When you force supply chain leaders to pick a side, it turns out most of them already have.
At Manifest 2026, FourKites staged three courtroom-style debates on the biggest questions in supply chain AI. Former executives argued assigned positions (harder than it sounds — you have to do the homework), cross-examined witnesses, and let the audience vote. In the first trial alone, 90% of the room voted that AI should act autonomously on routine decisions, not just advise.
Ninety percent. In a room full of practitioners, not analysts.
And yet most of their companies are still running manual exception management, still requiring human approval on routine decisions, still staffing teams to do work that AI agents are already handling at companies like Church & Dwight and US Cold Storage. The conviction is there. The deployment isn’t. That tension ran through every session — but so did practical guidance on how to close it.
Brad Blizzard of Bridgestone Americas painted a picture of a company dealing with millions in annual waste from slow response times, siloed systems, and manual bottlenecks — and when asked if that company was using autonomous AI, he said they weren’t sure where to start. He was arguing for the prosecution, the side that said AI should advise, not act. But his testimony landed differently next to companies already in motion. A Church & Dwight supply chain leader said in a video deposition that their AI agent Tracy saved 25 hours a week. US Cold Storage hit 96% accuracy on automated appointment scheduling across 600-plus shipments.
Both sides agreed that you should start with high-volume, low-risk decisions with clear business rules like appointment scheduling, detention alerts, and document collection. These are the tasks where the decision logic is well understood, the financial exposure per decision is limited, and the frequency is high enough to prove value fast. Nobody argued for handing strategic customer exceptions to an AI on day one.
The debate was about whether you earn governance through endless pilots or through controlled deployment with guardrails. Running a fifteenth pilot won’t produce governance. That takes intentional work — sitting down with IT, operations, and end users to define the guardrails before you turn anything loose.
Watch the full debate on autonomous action vs. decision support.
The second trial asked whether shipment tracking is enough. The defense argued it is — master what you have, don’t chase complexity.
Then James Glover of CHEP described a retailer with 250 SKUs on a delayed truck. Before orchestration, they’d air freight the entire load for $13,000 or do nothing and risk $50,000 in stockouts. With orchestration, the system identified that only 8 SKUs were critical. They moved those 8 for $1,800. As Luis Solana put it to the jury, that’s $10,000 saved on a single exception. Even if only a fraction of your monthly exceptions carry that kind of cost, the savings add up fast.
The jury voted 80-20 for orchestration. But the more useful thing here is the diagnostic. Say your team is regularly expediting entire shipments because nobody can quickly tell which SKUs are critical, that means that you’re overpaying on every exception. Or a delayed truck triggers hours of cross-functional investigation before anyone can act. That’s time and money spent assembling information, not making decisions. Either one is a sign that tracking alone has gone as far as it can go.
Watch the full debate on tracking vs. orchestration.
The third trial — build your own AI or buy — was the only one that was close. The jury voted 56-44 to buy, and when someone asked about hybrid approaches, roughly half the room raised their hands.
That’s because this is the question where conviction meets budget, org charts, and IT roadmaps. The build side made a legitimate case for control, data ownership, and long-term flexibility. Their witness described a failed vendor implementation and a successful in-house rebuild. The buy side countered with a CPG company that spent 18 months building an internal control tower with adoption stuck at 30%. FourKites deployed prebuilt agents in 12 weeks and the company recovered millions in working capital in six months.
Both sides, stepping out of their assigned roles, agreed on a practical framework: build what’s core to your competitive advantage, buy what gets you speed on everything else. Say you’re a retailer whose edge comes from how you run promotions — maybe you build an agent to optimize those workflows yourself. But if appointment scheduling and document collection are eating your team’s time, buying a purpose-built solution gets you to value in weeks instead of quarters. The mistake most companies make is treating it as an all-or-nothing decision when it doesn’t have to be.
What’s most important is deployment. Get that experience. Progress over perfection.
Watch the full debate on build vs. buy.
The votes were decisive. The examples are real. Coca-Cola cut customer response times from a 90-minute SLA to seconds. Arrow Electronics is using an AI agent to collaborate with 1,800 suppliers in multiple languages. US Cold Storage automated dock scheduling with over 90% accuracy.
Ninety percent of the room said act. The companies already acting are proving them right.
The prosecution wants you to fear autonomous action because it could fail. I’m asking you to fear inaction because it’s guaranteed to fail.
The only question left is what the other companies are waiting for.
All four sessions are available to watch on demand.
