Ex-Nasuni AI head’s AI implementation prescription

Jim Liddle, former Chief Innovation Officer at Nasuni and founder of Pract AI, warns that organizations lack the expertise to safely deploy agentic AI, citing the rapid evolution of the field since ChatGPT's late-2022 arrival. He prescribes rigorous digital employee discovery, identity, access controls, and monitoring, stating, 'Safe AI agents are not created by trust alone. They are made safer by identity, limited permissions, approval gates, and clear audit trails.'

Ex-Nasuni AI head’s AI implementation prescription Jim Liddle has decided opinions about agentic AI implementation, the deployment of digital employees. He was Nasuni’s Chief Innovation Officer https://www.blocksandfiles.com/ai-ml/2025/06/23/nasuni-ai-needs-edge-to-cloud-access-and-resilient-data/1593562 looking at Data Intelligence and AI, until new CEO Sam Kig rejigged her exec team in 2025 and the role went away. Liddle founded Pract AI https://pract.ai/ in January to bridge the gap between the potential of AI and the implementation actuality. Liddle reckons that, unlike traditional IT where an organization’s IT buyers and managers have broad general IT background knowledge, the AI field is developing so fast and its wider spread of buyers and managers do not have the same level of background knowledge and expertise. That’s not surprising as ChatGPT, the first generally available and powerful Large Language Model LLM , the precursor of today’s developing agents, did not arrive until the end of 2022, not quite 4 years ago. Becoming familiar with agentic AI capabilities and limitations, due to IT system hardware and software features, such as GPU memory capacity constraints, what they entail, and what system work-arounds, such as KV Caching https://www.blocksandfiles.com/ai-ml/2026/03/30/nvidia-and-its-partners-kv-cache-extenders/5209284 , involve, requires exec and manager learning. Without that, judgements made by them about AI agent system use, selection and deployment will be unreliable. If you don’t know about the features and capabilities of an electrically-driven car then you can’t sensibly choose between an internal combustion engine ICE vehicle and and a battery-powered car. Agentic AI buyers need to understand KV Cache and context memory basics and then move on to agentic AI access controls, as agents are, effectively, digital employees who can work 24x7, never get sick or need holidays, and execute IT command sequences at speeds a human worker couldn’t begin to contemplate. Liddle has produced several educative explainers: How KV Cache speeds up AI responses KV Cache and Context Memory in modern AI, AI and Agent access to enterprise data, Understanding AI Agent access control Understanding Context Memory in AI AI Agent access control at scale, KV Cache and context memory; although relatively new, are basic AI concepts. AI Agent access to enterprise data requires access controls and these, in the human sphere, are reasonably well understood. Although AI agents, being digital employees, require human-equivalent access controls, their implementation needs to coexist with thousands of agents acting and interacting at scale and operating with microsecond latencies. Their potential blast radius can, when they go wrong or make a mistake, expand massively and virtually instantly. The control on their access privileges need to be fine-grained and monitored constantly. Liddle says an AI agent should have the least amount of access privilege needed; give it only the access level it needs, for only as long as it needs it. This implies, of course, that all of an organization’s agents are known, identified and registered for control. The idea of shadow IT agents operating behind the scenes is horrifying. He says “Safe AI agents are not created by trust alone. They are made safer by identity, limited permissions, approval gates, and clear audit trails.” Do not enter the world of agentic AI deployment without first having a rigorous digital employee discovery, identity, regulation, control and monitoring program that can operate at scale.