ICON 2026 – Blue Yonder’s Chris Burchett on the AI reality in supply chain

As I’ve noted across much of my ICON 2026 coverage, Blue Yonder has used its annual user event to have a refreshingly honest conversation with both its customers and its partners – including CEO Duncan Angove saying that ‘the agent is the app’, that systems integrators are about to become a product feature (ouch), and that planning as a function is an artificial human construct that machine intelligence won’t need. 

But vision has to be built by someone and at Blue Yonder, a significant part of the engineering answer sits with Chris Burchett, SVP of Generative AI, who has spent the last 18 months building the AI layer that the cognitive platform depends on. I sat down with him towards the end of ICON to understand the build logic behind the decisions Blue Yonder has been making – and ended up having what was an equally direct take on what customers need to hear.

Not renting the intelligence layer

The model factory announcement I covered earlier this week, outlined how Blue Yonder is working with NVIDIA to create smaller, domain-trained models that focus specifically on supply chain intelligence. 

Burchett said that this partnership was driven by what’s happening in the frontier model market, whereby coding has driven token demand through the roof, and model providers – hardware constrained – are continuously retraining and quantising their models to manage capacity. As Burchett put it, the behaviour of those models ends up different from what it was before, and complaints about performance have followed. For an enterprise software vendor building reliable AI products for supply chain customers, that creates a specific problem around consistent quality. He described how months of hard engineering work – context engineering, harness engineering, building the evals, keeping agents grounded – can all end up being sensitive to model changes. He said:

It becomes a whack-a-mole situation if they keep racing to meet demand while we’re trying to serve the enterprise market.

The case for owned intelligence is therefore an engineering response to a genuine reliability problem. A smaller domain-trained model that inherently understands supply chain, Blue Yonder’s API surface area, and the decision processes involved in planning, transportation and warehousing doesn’t carry the same context engineering burden. It doesn’t need to be told what it doesn’t need to know. Burchett described the hybrid architecture that results:

I can have the frontier model handle the human interaction because it’s really good at that. It determines human intent and hands off to the smaller model, which does all the detailed supply chain tasks. Now I’ve taken ownership of the intelligence layer, because I’m no longer beholden to just what the market provides.

I have a hybrid architecture – use the big models for as much as you can, but where you can also drive better behaviour, simplify your development, and make it more reliable and consistent, that’s the owned intelligence strategy.

I asked whether this is structurally similar to what some other vendors are doing – using frontier LLMs to capture intent, but coupling this with deterministic workflows to execute at the core to ensure consistency and compliance. Burchett said that deterministic workflows break when supply chains produce situations slightly outside the normal flow – a new order from an important customer, a disruption that doesn’t follow the usual pattern. A model trained on supply chain reasoning can be the shock absorber in those moments, where a hard-coded deterministic workflow cannot. 

A new kind of job

One of the more interesting points in our conversation was around the changing nature of operating models across organizations due to the impact on AI – something that CEO Duncan Angove picked up on during his keynote. 

Burchett said that the time humans currently spend interacting with the mechanics of software won’t be needed in the way it is now. That frees people up, he argued, to serve customers more relationally and to focus on the higher-order decisions – providing a “premium”. What he believes will be new is the class of jobs that comes with agent-scale operations – what Burchett called the “agent wrangler” or “agent manager.” He said:

If you think about the control plane needed for enterprise agents, there’s going to be jobs for: how do you manage those agents? How do you give them new skills and capabilities as your business changes? How do you react to model updates that are coming from vendors constantly?

This connects to what Angove told me about the artificial constructs of organizations – the planning departments and cycles that exist because of human limitations. Burchett is describing the workforce component of this, whereby the new roles are shaped by the demands of operating at agent scale rather than human scale. That’s a big shift.

Trust before autonomy

Which led us nicely on to the hurdles to enabling an autonomous supply chain – namely, building trust. Burchett separated this into two categories. The first is technical – reliability, evaluations, the ability to undo agent actions, scaling underlying systems to handle agent-level interaction volumes. These are engineering problems, he said, that will get solved. Agents, as Burchett noted, don’t stop for coffee breaks and the underlying systems have to keep up.

The human trust problem is subtler, and in many ways is the harder problem to solve. Burchett’s argument is that trust in an agent has to be earned through accumulated experience in exactly the same way it is with a person – demonstrated competence, consistency, and memory. He said:

If the next time you interact [with an agent] and the agent has forgotten everything, your trust is lower than it would be if they remembered you, because now they’ve shown that you matter to them. If the agent is consistently right, or even recommending things you hadn’t thought of, that adds value to the relationship and builds trust. I don’t think you get to autonomy without a path to build that trust.

My take

As noted above, what I noticed across many of my conversations at ICON this week is Blue Yonder’s willingness to have the harder conversation with customers. When I asked Burchett about adoption challenges, his answer was that customers see the need, but what they’re struggling with is strategy – how this affects their workforce, what to plan for, build versus buy, which platform to place their bets on. 

The supply chain landscape many of these customers are operating in involves often dozens of siloed systems, each with their own data model and local metrics – each warehouse optimizing for its own overtime figures, not for corporate KPIs. Getting customers to understand that what Blue Yonder is proposing is not a warehouse project in isolation, but a strategic direction change, is a material part of the work.

Burchett described a Customer Advisory Board moment last year when customers started saying to each other ‘this is different’. And on building trust, I think a vendor that is willing to have these harder conversations – on jobs, on incentive structures, on what the operating model actually looks like on the other side – will fare far better than a vendor that focuses simply on selling as much software as possible.