Start Small Now. Build the Foundation as You Go.
I get the same question every week from supply chain leaders. Where do we start with AI? How do we begin? Maybe it is better to wait for someone else to figure it out, then follow?
These are not bad questions. They are honest ones. And they explain why AI adoption in supply chain is still so low compared to almost every other corporate function.
There is too much hype on the market. Startups preaching autonomous supply chains. Vendors claiming 99% accuracy on day one. Self-healing networks. Closed-loop planning. End-to-end visibility. Most of it is aspiration sold as reality. For a leader trying to make a real decision, it is overwhelming.
So leaders do what reasonable people do when they cannot tell the signal from the noise. They wait. They watch their peers. They commission another assessment from another consultancy. They quietly defer the decision because the risk of choosing wrong feels bigger than the risk of not choosing at all.
I understand the instinct. I think it is the wrong call. Here is why.
Take AstraZeneca, the company everyone keeps citing. They announced their self-healing supply chain at London Tech Week in June 2025. You would think they nailed it. But did they? Is their supply chain really self-healing already? No. AstraZeneca's own leadership has described their current state as "foundation rather than destination." It is a five-year, $60M+ program. The vision is real. The capability is being built. They are not finished yet.
But here is what AstraZeneca actually did, and what everyone else should learn from. They did not wait until they had perfect foundations to start. They announced the vision, started the program, and are now doing the foundational work alongside the AI deployment. That is the real lesson. Not "build a five-year master data plan first."
The companies waiting for perfect conditions will spend five years watching everyone else move ahead, and will never catch up.
Where to Start?
McKinsey recently published an illustrative matrix of 29 supply chain AI use cases, plotted on two axes: use case impact and use case maturity. It is worth studying. It quietly says the thing most vendors will not: high-ROI use cases are not where you should start. Why? Because the foundations are not ready yet.
Plot the use cases everyone is talking about against those two axes and a clean divide appears. The glamorous, high-ROI use cases are the ones the market has not figured out how to deploy reliably yet. The deployable ones are narrower and less exciting. Here is the split.
| Use case | Maturity | Verdict |
|---|---|---|
| Cross-functional orchestration | Low | Wait |
| Autonomous closed-loop planning | Low | Wait |
| End-to-end visibility with agentic alerts | Low | Wait |
| Digital twins for network scenarios | Low | Wait |
| AI/ML demand sensing for established lines | High | Start now |
| Agentic replenishment for stable demand | High | Start now |
| Inventory allocation and order promising | High | Start now |
| Transportation sourcing automation | High | Start now |
| Labor forecasting and DC scheduling | High | Start now |
| Dynamic slotting and pick optimization | High | Start now |
Source: McKinsey's illustrative supply chain AI use case matrix, plotted on impact versus maturity.
The use cases everyone is pitching as the future of supply chain are the ones you cannot deploy today.
What do the high-maturity use cases have in common? Narrow scope. Clear data sources. Documented workflows. Tested edge cases. The kind of use case where a company that has done the foundational work for that specific process can deploy AI and trust the output. The ones McKinsey labels high impact but low maturity are the ones companies want to start with, because the ROI projection looks beautiful. They are also the ones with the lowest probability of success today.
What "Ready" Actually Means
Ready sounds clear and means nothing until you make it specific. For one AI use case, it comes down to four things.
Quality master data
For demand sensing on one product line, the SKU, customer, and location masters for that line need to be clean. Just the slice this use case touches, not all your data.
Process documentation
If you do not know exactly how decisions get made today, you cannot replace or automate them tomorrow. Document the workflow, including the edge cases and the workarounds nobody talks about.
Tested edge cases
What happens when the supplier ships short? When the customer cancels last minute? During a promo uplift? If you cannot answer these for your process, you cannot expect the AI to.
Defined ownership
Who is accountable when the model gives an answer? Who owns the underlying data quality? If the answer is everyone or nobody, the project fails no matter how good the model is.
Not glamorous. Not getting written up in HBR. Also non-negotiable. And here is the practical implication: you do not need your whole supply chain to be AI-ready. You need one slice of it ready for one use case. Pick the slice. Get it ready. Deploy. Then move to the next slice.
How to Pick Yours
You can start with McKinsey's high-maturity zone as your shortlist. Demand sensing, agentic replenishment, inventory allocation, transportation sourcing, labor forecasting, dynamic slotting. These are the use cases the industry has figured out how to deploy. They are your candidate pool. Then filter the shortlist against your own organization.
Run the shortlist through four filters:
- Clean data - where your data is already in good shape, not a mess you would have to fix first.
- Narrow slice - one category, a few markets, two quarters. Protect it from scope creep.
- Fast feedback - signal in weeks, not once a month, so you can actually iterate.
- Modest ROI - a credible $5-8M target, not the $50M number your CFO is hoping for.
Run your shortlist through all four and the answer usually narrows itself to one or two real candidates.
The ROI Conversation
Once you have picked a use case, the next conversation is with your CFO. And this is where most leaders lose the room.
The first project is not about this year's P&L. It is about building the capability to impact next year's P&L. You are buying knowledge. You are buying a working data foundation. You are buying a team that has actually deployed something and learned what works. The savings come from the second and third projects.
ROI is what you measure at the end. Not what you optimize for at the start.
Most leaders avoid this conversation because it does not fit the standard ROI template. Promising transformational ROI on a foundational pilot is what causes pilots to be judged as failures even when they succeed at their actual purpose.
Why So Many Pilots Go Wrong
You have probably heard this from peers who tried and failed. Here is the honest take on what happened.
- They chase the ROI number. The CFO asks what the return will be. The leader picks the use case that promises a big one. Halfway through, the team realizes the data is not clean enough to support the model. Results are inconsistent. The pilot gets paused and rebranded as phase one of a longer journey. ROI was the selection criterion when it should have been the thing you measure at the end.
- They pick what the data cannot support. Glamorous use cases like autonomous replenishment and end-to-end network optimization need data from across functions. Most companies cannot produce a consistent SKU master across sales, planning, manufacturing, and logistics, let alone a harmonized demand signal. Picking a use case that requires data you do not have is the most common failure I see.
- They need cross-functional data the silos never reconciled. You can have the best master data in operations. If marketing, finance, and sales are working on different definitions of demand, no AI model will reconcile that for you. The model will surface the inconsistencies, not solve them.
- They choose the wrong vendor. Not the one promising 99% accuracy from day one. Not the one who says AI will figure out your data problems and fix them. Not the one promising an autonomous supply chain. The one who forces you to focus on the foundational work first.
Start small and do this right, and six months in you have:
- One use case running in production, with a model the team actually trusts
- A documented data foundation for that use case that you can extend
- A team that has genuinely built, deployed, and iterated on an AI model
- A clear read on where your data is good enough and where it is not
- Credibility with your CFO and CEO to fund the next project
- A realistic shortlist of which use cases are next for year two
The first step is unglamorous: get one slice of your data and processes AI-ready, then deploy against it. The leaders who start small now will run self-healing supply chains in five years. The ones waiting until their foundations are perfect will spend those five years watching the gap to AstraZeneca widen. See how blueclip gets your data and processes AI-ready →
