Every supply chain executive has heard some version of the AI pitch, from a chatbot that summarizes a dashboard to an autonomous agent that promises to forecast demand, plan replenishment, and run logistics on its own. Fewer have seen either fully deliver.
We put that gap in front of three people who see it from different sides:
- Bryan Audic, Chief Data Officer at RCM, Europe's largest automotive distribution group
- Vineet Sharma, Professor and Director of the Center for Supply Chain Excellence at Saint Louis University
- Prabhat Rao Pinnaka, a product and supply chain transformation leader with experience across Fortune 50 fulfilment and warehouse operations.
Mia Guo, Product Marketing Manager at Knime, moderated the conversation. Their answer converged fast: the failures they'd seen were never about the model. They were about trust.
3 Ways Trust Failed in Supply Chain
- Trust has to be earned in stages. Most AI stays stuck summarizing because it never earns the right to decide.
- Trust requires explainability. An accurate model still gets abandoned if no one can see how it works.
- Trust needs a foundation under it. Even a good algorithm can't be trusted if the data, context, and people around it aren't there.
Here's how each one played out, in the panellists' own words.
1. Most AI in supply chain still summarizes. It doesn't decide
Ask around and you'll hear the same story: a chatbot layered on top of existing dashboards, SOPs, or ticket queues. It demos well. It rarely touches a real operational decision.
Most AI in supply chain today kind of summarizes and recommends, almost none of it actually executes. It demos beautifully. Everyone nods in the executive room, and then it never really touches a real decision.
Prabhat Rao Pinnaka, Advisory Board Member, ISCEA (International Supply Chain Education Alliance)
The gap is readiness. Prabhat's advice for where to start: pick one high-volume, unglamorous process (resolving receiving variances at a distribution center, for instance) before touching anything more ambitious.
Most organizations skip that first process and go straight for the ambitious pilot. So the AI stays exactly where it started: a chatbot summarizing a dashboard, never handed a real decision to make.
2. An accurate model still has to be understood
The most instructive story from the panel is about a model that worked, and still got abandoned, because no one could explain it.
Vineet described a lead-time forecasting model his team built that hit 80% accuracy. When business leaders asked how it worked, the answer was: "It's too complicated. Just use it."
"That doesn't go very well with business," Vineet said.
Business leaders will not use a model or any solution unless they understand what it is and how it was developed.
Vineet Sharma, Professor & Director, Center for Supply Chain Excellence, Saint Louis University
The model was accurate and was also abandoned. His fix was inclusion.
This is where close collaboration between business and the technical teams comes into picture. And that's where platforms like Knime can be very helpful, because they can do a quick proof of concept with business people involved intrinsically, versus on the outside of the project
Vineet Sharma, Professor & Director, Center for Supply Chain Excellence, Saint Louis University
In other words, let the people who'll have to defend the number help build the solution behind it.
Prabhat saw the same failure pattern play out at a large fulfilment operation, where AI auto-generated SKU-level forecasts and triggered labour and replenishment calls.
Floor managers didn't trust the confidence scores, so they quietly verified everything by hand. One floor manager told him: "In the beginning, I didn't understand what confidence scores meant, so I verified everything myself."
A recommendation people cannot see into is a recommendation they'll override, and an overridden system stops getting the corrections that make it smarter.
Prabhat Rao Pinnaka, Advisory Board Member, ISCEA (International Supply Chain Education Alliance)
3. The real bottleneck is context, data, and people
Every panellist, unprompted, converged on the same point: AI fails on foundations most companies haven't finished building.
Prabhat named it precisely: the "context graph" problem. Most operational decisions never make it into a system of record. "A lot of decisions today are made within Teams chats and emails," he said.
When AI can't see the contractual exception, the supplier restriction, or the judgment call buried in a Teams chat, it produces recommendations that are "locally optimized but globally disruptive." In other words, technically correct but operationally wrong.
Vineet framed the same gap in organizational terms. Three prerequisites he sees missing again and again:
- talent readiness ("the analyst who understands both the business logic and the model is the scarcest resource in most organizations right now")
- data maturity
- governance
Without all three in place, he argues, "the technology is irrelevant."
Bryan's experience at RCM makes the case with numbers rather than theory. His team automated a simple rule: flag vehicles too old to sell to a customer, and emailed the responsible person each morning when the rule was broken.
"Ninety percent of the time, you can just automate things," Bryan said. "People will be happy because it saves them time, and building AI on top of it wouldn't be the worst thing in the world, once the data and the rules are right.”
Bryan added: "In my opinion, you don't replace people. You replace tasks."
Most AI strategy conversations skip that distinction. Any executive who gets asked, "Can we save money with AI?" needs this distinction ready.
What a good AI strategy actually looks like
Asked for one piece of advice each, the panel didn't converge on a tool. They converged on a sequence.
Start with the business problem, not the technology. "It starts with the business use case," Prabhat said, "but do not convert it into a use-case shopping list for the tool." From there, decide, decision by decision, what the machine can act on alone and what gets escalated to a human.
He calls this bounded autonomy, and it's earned in 3 stages:
- assisted automation
- guided autonomy
- adaptive agency.
Skipping to stage three on day one is how AI projects end up in the failure pile.
Define the destination before you touch the model. "You have to decide, at least at a high level, where it is that you want to get to," Vineet said, meaning what accuracy, what outcome, working backwards from there. Bryan's version of the same discipline, sharpened by experience: "If you don't know where you will go, you will just hit the wall harder."
The common thread: it was never a technology problem
Bryan, Vineet, and Prabhat came from different industries and different levels of AI maturity, and still landed on the same conclusion. As the session wrapped, Mia Guo, Product Marketing Manager at Knime, who moderated the webinar, summed it up: "The hard part is not AI. It's not technology. It's trust." Trusting the data, understanding where a number came from, and keeping people in the loop while the foundation gets built.
That's the tension this panel kept circling back to: speed is easy to buy. Certainty has to be built: one visible, explainable, auditable step at a time. The organizations getting AI right in supply chain aren't the ones with the flashiest models. They're the ones that did the unglamorous work first: clean data, clear rules, and a workflow every person downstream can actually see into.
This is the trust gap Knime was built to close. Every example in this panel comes down to the same root cause:
- A model no one could explain
- A confidence score no floor manager trusted
- A decision buried in an email thread
Work done where no one else can see it, with no guardrails on who can act on it. That's the gap Knime was built to close.
Knime turns that work into a visible, governed workflow instead: every step from data to decision is inspectable, every output traceable back to the logic and the data behind it, and every workflow subject to the access controls and audit trail that needs to be signed off.
That combination is what makes the AI reliable in practice, not just accurate on paper: the supply chain manager who built the forecast and the executive who has to stand behind it are looking at the same trail of evidence.
Not a black box you're asked to trust. A workflow you can govern, inspect, and actually check.