Resilient Supply Chain Podcast: AI Accountability and the Limits of Autonomous Operations

Episode 121 of the Resilient Supply Chain Podcast, hosted by Tom Raftery, features Simon Bezrukov, Chief AI Officer at Bristlecone. The discussion examines how agentic AI, LLMs, digital twins, forecasting, simulation, and governance are reshaping supply chain decision-making. Rather than treating AI as a replacement for operational expertise, the episode focuses on where automation creates value, where it introduces risk, and why human accountability remains central. The full episode is available at www.resilientsupplychainpodcast.com

Automation Is Not Autonomy

The central tension in the conversation is the distinction between automation and autonomy. Bezrukov argues that agentic AI has a useful role in supply chains, but only within clear boundaries. It can open tickets, retrieve missing data, draft supplier communications, propose replans, or support defined playbooks when options are already understood.

The risk emerges when organisations confuse these bounded tasks with autonomous decision-making. Supply chains are not single-variable problems. They involve competing objectives across cost, service levels, inventory, carbon, customer commitments, contractual exposure, and operational risk. As Bezrukov puts it, “Agents are great at doing the paperwork of decisions, but they’re not yet great at owning the consequences.”

That distinction matters for procurement, operations, and supply chain leaders because accountability does not disappear when AI enters the workflow. It simply becomes easier to obscure.

LLMs Explain, But They Do Not Decide

Another major theme is the role of large language models. Bezrukov describes LLMs as powerful tools for retrieval, explanation, and accessibility. They can help planners query policies, contracts, SOPs, and planning systems in plain language. They can also explain why one scenario appears preferable to another after a calculation has been performed elsewhere.

Resilient Supply Chain Podcast, hosted by Tom Raftery and features Simon Bezrukov, Chief AI Officer at Bristlecone

But the episode draws a clear line between explanation and optimisation. LLMs are not built to manage constraint satisfaction, multi-objective optimisation, numerical reasoning, or complex operational trade-offs without grounding, tools, and governance. Tom Raftery summarises the point directly: LLMs are “brilliant explainers” but “not decision engines.”

Forecasting, Simulation and Resilience

The conversation also challenges the assumption that more data automatically produces better prediction. Bezrukov notes that machine learning depends heavily on the past resembling the future. In supply chains shaped by pandemics, geopolitical shocks, lane closures, labour constraints, and material shortages, that assumption often breaks.

The practical implication is that resilience requires more than forecast precision. Simulation and stress testing become critical because they help organisations prepare for a range of outcomes rather than a single forecast. This is particularly relevant as leaders move from just-in-time efficiency towards networks designed to absorb disruption.

Minimum Viable Models

The discussion on digital twins reinforces a similar message. Bezrukov cautions against modelling everything simply because it is technically possible. A smaller, governed model tied to a real decision can create more value than a comprehensive digital twin that nobody trusts or uses.

The operational test is whether the model changes action. If a signal is too weak, too costly to act on, or disconnected from a business decision, modelling it may add complexity without improving performance.

Bottom line

The strategic takeaway is that AI in supply chain is moving from experimentation into operational consequence. That shift raises the importance of governance, audit trails, safe rollout plans, explainability, and clear thresholds for automation. For senior supply chain leaders, the priority is not to pursue autonomy for its own sake, but to decide where AI can improve resilience, where human judgement must remain, and how accountability is maintained when decisions scale through software.

Interestingly, one of the episode’s most useful insights is also the least glamorous: sometimes the best model is the smaller one. In an era obsessed with digital twins and agentic systems, restraint may become a competitive capability.