01:29

Nipun Sibal: Within AWS, if you're familiar, we are Amazon Web Services. We provide infrastructure, cloud, and technology to pharma companies. Nineteen of the top 20 pharma companies work with AWS, on infrastructure and on data. Over the last four or five years, all of them have been pushing to bring new modalities and new drugs to market.

02:00

Nipun Sibal: There's a lot of commitment out there. By 2030, we expect 19 to 20 new drugs. What does that mean for operations? What does that mean for supply chain? That's where it becomes very exciting to share our thoughts. I'm going to cover three things in the next 25 minutes. One, why AI-native applications and use cases are failing early. Two, some examples of AI orchestration in supply chain. Three, how we run this at scale at Amazon.com. I'm very open to your feedback as I go. Do we have any planners in the room? Good, I'm a bit safe then.

02:57

Nipun Sibal: From a supply chain perspective, we see three different areas. We work with multiple customers, starting with the commercial supply chain, where the requirement is making data visible. That's the theme of my talk: data and AI, and going much deeper into it. Clinical supply chain is very similar. But now we're also looking at biologics, like CAR-T, where you take blood from a patient to a laboratory, mix your secret recipe, and within 24 hours you have to deliver it back to the patient. That is very complex. The temperature control points mentioned earlier all apply to these new modalities.

03:46

Nipun Sibal: I'm going to go straight into a scenario. Forty pallets of 20 mg tablets manufactured in Mumbai are heading to Edison, New Jersey, and the port is closed. The briefing is to find the cheapest way to keep Edison supplied. I know there are no planners in the room, which is why I asked earlier. So what do you do? I'm sure you've come across this scenario before, maybe for a different molecule or a different route.

04:17

Nipun Sibal: Traditionally, you go through a lot of data and a lot of emails, forwards and backwards. You access your warehouse management system. Some good planners know where the stock is sitting and call that warehouse directly. A lot of time goes into this. Some good planners will get it done in minutes. They'll find the nearest stock, or take the safer bet and send air freight to keep Edison supplied.

05:14

Nipun Sibal: Now put orchestration into this, if we have the right data. Right data, right orchestration. I fire my inventory agent, and it comes back and says Edison has 14 pallets in stock against a safety stock of 20, so it's below safety stock. The next anticipated demand is 34, so the real gap is 20 pallets. Then I fire my logistics agent. It asks around all the nearest fulfillment and distribution centers and comes back: Chicago has 22. Can you deliver? The agent recommends a carrier at a cost of $1,690. And it's up to the human to decide.

06:12

Nipun Sibal: That's the message here. You'll see more and more of this in the coming period. The decision moves from days and hours to seconds, and orchestration enables that. From Gartner's perspective, 63% of organizations don't have the data practices or the accessibility to data, including data sharing with partners, as we heard in the last session. Only 17% of supply chain organizations are pursuing AI transformation. And $35 billion is lost to temperature control and logistics deviations and excursions. That isn't a technology issue. The technology exists. It's an orchestration and data issue.

07:22

Nipun Sibal: When we work with supply chain leaders across the industry, we keep hearing a pattern of things. The key one, from a VP of planning, is that sales data, production data, inventory data, and internal and external data have no single layer. Sometimes planners have to guess, which is not a good sign in a GDP or GxP environment. There's a lot of constant firefighting.

08:07

Nipun Sibal: There's also an organizational shift coming. Traditionally you have supply chain officers, a VP of planning and operations, and a number of people in a hierarchical structure. Most of the time, these analysts work on things that don't require human judgment: status reports, data entry, emailing. The shift we see coming is a supply chain controller, a human on top, working with supervisor agents for planning, operations, and logistics, with a number of sub-agents running in the background and raising escalations when needed.

08:52

Nipun Sibal: To be clear, this is not a message about reducing headcount. It's about increasing the velocity of decisions. It's not about putting a human on top to do more work. We'll see more of that as we move toward this orchestration vision.

09:19

Nipun Sibal: A lot of companies are working toward AI orchestration in supply chain and manufacturing, but from an AWS lens they fail at an early stage. There are three challenges. One is to contextualize your data. Agents are very powerful, but they are only as good as the contextualized data underneath them. Two is to deploy agents with guardrails and observability. You need the right guardrails to make sure the agents you build give the right responses.

10:14

Nipun Sibal: Third, as you use more AI in your technology stack, you need explainability: why did the AI make this specific decision? That's going to be critical. In manufacturing, we're already seeing two key regulations coming. One is the EU AI Act, which classifies AI risks and how AI is implemented in your business. It's been pushed back, but it's coming. The other is Annex 22 of EU GMP. The current draft is very clear that you need explainability in how you use AI in your business. It's worth considering, so I'll go through each challenge in more detail.

11:09

Nipun Sibal: This is probably a familiar picture. You have many silos of data within your organization, and then external parties. Data accessibility is no longer the issue. The issue is contextualization. These sources were mostly used for analytics in the past, but now with AI you expect derived relationships. Take two factories producing one product. If one goes down, how do you communicate the impact across the supply chain in real time, meet demand across all these data sources, and give real-time suggestions?

12:14

Nipun Sibal: Here is how we solve this at AWS. At the bottom are the data sources. On top we build a semantic hub that transforms all your raw manufacturing, supply chain, and third-party data. We build a semantic data model that captures your context and your taxonomy, which is your vocabulary and how you run your processes. What you get is a knowledge graph that AI can consume. On top of that sits a governance layer, where you track how data has been transformed in real time and near real time, and how stale it is. Above that is the AI consumption layer.

13:27

Nipun Sibal: What you get is a comprehensive metadata catalog of your business, where you can interact with all your data sources much faster. There's also a continuous feedback loop that keeps improving your data, which has been a gap in the past. Going back to my scenario, you get a knowledge graph of what was produced, what the actual batch is, and the fact that the port is closed. You can quickly interact with multiple data sources and ask for the best recommendation and the historical performance of suppliers. It's the ability to overlay data from many sources and interact with it across the board.

14:32

Nipun Sibal: The second challenge is deploying agents with guardrails and observability. Here's an example. A demand signal comes in for a supply planner. We build a supervisor agent with a chain of thought, and sub-agents that go out to specialist agents. Can I fulfill this order? What inventory do I have right now? What's the supplier performance? Based on that chain of thought, you get a response on whether you can meet the demand.

15:17

Nipun Sibal: On top of this, we build observability. As you use more AI, what does that mean for token use? Which data sources is the agent reaching into? Are you providing the right responses within the right guardrails? You'll see a whole spectrum of agents. Level one is human-led, gen AI-assisted, like asking ChatGPT something and then taking an action yourself. Next is agent-assisted: the agent does the analysis and makes a recommendation, and the human takes the action. Then agent-driven: the agent does the analysis, makes the recommendation, and takes the action, with the human just monitoring. Finally there's fully autonomous, where agents act on your behalf and there's no human gate.

17:00

Nipun Sibal: My two key messages. First, because of regulation, in pharma we'll see more and more applications in the human-in-the-loop layer, where the agent gives a recommendation and you make the call. Second, as agents get to know your system, some applications might move to agent-driven. If I have a safety stock of 20 units, or 200, why do I keep asking the same person every time? Why can't the agent do it? At Amazon.com, 99.99% of our fulfillment work is fully autonomous. We don't have the regulatory challenge there, which lets us be more innovative.

18:04

Nipun Sibal: We've already seen some pharma customers using agent orchestration. I've sanitized this example, and it's the other GLP manufacturer, if that helps. In real life you wouldn't see agents with this much visual. This is for demonstration only. It's a supply chain command center. It shows the status and health of your supply chain, with multiple agents running orchestrated. A quick AI analysis of how my warehouses are doing comes back and says there's an anomaly in DC East.

19:08

Nipun Sibal: We run a further analysis across five different platforms to find the challenge. It says demand for this warehouse is 3,475 units and supply has been 4,100 units, so you're going to be out of stock. Then you run simulations. Based on a Monte Carlo analysis, at the current trend you're out of stock in three days. What's the cost impact? The recommendation is to move 8,000 units from Chicago, the nearest distribution center. A human in the loop makes that decision, and the system shows the impact and how much you've saved. We're seeing a lot of this already with some of the leaders in pharma.

20:22

Nipun Sibal: The third challenge is making AI decisions explainable and auditable. This is critical for regulators, and I'm already hearing that they're coming. They're also using AI in how they do audits, so it's a very exciting space. Let me show how we do this at Amazon.com, where we can trace every single package and how each decision was made. We deliver billions of packages a year across the world, some next day, some within two hours. There's a whole engine running behind that.

21:26

Nipun Sibal: When you place an order, we're trying to address two things. First, which fulfillment center, which is our warehouse, is closest, so we can fulfill fastest. Second, what's the cheapest way to deliver it. Behind that, there are more specific questions. What's the estimated arrival time, so we can notify the customer? What's our cost to serve? And can I fulfill this order from one fulfillment center, or do I need to consolidate across a couple?

22:15

Nipun Sibal: Let me bring this to life with a real order. No judgments, please: I bought some protein powder on a friend's recommendation. You can enter an order number and see what happens behind the scenes. When you're about to place an order on Amazon.com, it tells you when we're going to deliver. That's the promise we make to our customers, and every part of the supply chain we've built exists to fulfill it. In your case, that might be your patient promise.

23:00

Nipun Sibal: When I place this order, there are 50 variants of which fulfillment center to use and which is next. What's the stock level? How many people are working there? It evaluates that I live in north London and that London City is the nearest fulfillment center. It looks at what it costs to serve this specific item, where it sits in the fulfillment center, and the whole operation of picking it and getting it out the door. Overall, it calculates the end-to-end supply chain cost for this item. It took 11 seconds, with no human involved in our Amazon fulfillment decision.

24:02

Nipun Sibal: We've built this over many years of expertise, a fulfill-to-promise engine, and it's something we can share. It's the journey most pharma customers want to get to. To conclude, the value of AI and data comes in three parts. One is self-improving intelligence. You have data accessibility now, but you also need the right contextualization, and AI makes that faster than building it manually.

24:48

Nipun Sibal: Second is speed to operational insight. We can get decisions in seconds or minutes that used to take much longer. Finally, when you build with explainability in mind, you gain much more confidence from your regulator, and you see more accuracy in the answers as the AI improves in your systems. So that's me. Thank you so much for listening. I'm open to any questions.

25:34

Audience member: How do you connect this engine with the human factor? In the end, we still depend on a human at the beginning or the end.

25:50

Nipun Sibal: Thank you. The principle behind all of this is to augment the human. When we build these agents, whether in pharma or at Amazon.com, we ultimately augment the human process. We look at how people do a specific task today and how specific agents can improve it. We look at one process, such as in a fulfillment center, and ask how we automate it with agents. We get better as we go. It's about understanding the outcome you want and how these technologies can help.

26:49

Audience member: Let me elaborate. We still depend on a human in the background to deliver the package. We have all these computer systems and engines running, but drivers still come around the corner, and some packaging or handling of cargo can't be automated. How do you include that in your calculations?

27:28

Nipun Sibal: If I understand correctly, you're asking about the last mile. What I shared does include it. When we say we deliver end to end in two hours, that includes the time the driver takes. A driver has many packages, so we ask how to structure the load and what the best route is for this specific person. We provide that guidance, and there are specific use cases where we improve the experience for the driver, which reduces the time.

28:24

Audience member: In the presentation there was a good example of an agent coordinating delivery from the distribution center, deciding between a trucking route and an airfreight route. What's the data source for that decision? Is it external? How do you validate the quality of the data, and how does that work behind the scenes?

28:56

Nipun Sibal: When we work with pharma companies, we provide the tooling: the knowledge graph tooling and the ability to store the data. We can review data quality. You tell us what a specific data point is supposed to look like, we set parameters, and we train either the tooling or agents to check the quality and staleness of the data and how real time it is. We build a validation framework. Ultimately, pharma companies, manufacturers, and suppliers know that data better, but we provide the infrastructure. The challenge we heard this morning is how to share data between suppliers and manufacturers so that federated learning can support decision-making.

30:04

Audience member: I think we all know that the pharmaceutical industry can be slow to adopt new technologies. And with Annex 22 coming in a few years, there will be some limits on the use of AI. What challenges do you see, and how does the mindset in pharma need to change to make full use of AI in a compliant manner?

30:40

Nipun Sibal: That's a really interesting question, and my answer will be more philosophical. AI is transforming industries. At AWS we're privileged to see not just pharma but other industries, and it's transforming banking significantly. When your bank calls to ask whether you're really in this country, an AI engine is running behind that. We're seeing this happen in pharma over the last 18 months. If your digital teams build agents with the right mindset, the question is always what outcome you want to achieve. At AWS we say work backwards from the outcome. Is it reducing deviations? Is it lead time? How can you do it differently with AI? Then build with the right model framework, so that when these regulations come into practice, you can demonstrate compliance.

32:09

Audience member: Thank you for your presentation. I followed the contextualization of data, but when you delegate actions to an agent, you enter the field of processes. In logistics, these processes may not always be well documented, and may live in people's heads. How are processes defined for these agents, and how are they maintained?

32:51

Nipun Sibal: Are you thinking of the logistics processes themselves, or the processes for how we transform the data?

32:59

Audience member: My context is air freight logistics, but in general there may be a discrepancy between the documented process or SOP and the reality of day-to-day operations. How does an agent deal with that, and how do you keep it traceable?

33:14

Nipun Sibal: Absolutely. The low-hanging use cases are where you have standard operating procedures for specific processes. You can use AI to validate options such as air freight, truck, or road. If you feed in the right historical data, you quickly get an indication from AI. We're also seeing a lot of document creation and report creation, where knowledge management and AI accelerate decisions so you can focus on what matters. Route optimization is another area, based on multiple inputs, including external ones. Someone mentioned a heat wave earlier. How do you bring that data into practice and let the supply chain command center recommend what to do? It comes down to the quality of the data you feed in.

34:37

Audience member: One last question.

34:41

Audience member: Thank you for the presentation. Everybody knows Amazon, and what you showed us is general cargo. The majority also know Amazon works mostly with its own assets, though it outsources part. Have you transferred this system to a forwarder or other provider who doesn't have its own assets, so that it works for the entire industry?

35:15

Nipun Sibal: Traditionally, that's been the case. Amazon is peculiar: our applications have been built as monoliths over many years. But things are changing. In the US, we deliver directly to patients through Amazon Pharmacy, which includes PillPack. As a patient, I might need specific tablets on Monday and Tuesday and others on Thursday and Friday, and we deliver them directly. This year, Amazon Logistics also opened for outside business, so we can manage logistics on behalf of customers, though not for pharma at the moment because of the cold and temperature requirements. Some of the services that power Amazon.com come from AWS. When we talk to large pharma customers, what we provide is infrastructure, elasticity, compute, and AI models.

36:35

Audience member: Thank you.