers in simple terms about the data processes and security required for each customizable feature they select.
Actionable steps for choosing the right techstack
Business leaders must weigh up the project's specific tasks, data processing needs, scaling and security requirements with the edge devices' capabilities. This includes the hardware's processing power and storage, along with software language processing speeds and data structures.
Let’s say a large retail chain wants to implement a highly accurate real-time inventory management system at all its stores. The system uses edge AI to analyze data from weight sensors embedded in shelves to track individual item inventory levels.
In this scenario, they might consider a language like C++ or Rust for their speed and efficiency in handling sensor data streams on the back-end. Frameworks like TensorFlow Lite or PyTorch Mobile can be used to deploy pre-trained AI models for weight estimation on edge devices. In terms of the architecture, a time-series database like InfluxDB or TimescaleDB would be suitable for storing and analyzing high-volume sensor data with timestamps.
Or perhaps a clothing retailer wants to use AI at the edge to analyze customer behavior in-store through video cameras. The system aims to provide personalized recommendations based on browsing patterns and identify potential shoplifting attempts.
This case requires a tech stack that can handle video processing, real-time analysis, and sensitive data privacy considerations. A language like Python with OpenCV or a framework like TensorFlow Lite for video processing on the edge device is a good option. Python’s readability, clear syntax, and extensive libraries make it easier for developers, while OpenCV's real-time video processing capabilities and TensorFlow Lite's efficiency enhance on-device model execution.
In all situations, retailers must ensure that they use techniques like anonymization or federated learning to address privacy concerns. They will likely consider storing anonymized video data for analysis and customer profile data (with consent) in a relational database like MySQL or PostgreSQL. These databases are well-suited for storing structured data like customer profiles and allow for efficient querying and retrieval of specific data points for analysis and generating recommendations.
AI at the edge offers retailers a potential secret weapon for personalization, but it requires a delicate balancing act. Business leaders must carefully consider project goals, data needs, security, and edge device capabilities to unlock the power of real-time customer insights and deliver the unique experiences that today's customers crave.
Ravi Narayanan is Global Practice Head of Insights and Analytics at Nisum, which provides consulting, development and strategy services to its customers.