AI Talk: Observability in the AI era – unlocking innovation and driving performance 

Views on news 

As AI systems and agent workflows become more complex and autonomous, the associated risks grow. Business and process owners need confidence that AI is operating as intended, aligned to policy, and making decisions that are fair, safe and reliable. One capability needed to help monitor AI-related risks is observability, in complement with an evolved AI governance framework, holistic testing practices and clear management criteria, among others. Observability tools collect and analyse meaningful signals including logs, traces, model outputs, and data flows throughout its life cycle. These signals are interpreted into metrics and alerts relevant for business leaders, helping to turn technical data into actionable business insights. Over the past few years, observability shifted from a reactive tool to a proactive intelligence layer. It can also act as a control pane for AI initiatives. However, anomalies are often spotted by users earlier than they are flagged up on a dashboard. Finding the right observability solution starts with understanding the business and the outcomes that it wants to achieve with the solution. The business must also establish how much of the AI estate they are willing to observe and what spend they have attached to it, as well as timelines. Most frequently monitored outcomes include speed and accuracy of responses, customer satisfaction and reduced wait times.  

 

Meta agents monitoring autonomous agents 

Monitoring is nothing new. So when businesses want to deploy observability tools, first they should decide what should happen to their existing monitoring systems. Then it must be defined what good and bad looks like. Even an AI agent working too well can negatively impact downstream workflows. Teams with successful observability programmes have it built into the design from the very start. As observability technology is in its infancy and works in the background, it’s hard to measure its RoI. Where there are no humans in the loop, AI agents can be monitored by other supervisor agents – especially where interpretability also plays a role. In these contexts, agents can reveal not just what’s happening but also why.  

 

Sometimes the observability layer looks great but the semantic drift and the query the model ran were completely wrong, although it did connect to the right tool. Interpretability is particularly key when humans are being incrementally removed from the loop.  Without interpretability, the model’s thinking tokens can’t be seen. Troubleshooting is easier now, when LLMs write in English. But this may change in the future if they shift to non-natural language formats. Observability solutions monitor grounding relevance, toxicity, guardrail executions. Beyond the LLMs performance, it must also be checked whether the model is improving through learning.  Fully autonomous agents can have meta agents monitoring them, which burns more tokens but offers more observability and interpretability. There are cases where troubleshooting agents could solve a critical, high-impact (P1) issue, which human experts can find hard to beat. P1 and mean time to repair (MTTR) are now also key metrics of agentic system performance. Improvements in this area may mean a 7 minute instead of a 30 minute downtime – which can be a huge difference. A metric even more important than performance is the model’s ability to learn – and observability is a handy tool to monitor that too. What LLM users should also bear in mind is LLMs’ tendency towards sycophancy, heroism and gaslighting. They also tend to mirror their user’s issues, which poses high risk when they are used for decision making.  

 

The panel’s advice 

• Currently 69% of agentic workflows has a human in the loop.  
• The three criteria key to RoI are observability, interpretability and token economics. 
• Observability is the control plane that aligns AI with business outcomes and builds trust by providing evidence. 
• If you can’t tie a business metric to an outcome, you’d better not track it at all.  
• There is now insurance for agents – which can enhance adoption in customer-facing deployments where the brand’s reputation is at stake. 
• There should be more talk about culture and ethics in relation to LLMs. 
• LLMS may perform well at benchmark tests and then perform differently in real world scenarios. Observability can make the model “aware” that it’s being tested.  
• For Dynatrace’s Observability Playbook, click here.