Technically, an accountable agentic architecture looks like modern QA + auditability. You need (1) automated evaluations and unit-style tests for each agentic node, (2) end-to-end tracing of each agentic decision and action , (3) runtime policy enforcement and permission boundaries, and (4) continuous monitoring of each component with clear ownership.
The Shift to “Action”: When AI moves from generating text to executing tasks (ActionAI’s focus), what are the new security and safety vulnerabilities that cloud architects need to be aware of?
Cloud-Native Safety: For a CloudTweaks audience, how does the underlying cloud infrastructure impact the ability to monitor and mitigate AI bias in real-time?
Don’t confuse a compelling POC with a production system. The principle a CTO can’t ignore is: reliability must be designed end-to-end—before you scale adoption. That means your stack needs built-in evaluation, action gating, tracing/observability, and operational controls (versioning, rollbacks, policy enforcement) across the entire agentic lifecycle: building, testing, debugging, and monitoring. If you treat reliability as a bolt-on at the end, you’ll get stuck at the last mile.
Advice for Tech Leaders: If a CTO is currently rebuilding their AI stack for 2026, what is the one “safety-first” principle they cannot afford to ignore
In this CloudTweaks email interview, we correspond with Shai Dekel, Head of Reliable Intelligence at ActionAI and an Associate Professor at Tel-Aviv University for over 25 years. Shai brings a quarter-century of academic rigor to the challenge of “Reliable Intelligence” the vital bridge between high-level ethics and low-level code.
The Role of the Human-in-the-Loop: In the future of “AI that acts,” where do you believe the human should sit in the workflow to ensure safety without sacrificing efficiency?
The Academic Rigor: You are a Visiting Associate Professor at Tel-Aviv University. How does your academic research into complex algorithms translate into the “safety layers” being built at ActionAI?
The Black Box Problem: Enterprises are often hesitant to use AI for critical decision-making because of the “black box” nature of LLMs. How does ActionAI provide the transparency required for institutional trust?
The Collaboration Dynamic: You and Miriam represent two different worlds, enterprise tech and social activism. How does this “clash of perspectives” lead to a better technical product?
Two scaling traps show up repeatedly. First, some enterprises move too slowly because they don’t have the expertise to quantify risk and ROI—so leadership waits for a few ‘hero’ pilots to prove value. Second, others try to scale quickly, spend heavily, and then stall at the production boundary because reliability isn’t engineered end-to-end. Underneath those two patterns are recurring technical traps: fragmented data and eval pipelines, no drift detection, weak guardrails and rollback strategies, and security models that give models more authority than their confidence warrants. That’s why many visible wins today are still assistive (chatbots/copilots) rather than fully automated workflows—automation forces you to solve reliability, observability, and governance in a way pilots don’t.
In cases of system errors or agentic low confidence, our ExEx mechanism routes the case to the right stakeholders. So for example, in a financial auditing use case, system exceptions are routed to the IT department and agentic (logical) exceptions to the auditing team.
My main area of research is the mathematical foundations of machine learning. This rigorous mindset led us directly into how we build safety layers: we treat agentic behavior as a system that must be measured, stress-tested, and bounded. We apply machine learning principles of using mixture of experts, fine tuning for the specific use case, exception mechanisms in case of low confidence and integrated deterministic tools that reduce the risk associated with stochastic processes.
We have in fact built the ActionAI platform to be cloud-agnostic. However, there are some areas where the cloud infrastructure matters because it determines how well you can enforce policy and observe behavior at scale. Identity and access management governs which tools an agent can call; network controls and secret management prevent exfiltration; centralized logging and tracing enable auditability; and streaming telemetry lets you detect skew, drift, and bias signals in near real time. We keep the platform cloud-agnostic by abstracting these primitives, but the implementation details—logging sinks, key management, and identity providers—vary by cloud and affect how quickly you can detect and remediate issues.
In agentic orchestrations there are some agents that are custom code agents that simply execute code. This is a potential vulnerability that requires built in guardrails to avoid code injection. Recently a competitor’s open-source platform was hacked exactly in this manner.
Following our discussion last week with ActionAI CEO Miriam Haart on the ethical mandates of autonomous systems, we continue our exploration into the mechanics of trust. While Haart addressed the cultural and ethical foundations of AI, the practical implementation of such systems requires a level of mathematical certainty that defines the next frontier of the industry.
That diversity is exactly what improves the product. I tend to push on production-grade reliability, governance; Miriam pushes us to design for real-world constraints, user trust, and the human workflows around edge cases. The result is a platform that’s both technically rigorous and grounded in how organizations actually adopt change.
The Wix Perspective: CloudTweaks readers understand scale. Having led AI at a global giant like Wix, what are the most common “scaling traps” companies fall into when deploying AI today?
Moving to Agentic AI: We are seeing a massive shift from passive chatbots to “Agentic AI” that actually executes workflows. From a technical standpoint, how do you build a safety layer that is fast enough to intercept a biased “action” in real-time without causing massive latency issues for the user?
Our ActOne assistant supports reliable agentic solutions with triage where the slower reasoning processes are only applied on the more challenging cases, leading to a faster average time of processing. In some mission critical use cases, where previously long, intensive and error prune manual labor was applied, it is acceptable to allocate even parallel reasoning processes of minutes to ensure the high accuracy that is needed.
ActionAI’s platform has built in observability tools such as evaluation and monitoring of each component. Our unique ExEx mechanism of explainable exceptions allows to easily configure routing of edge cases the AI was not confident about to the right stakeholder, along with valuable information on missing information, key parameters the agents were not able to extract, etc.
Defining Accountability: We hear the word “accountability” often in AI. For you, what does a truly accountable AI architecture look like on a technical level?
