Supply Chain Analytics on AWS That Reaches Production

Key takeaways

• The model is rarely what holds these programs up. At the foods company in this guide, the first real problem would be two warehouses that count stock differently, and no forecast can fix that.
• Some services in older AWS designs are gone or closed to new customers. Amazon Forecast stopped taking new customers in 2024, and AWS now calls Amazon Connect Decisions the next generation of AWS Supply Chain, so we'd recheck any plan written before this year.
• Whether to buy, build, or mix the two shapes everything after it. We'd expect a company like the foods maker to buy for its dry goods and build for its fresh-cut line.
• Food traceability rules alone can reshape the data model. Under FSMA 204, the foods company would have to hand over a lot's traceability records within 24 hours of being asked, and we'd plan for that in the first sprint.
• We'd plan on 12 to 16 weeks to get the first predictive use case live, provided the source data can be reached and every phase closes with a signed acceptance.

We'll use one illustrative company throughout this guide, a mid-market packaged-foods maker. It has three plants in the US, two distribution centers, about 2,800 SKUs, and a small sales office in the Netherlands. Orders and inventory sit in SAP, each distribution center has its own warehouse system, the retailers send point-of-sale numbers over EDI every week, and nobody can see where a truck is without logging into the freight broker's portal. Its largest customer is a national grocery chain that fines every delivery that turns up late or short, and the fines have gone up for three quarters running.

We've built data and AI systems on AWS since 2008, a lot of them for companies where a wrong number ends up in a board deck or in front of a regulator. Over that time we've watched plenty of supply chain programs stall, and it almost always happens the same way. The dashboards ship and the forecasting pilot gets a good reception in the demo, but somewhere between the data team and the planners who were supposed to use it, the whole thing fades out.

Budgets have moved faster than results. According to Gartner, 67% of supply chain digital investment is now allocated to AI, yet 55% of chief supply chain officers say they aren't clear on the return. We wrote this guide for the people signing off on that money, the COO, the CFO, and the VP of supply chain who'll live with whatever gets built, and we'll keep coming back to the foods company to show how each decision plays out.

What Does Supply Chain Analytics on AWS Actually Deliver?

Supply chain analytics is the use of data to make supply chain decisions earlier. It covers everything from plain reporting on what happened to forecasts and recommendations about what to do next, and the ladder further down shows how those stages build on each other. Most companies we meet are good at the reporting and patchy at the rest. On AWS, we'd pull the ERP, warehouse, transport, supplier, and IoT data into one governed foundation, run forecasting models and AI agents on it, and look for the difference in OTIF, inventory, and freight spend.

Suppose a truck to the grocery chain arrived 60 cases short last month. In the reactive version, the fine notice showed up three weeks later, customer service forwarded it to planning, and two planners spent an afternoon working out from old emails that the salad-kit line had run behind. With anticipating analytics in place, that order would have been flagged on the Tuesday, when the line was already slipping and the carrier still hadn't confirmed a pickup. A planner would have had a recommendation to split the load with two days to spare. The table below sets the two side by side, using the foods company's own week.

 

Why Do Most Supply Chain Analytics Programs Stall Before Production?

In our experience, it's rarely the model. Gartner asked chief supply chain officers about this in a survey published in April 2026, and 56% of them said integrating AI with legacy systems and processes was a major challenge, while half admitted they didn't have the in-house expertise to implement and manage it. McKinsey's numbers from late 2025 are starker still, with only 19 percent of companies saying they deploy AI tools at scale.

Say the foods company tried a forecasting pilot two years ago. A data scientist pulled a year of SAP orders into a notebook, the model beat the planners' spreadsheet in a demo, and everyone was pleased. Then it stalled. The two warehouse systems counted stock differently, so the inventory numbers never reconciled, and nobody had been given the job of getting the forecast into the planners' Monday routine, so they kept using the spreadsheet. The security review found that no one had thought about access controls or retention until then, which cost another quarter. The integrator had billed by the hour for discovery and mapping, so there was no particular pressure to finish, and since none of the phases had a signed exit, the go-live date just kept sliding.

We see those five problems again and again: source systems that were never connected, no owner for the move into planning, governance that arrived late, a delivery model paid by the hour, and no exit criteria. Gartner's research points the same way, finding that organizations with successful AI initiatives invest up to four times more in data and analytics foundations as a share of revenue, which is why we treat source readiness as week-one work.

What Does an Anticipating Supply Chain Look Like in Practice?

We describe it as a ladder with five rungs. Each rung shortens the gap between something happening in the supply chain and someone doing something about it, and most of the mid-market and lower-enterprise companies we meet are on the second one, with reasonable reporting, a bit of root-cause analysis, and a couple of forecasting experiments that never made it into the plan. The pressure to climb is growing. In the World Economic Forum's Global Value Chains Outlook 2026, 74% of executives said they view resilience as a driver of growth.

It helps to follow one short shipment up the ladder. At the bottom, the foods company's planners would learn about it on Monday from a dashboard in Amazon Quick that shows last week's fill rate. One rung up, an analyst could query the order history in Amazon Athena and see that most short shipments came out of the same distribution center. The middle rung adds prediction, where a forecast on SageMaker AI, built on Chronos-2 or AutoGluon, would warn every day which SKUs looked likely to run short. The fourth rung adds optimization and scenarios, so the system could suggest which distribution center should cover the order. At the top, an AI agent on Amazon Bedrock AgentCore would make that change itself, within an approval limit, minutes after the risk showed up.

Which Use Cases Pay Off First?

We'd start where the data already exists and a miss is easy to price. For most manufacturers, distributors, and life sciences companies, that leaves five candidates, and at the foods company we'd rank them roughly like this.

Demand forecasting on AWS would come first, mostly because the data's already sitting there in two years of SAP orders, the retailers' weekly point-of-sale files, a promotions calendar, and the weather. A good forecast here gives a range, not one number. That matters for salad kits, where a cool week can leave a distribution center full of product with six days of shelf life.

Next we'd look at shipment ETA prediction, which pulls carrier, telematics, and port data together. When people talk about real-time supply chain visibility, this is usually what they mean, and it's the piece that would have caught that short truck in time to reroute it. Supplier risk scoring is third. It rolls lead-time drift, quality escapes, financial signals, and exposure further down the chain into one score that procurement can act on.

The last two depend on the first three working. The first sets a service level for each SKU at each location, which is usually where the working capital hiding in slow movers turns up. The second is where supply chain AI agents earn their place: an agent reads the exception, checks stock and alternate sources, writes up a fix, and passes it to a planner if the impact is over a set limit.

Gartner thinks agent-led exception handling will go a long way, predicting that 60% of supply chain disruptions will be resolved without human intervention by 2031. We'd treat that number with some caution. Still, companies that have governed agent workflows running before then will have a decision history and a set of controls that latecomers will have to build from scratch. When we scope AI agents, the first thing we pin down with the client is what an agent may do alone and what needs a person's signature.

Plants add another layer. Our manufacturing practice connects legacy MES, ERP, and SCADA systems to cloud data platforms, so forecasting and supplier risk analytics can see what the lines are actually making. At the foods company, that's how a planner would find out the salad-kit line was running two hours behind before the truck showed up to load.

Which AWS Services Belong in a Supply Chain Analytics Architecture?

The diagram below shows the pattern we deploy most often. We keep the foundation deliberately conservative, since that's the layer where audit findings, cost overruns, and model drift tend to begin.

Two choices matter more than any single service on it. We'd keep the core data in an open table format, Apache Iceberg on Amazon S3 Tables, so you can change query engines later without moving anything. And we'd organize the supply chain data lake on AWS as data products that the business owns, with the demand, supply, logistics, and supplier teams each standing behind the quality of their own data. A lot of people would call the result a supply chain control tower. We'd hold off on that name until it can recommend and carry out actions as well as show status.

How Does ERP, Warehouse, and EDI Data Get Into the Platform?

This step decides most timelines, so here's what it would involve at the foods company. SAP orders and inventory would come in through the AWS Glue zero-ETL integration for SAP, which reads through SAP's OData interface and picks up changes as often as every 15 minutes. The two warehouse systems sit on ordinary relational databases, so AWS DMS change data capture could replicate them without anyone touching the applications. The retailers' weekly point-of-sale files are X12 EDI 852 documents, and AWS B2B Data Interchange would turn them into JSON in Amazon S3, and it can also produce the 856 advance ship notices going back out. Carrier and telematics feeds would arrive through Amazon Kinesis or AWS IoT Core.

Getting the data in is the easier half. At the foods company, one warehouse counts in cases and the other in single units, a few hundred SKUs have two different descriptions, and the grocery chain's store numbers don't match the ship-to codes in SAP. We'd sort out item, location, and customer master data before anything else, add AWS Glue Data Quality rules to the pipelines, and give each data product an owner who signs off on what its fields mean. A forecast trained on mixed units would still produce neat-looking numbers, and they'd be wrong.

What Happened to Amazon Forecast and AWS Supply Chain?

AWS has retired or repositioned several services that still show up in older supply chain reference designs, and a lot of guidance online hasn't caught up. If you're evaluating an Amazon Forecast alternative or reviewing a vendor's proposed design, we'd check it against this list before approving anything.

 

A roadmap that still names any of these was probably written from an older playbook, and it's far cheaper to fix that on paper than after the first pipeline is built.

How Does Demand Forecasting Work on AWS Now?

With Amazon Forecast closed to new customers, demand forecasting on AWS usually runs on one of three paths. Analysts who want a no-code tool get SageMaker Canvas. Teams that want accuracy without building models from scratch start with Chronos-2, the time series foundation model Amazon released in October 2025, which handles known future inputs such as scheduled promotions and weather forecasts and is available through SageMaker JumpStart and AutoGluon. Teams with a data science group and unusual demand patterns train their own models with AutoGluon on SageMaker AI.

For the foods company, we'd start with Chronos-2 at the SKU-and-distribution-center level, with promotions, price changes, and temperature forecasts as inputs, because salad kits sell differently in a heat wave. Before anyone relied on it, we'd backtest it against two years of history, check it separately on new items and on slow movers that sell in fits and starts, and then run it in shadow mode next to the planners' spreadsheet for a few weekly cycles. If its error is lower on their own SKUs, the planners tend to switch without being told to.

Should You Buy Amazon Connect Decisions or Build on AWS?

This is the most consequential call in the program, and it's often made too early, by whoever has the strongest vendor relationship. Amazon Connect Decisions, generally available since April 28, 2026, is what AWS calls the next generation of AWS Supply Chain. It comes with demand, supply, and onboarding intelligence, plus AI teammates that triage exceptions and run root-cause analysis, and it's priced by the type of AI teammate rather than per user. A custom build on SageMaker AI, Amazon Bedrock AgentCore, and Amazon S3 Tables gives you full control of the models and data instead. Plenty of companies end up somewhere in between.

Here's how we'd lay the options out for a leadership team:

 

How Do You Decide Which Route Fits?

If your planning looks a lot like your competitors' and speed matters more than being different, we'd lean toward buying. If the planning logic is part of how you win, you run several plants with complicated constraints, or a regulator expects to trace each forecast back to its source records, building is usually worth it. Hybrid covers the middle, with routine workflows on the packaged product and custom models where the margin is made.

Our guess is that the foods company would end up with a hybrid. Nothing about replenishing its dry goods is unusual. The fresh-cut line is different, with a shelf life of days, FSMA 204 traceability, and a customer that fines by the truck, so that's where we'd spend the custom-model budget. Whichever way it goes, the data foundation is the same piece of work, which is why teams that skip it to get a quicker demo tend to find themselves starting over a year later.

 

How Much Does Supply Chain Analytics on AWS Cost?

We can't price your program from here, because supply chain analytics implementation cost depends on how much data you have, how fresh it needs to be, and how many source systems have to be connected. What we can do is show what drives the bill, so you can test any proposal you've been sent against it.

 

Spending more rarely fixes it. Gartner's September 2026 research found that 83% of surveyed organizations had spent at least $3 million on automating supply chain planning, yet it predicts that only 5% of organizations automating planning will make at least 10% of their planning decisions autonomously by 2030. Most of that money went into tools and services hours while the decisions stayed manual. Gartner is more encouraging about adoption, though, and expects companies that rightsize change management to get twice the ROI on their AI initiatives by 2030.

Where Does the Budget Come From?

In our experience, a good share of the budget for predictive supply chain analytics is already inside your IT spend. Legacy data warehouse licenses, on-premises Hadoop clusters, and integration middleware that bills hours just to stay alive are all candidates for reallocation. On the Plume engagement, moving a telecom SaaS provider's analytics onto an AWS lakehouse cut its analytics run-rate by 58%, and the move went from kickoff to go-live in 18 weeks.

We'd also put security risk in the same spreadsheet. IBM's Cost of a Data Breach study for 2026 found the average breach now costs $4.99 million, against $4.44 million the year before, and a platform holding supplier contracts, prices, and customer data is worth attacking. That's why access control and lineage are priced into our cost model from day one, since adding them after go-live costs more.

How Do You Keep Supply Chain Analytics on AWS Compliant in the US and Europe?

Compliance tends to add the most unplanned months to these programs, usually because the architecture was approved before legal and quality got a look at it. The foods company is a good example of why we'd involve them early. Its fresh-cut salad kits are on the FDA's Food Traceability List, so it'll need FSMA 204 records once the FDA starts enforcing the rule, which won't happen before July 20, 2028. Its Dutch sales office also holds personal data on customer contacts and drivers, which brings GDPR in.

Only two rows of the table below, FSMA 204 and GDPR, would apply to the foods company. A drug distributor or a defense supplier would need different rows, which is why we've left them all in, each with the question an auditor tends to ask and what we'd show in reply. This isn't legal advice, and your counsel should have the final word on what applies to you.

 

Two rows have moved recently. Omnibus I cut CSRD back to companies with more than 1,000 employees and €450 million in turnover, which would probably leave the foods company out, and CSDDD now applies only to much larger companies, from July 2029. On product passports, the first one that's mandatory is the battery passport under the Batteries Regulation, while passports for other product groups under the Ecodesign Regulation will come later.

What Guardrails Should AI Agents Have in a Regulated Supply Chain?

Say an agent at the foods company noticed that Thursday's truck to the grocery chain was going to be 40 cases short and suggested pulling stock from the other distribution center. The move was under its approval limit, so it went ahead. Had the gap been 400 cases, a named planner would have got the decision instead. Either way the recommendation stays on file with the records and model version behind it, which is what an auditor would ask for later. We'd also set up the dull parts on day one: the agent's own scoped credentials instead of a shared service account, a log of every action and override for audit and retraining, and a rule that personal data from the Dutch office stays in an EU Region.

Those five guardrails are fixed rules in our enterprise AI strategy work. It's the same pattern we used for a renewable grid operator, where agents reason over real-time signals and act within bounded authority.

What Does a Realistic Roadmap for Supply Chain Analytics on AWS Look Like?

We break a mid-market or lower-enterprise program into phases, and nothing moves to the next phase until the customer has signed off on the last one. The week numbers in the chart move with data readiness and how many source systems there are, so they're a starting point for negotiation.

If the foods company kicked off in January, the first three weeks or so would go on agreeing baselines for OTIF, fines, expedited freight, and forecast error for every SKU at each distribution center. By around week eight, SAP, both warehouse systems, and the EDI feeds would be in Iceberg on S3 with access controls in place. The Chronos-2 forecast would then run in shadow mode until about week twelve, and the first agent would take on one kind of exception, short shipments to the grocery chain, with a planner approving anything above the limit.

Each of those exits is signed by a named approver, which is central to how we work, and a forward-deployed engineer owns the judgment calls along the way. We'd push hard to keep shadow mode in the plan, because it's the step teams most often drop under deadline pressure, and it's what gives planners proof on their own SKUs and auditors a clean comparison. We followed a similar path with a manufacturer whose supply chain was good at responding and slow at anticipating, and the supply chain modernization story describes the real-time data and agentic platform we built for it on AWS. Time to value there was measured in weeks.

Which KPIs Prove You Have Moved From Reactive to Anticipating?

These are the measures we'd agree on at the foods company during discovery and report at every phase exit. Finance and operations already track most of them somewhere. The hard part is getting everyone to agree on one number for each before anything changes, because otherwise every result gets argued over later.

 

How Would You Build the Business Case?

With a CFO, we'd build the case from numbers the company already tracks. The foods company's numbers might look like this: $2.1 million a year on expedited freight, about $900,000 in OTIF fines to the grocery chain, and 38 days of supply on hand. With roughly $180 million a year in cost of goods sold, one day of supply is worth about $490,000. Cut five days and around $2.5 million comes back as working capital, which is arithmetic any CFO can check on the back of an envelope.

We'd agree up front which of those numbers each phase is meant to move, and we wouldn't treat any of them as a promise, since the size of the gain depends on where you start. It's also the reason we measure forecast accuracy, OTIF, inventory, and freight before anything ships, and then report against those same numbers at every exit.

What Should You Ask a Supply Chain Analytics Partner?

If you're bringing in supply chain analytics consulting help, these are the questions we'd want the foods company's leadership to ask us, in their own words:

• Our planners have already ignored two forecasting pilots. Who on your team will sit with them until this one is part of their Monday routine?
• We have SAP, two warehouse systems, and a broker portal nobody has integrated. Which of those are in phase one, and what happens to the date if one is messier than you thought?
• Our board has been told this goes live next fiscal year. Is that date in the contract, next to the fee? And if it slips, who pays?
• After you leave, who owns the pipelines, the models, and the code? Will they still run in our own AWS account?
• If the grocery chain disputes a fine six months from now, how do we prove what the system knew and recommended on the day?

 

Our own answers are in our delivery commitment, which puts a fixed date, a fixed fee, and customer-signed acceptance at every phase into the contract.