Your AI can write a tidy email summary, but that is not the job. The real leap is from passive text generation to agentic AI that can read context, plan a sequence of steps, use tools through APIs, and execute actions inside real enterprise systems. That leap is thrilling, and it is also where most organisations hit the wall: plenty of pilots, very little production impact, and a growing fear of what happens when an autonomous agent is allowed anywhere near procurement, customer data, or payments.

TL;DR:

why AI investment keeps rising while production success stays low
the scaling wall: latency, compute cost, fragile error handling, messy data
the trust gap when autonomous agents can touch procurement, payments, and live systems
process inertia and the trap of paving the cow path
pragmatic AI mindset: hyper-specialised utility over sci-fi general intelligence
six pillars of agentic AI: tool use, action, memory, perception, planning, orchestration
multi-agent systems as modular digital specialists that isolate risk and raise accuracy

We use Google Notebook LM Agents to take insights from a Deloitte AI Institute report produced with Google Cloud to unpack why scaling enterprise AI is so hard and what actually changes when you build goal-oriented agents. Google Notebook LM Agents break down the practical architecture behind autonomous digital workers, including memory and reflection, multimodal perception, and planning that turns an ambiguous goal into an executable workflow. They also dig into multi-agent systems, where specialised agents work like a kitchen brigade rather than one giant generalist model, and why that modularity improves accuracy while reducing the blast radius when something fails.

Autonomy without governance is just risk at speed, so we get specific about controls: an agent OS hub-and-spoke model for visibility, FinOps guardrails and kill switches to stop runaway compute spend, and a defence-in-depth approach to security. That includes linguistic guardrails against prompt injection, sandboxing, semantic checks with constitutional AI auditing before actions execute, and infrastructure-level threat hunting. We also cover IDAMA, identity and access management for agents, so permissions stay least-privilege and accountability stays human-owned.

Finally, we bring it back to reality: change management, process redesign, and data gravity. You will hear concrete case studies in accounts payable automation and an agentic knowledge assistant with citations, plus why Apache Iceberg and cross-cloud lakehouse patterns matter for querying data where it lives. Subscribe, share, and leave a review if this helped, and tell us what task you would trust an agent to run first.

Support the show

𝗖𝗼𝗻𝘁𝗮𝗰𝘁 my team and I to get business results, not excuses.

☎️ https://calendly.com/kierangilmurray/results-not-excuses
✉️ kieran@gilmurray.co.uk
🌍 www.KieranGilmurray.com
📘 Kieran Gilmurray | LinkedIn
🦉 X / Twitter: https://twitter.com/KieranGilmurray
📽 YouTube: https://www.youtube.com/@KieranGilmurray

📕 Want to learn more about agentic AI then read my new book on Agentic AI and the Future of Work https://tinyurl.com/MyBooksOnAmazonUK