Some time ago, the head of one of the world’s largest tech companies explained in a podcast that organizations using artificial intelligence (AI) to boost productivity and stimulate the economy would be the real winners of this revolution.
This statement highlights the tangible progress observed with generative AI, particularly with small language models (SLMs) and AI agents. Less visible than the large language models (LLMs) that ship on laptops and smartphones, SLMs offer remarkable advantages and concrete applications for field teams, especially in sectors such as retail and distribution.
A curated set of dedicated SLMs, integrated into a suite of AI agents, can be optimized efficiently for intelligent automation of specific tasks. These AI capabilities empower field teams to capture the context of their workflows with ease and then embed it directly into a mobile terminal equipped with AI agents to boost productivity, enhance the customer experience, and strengthen asset visibility.
Rendre l’IA réelle
SLMs are also ideal for on-device AI capabilities. They bring this technology directly to mobile devices, portable terminals, and other endpoints with limited resources, enabling features such as offline voice assistants and real-time translation.
AI agents based on SLMs enable the deployment of edge computing applications, processing data as close as possible to its source, which reduces latency and bandwidth consumption.
This technology offers meaningful benefits to field teams in distribution, warehouses, and logistics by improving real-time decision-making and operational efficiency. Here are a few examples of AI agents generated by SLMs:
1. Knowledge Agent : capable of interacting in natural language with training materials and standard operating procedures to facilitate onboarding and provide the information they need when they need it.
2. Sales Agent : helps answer questions from customers and colleagues, interrogates live stock levels and prices, and offers cross-sell or add-on recommendations.
3. Merchandising Agent : combines onboard image recognition and computer vision to automate the analysis of shelf conditions, identify stockouts, placement errors, planogram non-conformities, and pricing and signage errors.
L’IA au bon moment, et sans cloud
Embedded SLMs offer particularly compelling advantages for IT, innovation and engineering teams, especially in terms of privacy:
● Enhanced privacy: user data never leaves the device, which reduces the risk of data breaches and ensures better control over personal information.
● Low latency: processing occurs locally, with no need to query a distant server. Responses are therefore nearly instantaneous, which is essential for real-time applications such as voice assistants or translation.
● Reduced bandwidth and cloud costs: since data is not systematically sent to the cloud, local processing lowers mobile data usage and the expenses tied to cloud-based LLM computing.
● Offline capabilities: when LLMs are integrated into devices, AI can continue to operate without an Internet connection, which is particularly useful in areas where connectivity is limited or unstable.
The future is multimodal AI agents
The future of AI is intrinsically multimodal. Humans do not experience the world solely through text; they engage all their senses. AI must do the same, leveraging all of these “senses” to truly understand and interact effectively with the world.
The good news is that SLMs and AI agents can be multimodal, as in the merchandising agent mentioned earlier. To fully unlock their potential—especially when deployed on edge devices—they must indeed be multimodal and not limited to text processing and generation. Two main approaches enable achieving this goal:
● Integrated multimodal SLMs are designed to handle several modalities directly. This approach is the most efficient, but it requires particularly meticulous design and training work.
● Modular multimodal systems combine an SLM with distinct specialized models (for example an image recognition model or a speech-to-text transcription model). The SLM then acts as the coordinator, processing the text and interacting with the other models as needed.
The trend is toward more integrated multimodal SLMs as technology evolves and the training of even complex models becomes more efficient. However, a modular approach often remains simpler and more cost-effective in the short term.
The future will likely hinge on a combination of both approaches, depending on use cases and available resources. Ongoing R&D will enable the creation of integrated multimodal SLMs and more powerful AI agents, while developing robust modular systems that are easy to customize and deploy across a wide range of devices.
The objective is to empower AI systems to understand the world through multiple lenses, in order to offer interactions that are more natural, intuitive, and effective with people and their environment. AI that improves daily work will be the true winner of tomorrow.
*Andrea Mirabile is Zebra Technologies’ Global Head of AI Research
