Most organisations have moved past the AI pilot stage. Users have access to the tools, security guardrails are in place, and data is being collected. But many firms are discovering that what worked in a controlled pilot bears little resemblance to what happens when real people use AI in their day-to-day work.
“AI is in full production now. Organisations have given users access to these tools, they’ve checked boxes on the security side, they’ve done some form of governance and guardrail implementation,” says Esteban Lopez, Senior Manager of Product & Technical Marketing at Theta Lake.
Related:
Why AI Pilots Don’t Prepare You For Production
The gap between a pilot and full production isn’t just about scale but also about behaviour. In a pilot, security teams do their best to anticipate how users will interact with AI. In production, they find out.
“You can only do so much in a pilot, and you’re only assuming how you think a user is going to use AI,” says Lopez. “In production, for the first time, you really understand how users are interacting with AI, how they’re trying to manipulate it, what it’s returning.”
As Dan Nadir, Chief Product Officer at Theta Lake, puts it: “With AI, we really are in new territory with respect to user behaviours. In the old world, there’s hardly anything new under the sun for most legacy platforms, like email. But AI is different. When you give a user a tool that has access to all your data, you just never really know what they’re going to do.”
The result is that risks which never appeared during testing start to surface. A user who crafts a question in just the right way can get AI to return information that is problematic from a compliance and governance perspective.
What AI Governance Needs To Look Like At Scale
Firms that are managing AI well tend to have followed a similar path. It starts with understanding where users are going and what tools they are using, moves through data hygiene and basic controls, and arrives at the harder question: How do you monitor what’s actually happening in those AI conversations?
“Monitoring AI interactions and communications becomes critically important,” says Lopez. “You need to know how users interact with their AI tools and what information these tools return. Behavioural visibility is the foundation for organizations to gain a deep understanding of their AI technology.”
The challenge is that traditional monitoring tools weren’t built for this. They look for known, structured risks such as data leakage of account numbers or social security numbers. Today’s AI risks are often subtle, behavioural, and can only become visible over time.
“Applying classifiers to prompts and responses to detect problematic content is important,” says Nadir. “But it’s the behavioural analysis over time that’s really critical. Repeated behaviours aren’t necessarily going to get detected if you look at just one record at a time. You need to be able to see patterns over time to really understand how users are actually behaving.”
The Problem With Over-Blocking
One response to the uncertainty of production AI is to lock things down. It’s understandable, but it tends to backfire. Users who can’t get what they need from sanctioned tools will find other ways. These workarounds can create far bigger problems than the ones firms were trying to avoid.
“Over-blocking causes user friction and frustration,” says Lopez. “It creates shadow AI. Users will simply go to go to their personal device instead. The question is how we intelligently allow users to interact with AI while maintaining the ability to monitor those communications and surface any risk.”
Monitoring The Conversation, Not Just The Output
Where Theta Lake sees firms focusing now is on that last mile: understanding what is actually happening in AI interactions, not just flagging when something goes wrong.
A single prompt might look harmless. But a user who repeatedly tries different approaches to get AI to produce something it initially refused to do is exhibiting a very different pattern: one that only becomes visible when conversations are monitored over time.
“Sometimes it’s not even malicious intent,” says Lopez. “Users may just be doing information gathering. But having visibility into how that behaviour shifts over time is where firms are looking to build real control over AI communications.”
Theta Lake captures and normalises AI interactions across any AI tool, summarises them intelligently, and applies behavioural classifiers built specifically for AI. Designed to complement and integrate with AI guardrails, LLM gateways, and SIEM solutions, Theta Lake provides enrichment to the analysis and alerts generated by guardrails; providing the investigation view of AI interactions for the SOC.
“We take all that AI interaction, we normalise it, we allow you to look back over months of history, we identify the risks within these long and verbose AI interactions and make sense of them,” says Lopez. Adding efficiency to the monitoring and investigation of AI content and communications while improving risk detection effectiveness in this new arena is a core strength of Theta Lake.
From Data Collection To Real Insight
Many firms are already collecting AI interaction data. The gap, increasingly, is a plan for evaluating whether users are interacting appropriately.
“Now that I have all the AI data, what do I do with it?” says Lopez. “How do I add intelligent monitoring to all of this communication? What should I look for? Those are the questions that are still unanswered.”
The firms getting this right aren’t just flagging individual incidents. They’re building up a picture of how users behave with AI over time, and feeding what they learn back into their governance approach.
“Even though they have all the controls in place from the pilot, they need to expect that they’ll do additional verification and tweaking and tuning,” says Nadir. “But if they can prove that the tool they’re using is built for scale from day one, it’s a good start. Once they have that data, even after the fact, they can do additional analysis on it.”
AI governance in production is not a one-time project. It’s an ongoing practice, and the organisations building it properly now will be the ones best placed as AI becomes more deeply embedded in how their people work.
Read more:
