From the Operations Floor - Agentic Supply Chain

The Software Did Not Save Us.

By Mark Riskowitz, VP Operations, Caraway - Guest Writer · March 2026 · 10 min read

On the failure of enterprise software as a substitute for intelligence, and the arrival of something genuinely different.

01 - Software is a Tool. Intelligence is the Work.

Enterprise SaaS gave us better spreadsheets. ERP systems gave us integrated spreadsheets. BI platforms gave us pretty spreadsheets. And yet the average operations team still runs on a fragmented stack - an ERP and a TMS that speak different dialects of the same language, a WMS and an OMS that share data only after someone has mapped, cleaned, and reconciled it, a carrier portal that communicates in PDFs, and a Slack channel where someone manually escalates what no system could resolve on its own.

None of these layers changed the fundamental nature of the problem: supply chain operations require judgment, and judgment cannot be templated. Every "optimization" layer we added required a human to interpret its output, escalate its exceptions, and absorb the consequences of its blind spots. We automated the easy parts and called it a transformation. It was not.

02 - The Gap Between Data and Decision Has Always Been a Human.

Any operator running a brand at scale knows the real bottleneck: it is not data availability. We are drowning in data. The bottleneck is the conversion of data into action - the supplier invoice that doesn't match the PO and sits unresolved for three weeks, the carrier overcharge that slips through because nobody audited the contracted rates against what was billed, the inbound shipment exception that lived in someone's inbox until it became a stock-out.

Software surfaced these moments. It could not resolve them. That gap - between signal and response - has always been staffed by exhausted, talented human beings working against impossible complexity. The cost of that gap is not just operational. It compounds quietly on the P&L, in ways that rarely get traced back to their source.

The question was never whether machines could process supply chain data faster than humans. The question was whether they could reason about it — and act with the contextual judgment of a seasoned operator. That question now has an answer.

03 - Agentic AI is Not Another Software Layer. It is a Different Category.

The distinction matters enormously and is frequently obscured by vendors who affix "AI" to legacy architectures. An agent does not surface an exception for a human to resolve. An agent reasons about the exception, considers constraints, selects a course of action, executes it, and learns from the outcome. Critically, it does this across every system the operation runs on - not just the ones with clean APIs and IT-approved integrations.

True integration is not a connector. It is comprehension - the ability to navigate a system the way a trained employee would, learning its structure, its quirks, and its data on contact. That is a fundamentally different proposition than anything the enterprise software industry has offered before, and brands that internalize this distinction early will build operations structurally faster, leaner, and more resilient than those that do not.

04 - The Operator's Advantage is Still the Operator.

Agentic AI does not make supply chain experience obsolete - it makes it more powerful. The operator who understands lead times, supplier relationships, demand seasonality, and margin pressure is the one who can define the right constraints for an agent, validate its reasoning, and recognize when context has shifted beyond its training.

The experienced operator with an agent is not replaced. They are multiplied. This is the correct framing, and brands that adopt it will attract and retain the kind of operational talent that can use these tools to their full potential.

05 - Speed of Iteration Is the New Competitive Moat.

In a market where inventory decisions, logistics choices, and supplier negotiations happen faster than any quarterly review cycle can absorb, the operative advantage is the ability to act, learn, and adjust continuously. Not in weeks. In hours. And crucially - not with diminishing returns, but with compounding ones.

An agentic system that processes every shipment, every invoice, every vendor interaction is not just resolving individual tasks. It is building a living picture of how the operation actually works - patterns, anomalies, vendor tendencies, channel behaviors - that makes every subsequent decision sharper than the last.

Dashboards display history. Agents accumulate judgment.

06 - This Is the Moment to Build, Not to Wait.

The window in which agentic AI represents a differentiated operational capability is not permanent. It is open now. Brands that treat this as a future consideration - something to evaluate in the next planning cycle - are making a timing decision with lasting consequences.

The cost of early adoption is manageable. The cost of late adoption, in a market where your fastest competitor is running with agents while you are still running on tickets, is not.

We Do Not Need More Software. We Need Thinking Systems.

I have spent years running operations that software was supposed to simplify. I have seen the promise of each successive platform generation - and I have staffed the gap between what each promised and what each delivered. That gap is a permanent feature of passive systems.

Agentic AI is the first serious proposal to close it. Not with a better interface or a smarter dashboard, but with a system that can reason, act, and improve inside the actual complexity of operations. That is not a modest claim. It is also not an empty one.

This is the work. This is the moment. The brands willing to think clearly about what these systems actually are - and what they are not - are the ones who will build operations that last.