As agentic AI gets more advanced, how do we decide where its independence should start and stop?

In this episode of Tech Tomorrow, David Elliman speaks with consultant and author Sam Newman about setting boundaries for agentic AI in large-scale systems. They also discuss why planning for uncertainty is now a key issue for many business leaders, and how doing small experiments with AI is ultimately the best approach.

Sam points out that non-determinism in agentic AI is a major challenge because its results are not always predictable. When these AI workflows are connected, small mistakes early on can spread and impact later parts of the system. To handle this, Sam suggests breaking systems into smaller, manageable parts and adding checks between steps to catch problems early. He also highlights the importance of being able to trace issues and roll back changes, so teams can fix problems and recover from failures. These steps are only possible if boundaries are set early and humans stay in the loop throughout.

They also talk about designing systems, so AI does not become a complicated dependency. One way is to keep AI tasks separate, using clear boundaries and security measures, often treating them as their own services within specific business areas. This makes it easier to manage data securely and to swap out models or vendors as technology changes and providers rise and fall.

Of course, costs make things even more complicated. Token-based pricing models can lead to unpredictable expenses, much like the early days of cloud computing, where many businesses were shocked that the promise of cost-cutting wasn’t delivered on. Subscription models for AI software can also hide high computing costs, making it hard for decision-makers to know how much they are really spending on agentic AI.

Overall, Sam’s main point is clear: try small, controlled experiments with agentic AI, but do not let them manage your large-scale systems without oversight, clear boundaries, and a way to undo changes if something goes wrong.

Episode Highlights

01:17 – How are agentic AI agents defined, and what is determinism in this context?

03:56  – What kind of issues are Sam’s clients having?

07:13  – The shift to breaking down problems into lots of modular steps.

08:48  – David’s Thoughts: What happens when AI agents pass problems down the chain?

10:12 – How does Sam approach agentic agent deployment?

16:32  –  Sometimes it just makes sense to write the code yourself.

19:30  – David’s Thoughts: Lessons learned from the move to the Cloud.

21:06  – Where Sam thinks generative AI may be heading.

25:36 – Sam’s advice on agentic AI? Do lots of small experiments.

26:56 – Wrap up.

About Zühlke:
Zühlke is a global transformation partner, with engineering and innovation at its core. We help clients envision and build their businesses for the future – running smarter today while adapting for tomorrow’s markets, customers, and communities.
Our multidisciplinary teams specialise in technology strategy and business innovation, digital solutions and applications, and device and systems engineering. We thrive in complex, regulated sectors such as healthcare and finance, connecting strategy, implementation, and operations to help clients build more effective and resilient businesses.

Links:

• Zühlke Website
• Zühlke on LinkedIn
• David Elliman on LinkedIn
• Sam Newman Website