The Signals Loop: Fine-tuning for world-class AI apps and agentsÂ
Autonomous workflows that leverage real-time feedback and ongoing learning are becoming vital for boosting productivity and enhancing decision-making.
Initially, during the AI revolution, many applications were simply built on existing foundational models. However, as developers began to explore more intricate use cases, they soon noticed the drawbacks of just layering RAG over standard models. Although this method allowed for quicker deployment, it often didn’t meet the accuracy, reliability, or engagement levels necessary for more advanced applications.
That trend is changing fast. As AI transitions from merely assisting users to acting as intelligent collaborators, the underlying structure of these systems must advance. Autonomous workflows that embrace real-time feedback and continuous learning are now crucial for effective productivity and informed decision-making. AI applications that integrate continuous learning through ‘signals loops’ are emerging as a key strategy to create more adaptable and resilient systems over time.
Creating genuinely effective AI applications calls for more than just access to powerful LLMs. It requires a fundamental overhaul of AI architecture, prioritising continuous learning and adaptation. The ‘signals loop’ focuses on capturing user interactions and product usage data instantly, then systematically weaving this feedback into the model’s behaviour and product features, which leads to progressively better applications.
The rise of open-source frontier models is making fine-tuning—such as reinforcement learning—more accessible, rendering the construction of these signals loops increasingly practical. Technologies like memory retention are enhancing the capabilities of signal loops by allowing AI systems to gather context and learn from user feedback. This drives greater personalisation and boosts customer retention. With the increase in agent use, ensuring accuracy becomes even more paramount, highlighting the critical need for fine-tuning and implementing robust signals loops.
At Microsoft, we’ve witnessed the immense potential of the signals loop approach in action. Products like Dragon Copilot and GitHub Copilot showcase how signals loops can lead to swift product enhancements, increased relevance, and sustained user engagement.
How to Leverage Signals Loops for Continuous AI Improvement: Insights from Dragon Copilot and GitHub Copilot
Dragon Copilot is a tool designed for healthcare professionals, enhancing their productivity and the quality of patient care. The Dragon Copilot team has implemented a signals loop for ongoing product enhancement. They developed a finely-tuned model using a wealth of clinical data, which significantly outperformed the standard foundational model based solely on prompting. As usage has increased, the team employs customer feedback telemetry to continually refine the model. New foundational models are routinely assessed with automated metrics, and updates occur if benefits are substantial. This iterative approach leads to compounding improvements with each model update, crucial in a sector where precision is essential. The latest models now outperform foundational counterparts by approximately 50%. Such high performance allows clinicians to concentrate on patient care, capture comprehensive patient histories, and improve documentation quality efficiently and consistently.
GitHub Copilot was Microsoft’s first Copilot, generating considerable attention and setting a benchmark for AI-driven assistance. In just the first year, it soared to over a million users and has since surpassed 20 million. As the bar for code suggestion quality continues to rise, the GitHub Copilot team has shifted its emphasis to create a more robust mid-training and post-training environment, facilitating a signals loop that ensures continuous fine-tuning. The latest code completion models were trained on over 400,000 real-world examples from public repositories, with further adjustments made via reinforcement learning using custom, synthetic training data. Alongside these updates, the team introduced several client-side and user experience improvements, resulting in over a 30% boost in code retention for completions and a 35% enhancement in processing speed. These upgrades enable GitHub Copilot to predict developers’ needs and function as a proactive coding partner.
Key Insights for the Future of AI: The Importance of Fine-tuning, Feedback Loops, and Speed
The experiences gained from Dragon Copilot and GitHub Copilot highlight a significant shift in building and scaling differentiated AI products as we move forward. Several vital implications surface:
- Fine-tuning is essential—it’s not optional: Fine-tuning has evolved from being a niche activity to a strategic necessity that fosters considerable performance enhancements. Our products have shown that fine-tuning leads to remarkable gains in accuracy and overall feature quality. As access to open-source models broadens, the capability to fine-tune for specific applications will ultimately dictate product excellence.
- Feedback loops foster continuous advancement: As foundational models become more standardised, the sustainable success of AI products will hinge on how well these models learn from real usage. The signals loop—driven by genuine user interactions and fine-tuning—enables teams to offer high-quality experiences that keep improving.
- Firms need to adapt for scalable iteration, with an emphasis on speed: To facilitate regular model updates, companies must rethink their data pipelines, fine-tuning processes, evaluation systems, and team operations. Engineering and product teams must focus on rapid iteration and fine-tuning, alongside telemetry analysis and automated evaluation frameworks, to stay aligned with user needs. Organisations that enhance their systems and tools to integrate signals—from telemetry to user feedback—will be best equipped for leadership in the AI space. Azure AI Foundry provides the essential tools needed for sustained model and product improvement.
- Agents need careful design and ongoing adaptation: Building intelligent agents goes beyond mere model selection; it requires deliberate coordination of memory, reasoning, and feedback mechanisms. Signals loops enable agents to transform from basic responders into proactive co-workers that learn and improve from their experiences. Azure AI Foundry offers the necessary framework to support this evolution, aiding teams in developing agents that act responsively and adaptively.
In the past, fine-tuning AI models was costly and time-consuming. However, the introduction of open-source frontier models and techniques like LoRA and distillation has made tuning much more economical and accessible. Consequently, fine-tuning is now within the reach of more organisations than ever before. While ready-made models still have their place for general tasks like knowledge searches or customer support, more companies are exploring fine-tuning to cater to specific industry requirements, integrating their unique data into products and models.
The signals loop effectively ‘future-proofs’ AI investments by enabling models to perpetually enhance their performance as user data feeds back into the fine-tuned model, preventing stagnation over time.
Create Adaptive AI Experiences with Azure AI Foundry
To ease the implementation of fine-tuning feedback loops, Azure AI Foundry provides top-tier fine-tuning capabilities through a cohesive platform that simplifies the entire AI journey—from model selection to deployment—while ensuring robust compliance and governance. This empowers teams to create, adapt, and scale AI solutions confidently.
Here are four compelling reasons why fine-tuning on Azure AI Foundry is beneficial:
- Model Selection: Choose from a diverse range of both open and proprietary models offered by leading providers, with the option to select between serverless or managed computing solutions.
- Reliability: Count on 99.9% availability for Azure OpenAI models, along with guaranteed latencies using provisioned throughput units (PTUs).
- Integrated Platform: Utilise a complete environment that combines models, training, evaluation, deployment, and performance metrics, all in a single location.
- Scalability: Begin with an affordable Developer Tier for experimentation, and easily scale to production workloads using PTUs.
Join us in shaping the future of AI—where copilots evolve into co-workers, and workflows transform into continuously improving engines of productivity.
