If the last two years were about experimentation with generative AI, the next two will be about operational discipline. Across supply chain and logistics, the mood has shifted from curiosity to scrutiny. Leaders are no longer asking what a chatbot can summarise; they are asking what an autonomous agent can decide, and whether those decisions can withstand the volatility of global trade.

At a recent industry roundtable sponsored by Locus, senior supply chain and digital leaders shared candid reflections on where agentic AI is delivering value, and where it is colliding with operational reality.

Beyond content generation

For many attendees, the first wave of AI adoption was misunderstood. Generative AI tools demonstrated impressive fluency, but supply chain is not a content problem; it is a decision-making problem. As one participant put it, “We’re not here to generate text. We’re here to move goods. If we get it wrong, shelves are empty.”

That distinction matters. Agentic AI — systems capable of autonomous action within defined guardrails — promises more than summarisation. In theory, it can orchestrate workflows, trigger transactions, and reconcile data across fragmented systems. In practice, however, supply chains are often messy, distributed and highly regulated. Decision-making is rarely linear.

The data paradox

No theme surfaced more frequently than data quality. In deterministic machine-learning use cases — such as established demand-forecasting models — organisations have operated successfully for years. But when moving toward agentic systems that synthesise multiple datasets and take action, data brittleness becomes existential.

“Same model, same volume of data,” noted one executive. “One output is excellent, the other unusable. The difference? Context.”

The old adage of “garbage in, garbage out” remains brutally relevant. In agentic architectures, however, the stakes are amplified. Poorly contextualised inputs can propagate across integrated systems, influencing planning, procurement and customer communication before human intervention occurs.

Several leaders described failed or stalled pilots where disparate systems — ERP, transport management, customer service platforms — could not align sufficiently to support reliable agent decisions. The lesson was clear: controlling the source of truth dramatically increases the probability of success.

Human-in-the-loop: design, not default

While few believed humans would disappear from the supply chain, there was healthy disagreement about where they should sit in the loop, and what autonomy they should have. Prady Chowdhary, Chief Product Officer for Locus, noted “When a human makes a decision, we trust it; when AI does, the first questions is ‘Why?’ – and that’s a gap we need to close.”

In highly regulated industries — aviation, energy, tobacco — accountability cannot be outsourced to an algorithm. As one attendee observed “In certain industries, you’re going to be ultimately accountable for decision-making“. Those decision-making rights must be explicitly mapped. Where can an agent act autonomously? Where must it escalate? Who retains ultimate liability? Designing this governance architecture requires more than technical integration. It demands clarity around organisational risk appetite.

Yet there is also an uncomfortable truth: supply chains are built on tacit knowledge. In several organisations represented at the table, critical expertise resides in experienced planners’ heads. As demographic shifts accelerate retirements, knowledge transfer has become urgent.

Here, agentic AI shows promise. Knowledge agents trained on structured documentation and historical decision logic can support onboarding and continuity. One organisation reported early success building a chat-based internal agent to interpret process documentation and planning nuances — effectively creating an institutional memory layer.

The more ambitious question remains whether that tacit knowledge can be codified before it walks out of the door.

Execution versus planning

Another insight was cultural. Supply chain functions pride themselves on firefighting. When a vessel diverts, a port closes, or a geopolitical announcement shifts tariffs overnight, teams adapt. Improvisation is embedded in the DNA.

This execution bias, however, can inhibit architectural thinking. Agentic AI demands process clarity. Before delegating decisions, organisations must articulate what those decisions are. Indeed, one leader asked “What decisions do we actually make?”.

Without explicit decision mapping, agent deployment becomes reactive rather than strategic. This appears to suggest that the failure of many initiatives at proof-of-concept stage is because their underlying process remains opaque.

Fragmented ecosystems

Even where internal systems are stabilised, external dependencies complicate matters. Freight forwarders described operating across dozens of airlines and shipping lines, each with different data standards and varying digital maturity. Government customs authorities introduce further variability.

One executive described a case where automated milestone reporting conflicted with an airline’s manual data entry, forcing reversion to signed paper delivery notes to prove compliance. In such environments, fully autonomous orchestration remains aspirational. Agentic AI can optimise within the enterprise boundary, but cross-ecosystem standardisation lags behind.

This is particularly evident in areas such as real-time visibility, freight pricing intelligence and returns management in e-commerce. While marketplace models such as those popularised by Amazon have conditioned consumers to expect frictionless service, the operational backend remains highly fragmented.

Cost versus growth

Inevitably, the commercial lens surfaced, particularly in relation to the cost of investment into AI. One attendee noted “Unless you’re solving business issues and getting some kind of ROI, it’s very hard to justify it.”

Is AI primarily a cost-reduction lever or a growth enabler? Public markets often default to headcount reduction as a visible metric of AI “success”. Yet several participants cautioned that this framing is reductive. Automating an existing process may reduce labour intensity, but it does not necessarily create competitive differentiation.

A more nuanced view emerged: use agentic AI to absorb incremental growth without proportional headcount increases. In other words, redeploy capacity rather than remove it. This reframing potentially mitigates workforce resistance while aligning with shareholder expectations of productivity gains.

Crucially, many frontline teams fear displacement. Transparent communication and upskilling programmes are therefore essential. The objective should be augmenting human capability — freeing teams from repetitive system-hopping across multiple applications — not hollowing out expertise.

Governance as an enabler

Concerns around cybersecurity, data privacy and regulatory compliance were not positioned as blockers but as necessary guardrails. In an era of heightened scrutiny and high-profile breaches, uncontrolled experimentation is untenable.

One participant summarised it succinctly: “It’s not ‘no’. It’s ‘Yes, if.’”

If the use case is robust. If the data is governed. If the security architecture is sound.

Agentic systems, by definition, act. That action must be auditable. Particularly in consumer-facing environments subject to a variety of compliance regimes, uncontrolled data leakage would be catastrophic.

Defining “good”

Perhaps the most strategic challenge is definitional. What does good look like in 12 or 24 months?

Few organisations have articulated a clear AI north star. Without it, pilots proliferate but fail to scale. The art of the possible remains under-explored.

Leaders agreed that incremental experimentation is necessary — but so is vision. Where can autonomous agents genuinely transform lead time predictability, working capital efficiency or service responsiveness Where can they monitor global news signals 24/7 and escalate material disruptions before human teams wake up?

The technology is advancing rapidly. Multi-agent collaboration, tool-calling frameworks and contextual memory capabilities are improving steadily and consistently. What failed six months ago may now be viable. The question is whether organisations can evolve their operating models as quickly as the tools themselves.

From hype to discipline

The roundtable did not produce a single blueprint — nor should it have. Supply chains vary too widely in complexity and risk profile. But a pattern emerged.

Agentic AI is not a plug-and-play overlay. It is an architectural commitment that requires:

Clean, contextualised data
Explicit decision mapping
Designed human-in-the-loop governance
Clear value metrics beyond vanity cost savings
Cultural readiness to move from reactive firefighting to proactive system design

Those who treat it as a tactical bolt-on will struggle. Those who treat it as an opportunity to re-engineer decision flows may unlock material advantage. As Prady Chowdhary noted ““It’s not about choosing sides – it’s about finding what works for your industry and applying it thoughtfully.”

The era of hype is receding. The era of operationalisation has begun, which is where Locus come to the fore – as a globally-recognised provider of cutting-edge logistics solutions, leveraging innovation at enterprise level to drive real-world growth. Employing agentic AI to enhance and connect human visibility and decision-making in real-time, their aim is to enable businesses globally to transform their last-mile delivery.

To learn more, please visit: www.locus.sh

Source link