ASCI Connect

Supply chain leaders keep hearing the same promise: we're moving from reactive to predictive. AI will see things we can't. Planning will become proactive. Disruptions will be anticipated, not managed after the fact. And supply chains will move faster than ever.

I buy the direction of travel, but I'm not convinced we're there yet.

In fact, it's possible that for most organizations, AI has primarily helped us make decisions based on historical data faster, not fundamentally differently. What many companies call predictive today is often just accelerated forecasting. Models trained on historical and transactional data do a better job turning yesterday into tomorrow.

Is that useful? Absolutely. Is it truly predictive? Only if tomorrow looks like yesterday.

Once disruption enters the picture-weather events, geopolitical friction, labor shortages, regulatory shifts, supplier instability-the ability to model the future based on historical results is no longer guaranteed. If you need proof, revisit the COVID years.

Truly predictive supply chains anticipate risk before it shows up in the order, the shipment, or the inventory signal. But if our systems are built primarily on historical data, we are not anticipating risk, we are just reacting to it faster.

Which raises the real question supply chain leaders should be asking: Are we predicting the future, or just summarizing the past more efficiently?

Why historical data alone isn't predictive

If an AI model is trained mostly on historical performance, it can generate probabilities, but it remains anchored to patterns that have already occurred. As long as those patterns repeat, the model performs well.

But supply chains rarely operate in stable conditions for long.

When the world shifts through climate disruption, geopolitical events, labor instability, or upstream supplier shocks, historical patterns lose reliability. And when that happens, so does any AI model dependent on them. This is why I don't think it's accurate to describe many of today's systems as predictive at all. They are optimized for continuity, not uncertainty.

Prediction requires more than pattern recognition. It requires an understanding of exposure.

The missing ingredient: exposure data

Most organizations are very good at measuring performance. Far fewer are equally focused on measuring exposure. I spoke with Cindy Elliott, who leads Esri's business industry sector teams, at the recent NRF Retail Big Show in New York City. She emphasized that source and origin data are critical if AI systems are expected to model future scenarios with any degree of confidence.

"If you don't actually get the source and the origin data correct and the live feed data correct, AI isn't going to produce what you [expect]," Elliott said.

That comment stuck with me. AI systems can only reason over the data they're given. If that data doesn't include upstream supplier visibility, environmental conditions, or live signals beyond tier two suppliers, then predictive scenarios become educated guesses and their quality will vary widely.

Elliott offered the example of Nestlé's coffee supply chain. "They don't own the coffee farms. In fact, they don't even often own the distribution to get the raw material to the roaster," she said. Yet Nestlé invested in tracking environmental conditions across more than 100,000 growers to understand where land is under stress and how to respond long before "the bean didn't show up."

That's exposure-based thinking. And it's foundational to any predictive system.

Digital twins: promising, but not widespread

This is where digital twins come into play. At their core, digital twins represent the real supply chain-its physical assets, flows, constraints, and interdependencies. Yes, they rely on historical data, but unlike traditional models, they allow organizations to alter inputs, test non-historical scenarios, and observe how disruption might propagate across the network.

This is where predictive AI can genuinely add value.

But there's a catch: adoption is still limited. PwC's 2025 Digital Trends in Operations survey found that only 21% of respondents say their companies use digital twins, even though 97% of those users report that digital twins create value.

That gap matters because you can't predict future behavior if your system can only replay the past. Digital twins provide the sandbox where AI can explore "what if" scenarios, including those the organization has never experienced.

AI without a digital twin tends to optimize hindsight. AI combined with a live, flexible model of the supply chain has the potential to anticipate what comes next.

What predictive actually looks like in practice

In our March issue of Supply Chain Management Review magazine, authors Saravanan Venkatachalam and Arunachalam Narayanan offer a useful example of how this works in practice.

They describe a consumer packaged goods company operating with stable forecasts and balanced inventory and nothing in its S&OP forecast cycle suggesting risk.

"Five days into the month, however, the company's monitoring AI agent detected something unusual," they wrote. "Order intake and point-of-sale data in one region were trending above plan. External signals pointed to a competitor stockout and an unplanned regional promotion. Individually, none of these signals warranted action. Together, they raised the probability of a service-level breach to over 80% within two weeks."

Instead of waiting for the next planning cycle, the system reassessed the forecast, evaluated corrective options, and recommended a targeted inter-distribution center transfer. After approval, the action was executed, preventing a stockout with minimal cost.

That is what predictive looks like: early signal detection, scenario evaluation, and timely intervention before the problem materializes. It is also a scenario that could have been created inside a digital twin so if the situation arose, corrective actions would have already been mapped out.

The real takeaway for supply chain leaders

There is absolutely a role for historical data in predictive systems. But if we're serious about building supply chains that can anticipate disruption, not just respond to it, we need more than faster models.

We need exposure data. We need location and source-level visibility. We need systems flexible enough to model scenarios that haven't happened yet.

That requires a layer of data discipline and model adaptability that AI alone can't solve. The predictive future isn't something you install. It's something you earn.