Lenovo & Levi’s Cracked AI at Scale. Neither Started with AI.

Welcome to Between the Hype. A biweekly newsletter on where enterprise systems and AI actually intersect. Not the keynote version. The version you need when you’re back at your desk.

According to the IDC/Lenovo CIO Playbook, only 4 out of every 33 enterprise AI proofs of concept graduate to production. That’s an 88% failure rate. Separate research paints a similar picture: 78% of enterprises have at least one AI agent pilot running, but only 14% have scaled one to organisation-wide use.

So the question isn’t whether AI agents work. It’s why almost nobody can get them past the demo.

This issue, I’m looking at two companies that did. They’re in completely different industries, on completely different technology stacks, operating at completely different scales. But they arrived at the same answer. And it has almost nothing to do with AI.

Levi’s: 1,000 agents, one ERP

Levi Strauss walked on stage at Sapphire and mentioned they have more than 1,000 AI agents in production. Not in a pilot. Not in a sandbox. In production, across the business, handling real work.

Jason Gowans, Levi’s Chief Digital and Technology Officer, was refreshingly blunt about how they got there. The 1,000 agents didn’t start with an AI strategy. They started with an ERP consolidation.

Levi’s was running nine separate ERP systems. Nine. Different processes, different data models, different ways of doing the same thing across regions. Anyone who’s worked inside a multi-ERP environment knows what that looks like.

So they went all-in on RISE with SAP. S/4HANA Fashion, running on Microsoft Azure. Over the course of several years, they retired more than 90 legacy systems, standardised more than 80% of their global business processes, and put more than 2,600 employees on a single platform with a common data set. They also trained more than 4,000 employees hands-on with AI.

That’s not an AI project. That’s a multi-year transformation programme. It’s not sexy, but it’s the part that made everything else possible.

As Gowans put it: “Standardisation and agility don’t stand in opposition. Standardisation is what allows us to move with agility.”

Here’s the use case that got the most attention: wholesale order processing.

Levi’s moves a lot of product through wholesale. About 80% of those orders already flow through automatically — standard EDI, clean data, straight-through processing. No issues there.

But the remaining 20% is where things get challenging. Orders from smaller customers — submitted through handwritten notes, emails, unstructured documents. The kind of thing that needs a human to interpret, validate, cross-reference, and key in manually. That process used to take two to five days per order. With the agents built on top of SAP, it takes 20 to 30 minutes.

The agents work because the gaps were closed first. One system and one data model. Standardised processes. Then the agents had something reliable to operate on.

Levi’s — The 20-minute upgrade

There is another a detail in the Levi’s story that deserves attention.

SAP system upgrades in most enterprises take around 48 hours. That’s the industry norm — and anyone who’s lived through upgrade weekends knows it usually takes longer. Levi’s co-developed a process with SAP that completes an upgrade in approximately 20 minutes.

Twenty minutes. For an ERP upgrade. On a platform that just migrated 14 countries across East Asia Pacific and China in a single wave — with the Andes, Brazil, and Europe still to come by mid-2027.

This only works because of the clean core. No custom code blocking the upgrade path. No bespoke workarounds that need manual regression testing. The upgrade is fast because there’s nothing non-standard getting in the way.

Levi’s — It’s not just SAP

One thing I find interesting about the Levi’s story is how many vendors are involved.

The back-office runs on SAP. Last November, Levi’s announced a partnership with Microsoft to build what they’re calling a “super-agent” — an orchestration platform embedded in Microsoft Teams, running on Azure. One entry point for employees, with specialised sub-agents across IT, HR, and operations.

And then there’s STITCH — a store-associate AI tool built on Google Cloud with Gemini. Originally a hackathon idea, now deployed in 70+ US stores. It gives frontline employees on-demand access to product information, procedures, and training materials.

So that’s SAP for back-office processes. Microsoft for employee orchestration. Google Cloud for in-store AI. Three platforms, each doing what it does best. This is the Joule-Copilot story from Issue #3 playing out in practice — except the reality is even more multi-vendor than the keynotes suggest.

Lenovo: seven years of data work before the AI

Now here’s a story that caught my attention yesterday, and the reason this isn’t just a Levi’s issue.

Harvard Business Review just published an article by Robert Handfield, a supply chain professor, about how Lenovo built its AI-powered supply chain. The core argument is one that should sound familiar by now: Lenovo spent years integrating operational data before building an enterprise-wide AI architecture. Most companies trying to do the same thing are starting with the technology. Lenovo started with the data.

They began in 2017. Not with an AI initiative — with a data integration project. The platform they built is called Supply Chain Intelligence, or SCI. Over the next seven years, they connected more than 800 individual data sources covering almost 80% of all their supply chain data inputs. The system now handles over 1,500 data-related tasks daily across procurement, logistics, manufacturing, and fulfilment.

Think about what that means. Seven years of connecting, cleaning, and standardising data across a supply chain that spans more than 30 manufacturing sites in 11 countries, producing around four devices per second for 180 markets. That’s not a quick win. That’s a long, unglamorous infrastructure programme — the supply chain equivalent of Levi’s retiring nine ERPs.

Lenovo — What it actually delivered

The results are significant. According to Lenovo’s own reporting, SCI helped reduce manufacturing and logistics costs by around 20% while improving service levels. Decision-making cycle times dropped by 60%. And the HBR article cites a 25% increase in inventory turns.

But the number that gets attention is the production planning one. Lenovo built an Advanced Production Scheduling system — developed in-house by a team of 10 experts in six months — that cut production planning time from two hours to two minutes. They’ve since deployed it across more than 20 facilities, with a 24% increase in production line capacity and a 3.5x improvement in on-time deliveries. That’s not a pilot metric but operational at scale.

And here’s the part I think is particularly telling: Lenovo is now selling this capability externally. They’ve packaged it as a product called iChain and are deploying it to customers outside their own industry. Yili Group, Asia’s largest dairy company, is among the first. The logic being: if the data disciplines and algorithms work for discrete electronics manufacturing, the core approach translates to continuous manufacturing too.

When your internal supply chain becomes a product you sell to other companies, that says something about how far ahead of the curve you are. Gartner seems to agree — Lenovo ranked 8th in the 2025 Supply Chain Top 25 and holds the number one spot in Asia Pacific for the fourth year running.

Levi’s & Lenovo — The parallel

Levi Strauss is a 173-year-old American apparel company running SAP. Lenovo is a global technology manufacturer that built its own platform. Different industries. Different geographies. Different technology stacks. Different scale.

Though, the playbook is the same.

Levi’s consolidated nine ERPs into one, standardised more than 80% of their processes, and retired more than 90 legacy systems — over several years — before deploying a single agent. Lenovo integrated 800+ data sources and spent seven years building a unified data platform before layering AI on top. Both companies invested heavily in the boring, unsexy foundation work that nobody writes keynotes about. And both ended up with AI running at a scale most organisations can’t get past the pilot stage. That isn’t a coincidence. It’s a pattern.

Why 88% fail and these two didn’t

I could leave this as two case studies. But the reason I’m writing about both is the gap.

88% of AI pilots fail. These two companies have AI running in production across their entire operations. What’s different?

It’s not the AI. The root causes of pilot failure are well documented: integration complexity with legacy systems, inconsistent output quality at volume, no monitoring tooling, unclear ownership, insufficient domain data. Every one of those is an organisational problem, not a technology problem. And every one is something both Levi’s and Lenovo addressed before deploying AI at scale — by doing the consolidation, the standardisation, and the data integration first.

The Sapphire customer stories reinforce this. Lockheed Martin called their programme “the largest transformation investment in the company’s history.” ExxonMobil talked about simplifying their organisational structure and addressing years of fragmented data. Aeropuertos Argentina went from a clean-core migration in 2023 to a production AI agent in twelve weeks, with projected savings of 16% in direct costs and 45 tonnes less CO2. None of them started with AI. All of them ended up there.

And the HBR article on Lenovo makes the same point from the academic side. As Handfield argues, most companies deploying AI in their supply chains start with the technology before understanding their data. Lenovo took the opposite approach. The years of data work weren’t a delay — they were the strategy.

What this means for your roadmap

If you’re reading this and thinking “we’re nowhere near Levi’s or Lenovo” — that’s fine. Most organisations aren’t. Both companies had multi-year head starts and the leadership structure to execute. But the lessons translate regardless of your stack or industry.

Start with the foundation. If you’re running multiple ERPs, fragmented data sources, or heavily customised systems, that’s your bottleneck — not the AI tooling. The agent will only be as good as what it can see and act on. Lenovo spent seven years on this. Levi’s spent several. There’s no shortcut.

Take clean core seriously. Aeropuertos Argentina went from clean core to production agent in twelve weeks. Levi’s gets ERP upgrades done in 20 minutes. Most organisations without a clean core never make it past pilot. If you’re planning an S/4HANA migration, build toward clean core from the start. If you’re already on cloud, audit your customisations.

Own your data integration. One thing that stands out in the Lenovo story is that they built SCI themselves. A team of 75+ researchers, including 22 PhDs. They didn’t outsource the understanding of their own data. Whether you build or buy, the organisation that understands its own data best is the one that gets the most out of AI.

Don’t skip the boring work. Retiring more than 90 legacy systems isn’t exciting. Connecting more than 800 data sources isn’t exciting. Standardising business processes across dozens of countries isn’t exciting. A 20-minute ERP upgrade is — but it only happens because someone did the boring work first. Production planning in two minutes is impressive — but it took seven years to get there.

The most important number at Sapphire wasn’t the more than 200 agents or the 50-plus Joule Assistants. It was the one hiding in the industry data: almost nine out of ten AI pilots never reach production. Levi’s and Lenovo are in the group that did. And the reason isn’t better AI. It’s better plumbing.

If this issue was useful, forward it to someone navigating the same decisions. That’s how Between the Hype finds its audience.

Dive deeper: platform-by-platform guides

Everything mentioned above — the embedded capabilities, the activation steps, the honest limitations — is covered in detail in my practitioner guides. Each one walks through what’s actually available today, how to enable it, and where the gaps are.

Executive overview: the strategic decisions that matter →