Q&A: Why Pharma Supply Chains Are Turning to AI

Pharmaceutical supply chains are under mounting pressure from multiple directions at once — geopolitical volatility, cold chain complexity, fragmented planning systems, and a pace of change that traditional forecasting tools were never built to handle. For some companies, the response has been reactive at best, patching together siloed systems and relying on institutional knowledge. Natalia Andreyeva, vice president of market strategy at 4flow, has spent her career trying to change that, offering insight on how legacy planning infrastructure is failing and where artificial intelligence (AI) is beginning to fill the void.

In a conversation with Pharmaceutical Commerce, Andreyeva identifies some of the industry's most persistent failure points: demand forecasting tools she says carry a 30% to 40% error rate on average, planning environments where handfuls of siloed systems routinely fail to communicate, and batch quality events that cascade through the supply chain with little real-time visibility. She makes a case for why logistics and disruption management are where AI is generating the clearest near-term value, and why, with roughly 70% of pharma products moving through temperature-controlled lanes, any logistics delay is not just a cost problem but a potential patient safety issue.

Andreyeva also describes a fundamental shift in how pharma companies are now approaching AI investment. The ask is no longer a point solution for a specific function — it is, increasingly, a request for organizational transformation designed to sustain AI adoption for decades to come. And while she sees automation eliminating much of the laborious, repetitive work that bogs down quality and compliance teams, she is clear-eyed that human oversight, especially in regulatory compliance, isn't going away anytime soon — and may not be for some time.

PC: What Supply Chain Problems Are not Being Solved Well by Traditional Planning Systems, and How Does AI Address Them?

Andreyeva: There are many. First of all, traditional forecasting tools that pharma is typically using are highly inaccurate. There is a 30% to 40% average of accuracy of that forecast, mainly because the environment changes too quickly. Regulatory approval shifts, demand shifts, the prescription habits and trend shifts, and many other events and outbreaks are just some of these changes. The traditional systems do not allow for quick adjustments of those plans. And usually, the signals are not even getting into the forecast. The planners are usually using something manual. They use other spreadsheets as an output to changes, and then they input those changes into the planning systems at best. It's just inflexible from that perspective.

The other thing that is important but is not really solved is long lead times in production. Because in pharma, there's always a lead time baked into the entire process, from manufacturing through distribution and delivery to the customers and patients. Even with those lead times, production is still sometimes unpredictable, because you can have a bad batch, especially for newer products where it's not ironed out yet.

It is very hard to adjust inventories on the fly— to understand how a change in production will impact your safety stock, what it would mean to you, what markets are impacted and so on and so forth. We're actually getting a lot of requests lately for batch tracking — the batch quality concerns, if that wasn't released, then what will happen to my supply chain? What is the impact? Do I need to worry about the safety stock? Do I need to worry about the shortage of the product? They want to see that impact, and they consider it a disruption, along with operational disruptions and geopolitical disruptions and so on and so forth.

And then there's also disintegrated planning. There are simply too many tools, and every tool has its own information in it. SAP IBP is a very popular tool for pharma, so they usually use it, but they don't necessarily always connect it to the other tools they have. When they use IBP on one hand and TM on the other side, they don't always speak to each other. So you create a lot of these little silos that a portfolio can easily be about ten systems at the same time to collect information. And when things change, either in execution or in planning, they don't know about that. It disconnects. That also creates a situation of misalignment that leads essentially to waste in pharma. It's typically 1% to 3% of all products, but it's still a really huge number that goes to waste because of shelf-life expirations.

Where Are Pharma Supply Chains Seeing the Most Measurable Impact From AI Tools Today?

I think the biggest and most measurable impact for pharma, as well as for some other industries, is logistics and risk and disruption management, because you can immediately see the impact on the supply chain. From a logistics perspective, it's always cost. For pharma, in addition to cost, it's product availability and quality of the product, because 70% of products are cold chain drugs. If there's a delay in logistics or transportation, it will very often mean that there's going to be a temperature deviation, which will have to be evaluated and will cause further delay, if not a discard of the product. Risk and disruption go hand-in-hand with this, because a lot of logistical problems are caused by delays that are caused by disruptions in the supply chain from an operational perspective.

AI tools and AI platforms have an ability to measure that impact immediately and apply it directly to the customer lanes. That's very important, because there was a belief in pharma a very long time ago that all lanes operate the same. A lane can be impacted by a geopolitical event — that's clear — but the impact of that for each organization would be completely different, because they'll be transporting a very specific product with very specific conditions, temperature requirements, and cost to a very specific market. It may vary from insignificant to very significant. That impact is what they're looking for, and this is what our platforms can now give them. And that's the most valuable piece — because you know what is impacted, how much it's going to cost you, and what it would mean if you don't react to that disruption.

Which Disruptions Are Still the Hardest for Pharma Companies to Predict and Respond to Quickly?

The disruptions that are truly unpredictable are anything I call external disruptions to the supply chain — geopolitical events, weather conditions, port congestions, freight capacity unavailability impacting air freight, truck freight and so on. Those events happen because something else happens in the world and the global economy, and that impacts the supply chain first. Then it ripples down to each business in each industry.

On the other hand, there are also operational disruptions that happen internally in the supply chain. In my opinion, those are a lot more predictable. Pharma is a great example because there is better documentation. Data patterns of suboptimal choices could be traced and highlight areas for improvements. The patterns of operational efficiencies can be seen in supply chain network. You can absolutely predict what's going to happen, because sometimes it's also due to seasonality — high seasons or low seasons for transportation of products in different geographies. You can predict those disruptions because you can see them in data. And you can also prevent them

I was doing a product discovery years ago and I was talking to a QA director at one of the big pharma companies — someone who had been with the company for more than fifteen years, sitting in Europe in a global distribution quality management role. He told me, "I know that I have to investigate shipments that go through the Port of Miami." I said, "Why the Port of Miami?" He said, "I've been getting too many deviations lately from the Port of Miami. We know it's a problematic port, but I'm getting more than normal." I asked, "Is it documented anywhere?" He said, "No. I know it just because they all come to me, and they keep checking them. So for me, it's a red flag once I count more than five." That's the pattern recognition and why doing it manually is still a problem today. It is people's experience, it is people's knowledge, and if you've been with a company for a very long time, that's great. But for anybody who is new and doesn't know operations, that is not data-driven.

Another challenge is that all of that is documented in free form. We looked at records with another former customer — also a big pharma — and we looked at the reasons for deviations or delays, and temperature condition changes. Everyone would use their own language, their own descriptions, and it would be impossible to compare. You just know you're comparing apples with oranges. That was also feedback directly from the team. That's why we implemented a batch release platform for them.

Over the Next Five Years, How Do You See the Role of AI Tools in the Supply Chain Expanding, and Into Which Areas?

Based on what we are seeing with 4flow for life sciences and pharma, the next wave of AI adoption will go beyond logistics. There are too many requests right now focused on demand sensing, regulatory compliance, network design monitoring, and planning. And what we are also seeing as a trend is that the boundaries of those functions are blurring — they kind of overlap and blend together. There is no clear cut anymore between where planning ends and where execution starts. It needs to be a system that is constantly learning and sending signals, informing the impact chain from planning to capacity, to logistics, to execution, to delivery, to quality and so on and so forth.

The other thing that is very interesting is that companies have already stopped treating AI as a point tool. They're not doing that anymore. They're looking at major transformations. The requests we are currently getting are not, "let me implement your solution and use it for logistics." The requests are, "help us transform the organization so that we can set up for AI adoption for the next thirty years." And that is a cultural shift and change for them as well. They're not looking at periodic planning or a tool that can do something tactically. They're looking at systems that can combine all of this in one, so that they can have a clear view of their supply chain performance and scale it. Because with the speed of change in the world, especially in pharma, they're not keeping up with human resources alone. They need measurements and metrics to be more agile and more flexible with the changes.

Which Supply Chain Decisions still Require the Most Human Oversight?

That one is easy — anything related to regulatory compliance or quality, in my opinion, will still require human oversight for a while. And I'll give you an example. A couple of years ago, I had the pleasure of working at a different company where we launched a first-of-its-kind platform that could do the evaluation for stability of a batch all the way through to the product — not just at distribution logistics, but all the way through the chain from manufacturing to the patient. Even back then, we could automate the entire process completely human-free. You would go through the entire chain, the system calculates, evaluates, gives you a recommendation, re-evaluates, recalculates, and gives you the release to use or not use a product.

And anybody I met with — more than twenty pharma companies — would tell me, "if we let the system do that, who's going to be responsible if it got it wrong?" That was the reason. The only check they were doing was literally a human person looking at the recommendation and saying, "okay, yes, I approve. I agree with this." And then it would still go and execute all the way through and inform other systems.

Automation was the point, because a lot of those checks are so laborious. It takes hours, and sometimes days and weeks, to do these checks and approvals — and that's not even people's main job. It actually frees up human capital from doing wasteful, repeatable tasks, but it still needs that small check and oversight from humans to confirm nothing is missing. I can also see that we could automate the oversight itself, but then somebody will still have to oversee the oversight — the idea of humans overseeing agents. I think it's still a little bit dangerous to allow complete automation in this area.