Shattering the AI pilot trap

Artificial intelligence has dominated supply chain conversations for several years now. Yet for all the enthusiasm around generative AI and agentic systems, organizations that have moved passed pilot projects remain small in number.

“I would argue that most enterprises are not yet adopting AI and going into production,” said Aadil Kazmi, head of AI at Infios, told Supply Chain Management Review during a recent discussion on the state of AI deployment at the Manifest conference.

While generative AI remains “top of mind for executives,” he noted, scaling it into day-to-day operational decision-making is far more complex than launching a proof of concept.

The pilot trap

Kazmi described two approaches for understanding AI adoption. Within Infios’ own customer base, some deployments have progressed rapidly from pilot to production in “a matter of two quarters.” But across the broader ecosystem, progress has been slower. The disconnect, he suggested, often begins with how organizations frame the problem they are trying to solve.

“A couple of weeks ago, one of our own customers came to us and said, ‘Hey, we’re looking to optimize how we build loads.’ And they kept on approaching it as a gen AI problem,” he said. Instead, Infios redirected them to a classical machine learning optimizer already embedded in its platform.

The example, Kazmi explained, is reflective of the approach many are taking when it comes to AI. Find a problem to solve rather than attacking problems and identifying the correct solution. Leaders feel pressure to “do AI,” often before validating whether AI models are even the right solution.

Kazmi offered a framework grounded in principles. “What changed three years ago?” he asked. Machine learning, natural language processing and OCR have existed for years, but the shift occurred when large language models could finally understand unstructured data and communication.

“If we agree that the fundamental change three years ago … was machines can now understand unstructured data using LLMs, then the framework should be … to look at workflows that already depend on unstructured text and data as the key blocker for automation,” Kazmi said.

In other words, AI agents are best applied where human reasoning over emails, voice calls, and free-form communication has historically prevented automation, not where traditional optimization or RPA already suffice.

Embedded agents vs. bolt-ons

Another design choice that shapes success is where agents live. Kazmi described two philosophical approaches. The first is vertical—automating a specific workflow such as driver check calls. The second is horizontal—connecting multiple systems (ERP, TMS, WMS) to create cross-functional automation.

Equally important is whether AI is embedded into existing software or layered on top as a standalone interface. “When we deploy our order entry agent … it natively lives inside [the system],” he explained. Users forward emails to the agent without leaving their core platform. By contrast, a bolt-on interface forces users to toggle between systems, increasing friction and limiting adoption.

Beyond chatbots: From data to action

Kazmi cautioned against stopping at conversational AI, though. “Chatbots are fantastic,” he said, particularly for querying reports, but frontline supply chain teams do more than retrieve information—they execute.

“Their core workflow isn’t just querying data, it’s taking actions,” he said.

To generate measurable ROI, AI agents must interpret data but also perform tasks such as updating order status, canceling shipments, or scheduling appointments. That requires integration into transaction systems and access to APIs, not just access to reports.

Guardrails, governance and autonomy

As AI agents move closer to operational execution, governance becomes critical.

Kazmi acknowledged the risks of agents deviating from business rules but said that preventing such failures is a matter of architectural discipline.

He introduced the concept of L1 versus L2 autonomy. “L1 being specifically gated to the task at hand … versus L2, which is true decision autonomy,” he explained.

Infios recommends starting with L1 task automation, where ROI can be substantial. “Some of our customers are seeing ROI of upwards of 60% with just L1 automation,” Kazmi said.  

Decision autonomy can follow later once governance, confidence thresholds and human oversight are mature enough to handle it. He also emphasized reinforcement learning through human feedback to prevent agents from institutionalizing exceptions.

Ultimately, he advised managing AI agents much like human employees: monitor performance, set guardrails, and manage risk tolerance.

Purposeful innovation at scale

One of the most difficult balancing acts for software providers is managing demand.

“We have a ton of customers coming to us with every single sort of problem,” Kazma acknowledged. But rather than building bespoke agents for every request, Infios maps workflows across its customer base to identify scalable use cases, in a sort of Venn diagram approach: identify common workflows across hundreds of customers, then prioritize those with the greatest repeatable value.

This “purposeful” strategy shortens deployment cycles over time, Kazmi said. Early pilots create learnings that compound across subsequent implementations.

Measuring ROI: Follow the workflow economics

Post-deployment, the question shifts from excitement to economics. ROI can vary, Kami pointed out, with some use cases generating immediate returns while others take months due to required training and change management.

To measure value, he recommends starting with workflow-level cost analysis.

“If you’re looking at a workflow to automate, start with what is the current process? How many humans are touching this … what is their average salary or hourly rate?” Kazmi suggests. From there, calculate cost per execution and compare it to the AI-driven alternative.

While time savings are often cited, Kazmi argues that cost-to-serve is the clearer metric, particularly in high-volume operational environments.

The human dimension

Finally, scaling AI requires more than technical integration. Infios provides pre-built templates for common supply chain workflows, reducing the burden on non-technical users. But it also pairs deployments with professional services support and change management.

AI agents may automate tasks, but humans still define guardrails, validate outputs and adapt processes, Kazmi noted. As with previous waves of automation, upskilling, not replacement, remains central to sustainable impact.

From hype to operational discipline

If there is a common thread in Kazmi’s perspective, it is this: AI success in supply chain is less about model sophistication and more about workflow clarity. Select the right problems. Embed agents where work already happens. Start with constrained autonomy. Measure ROI at the task level. And treat governance as a design principle, not an afterthought.

The industry may still be in the pilot-heavy phase. But the path to production is becoming clearer, one workflow at a time.

SC
MR