Levi's: cutting order processing from 5 days to 10 minutes with AI

In September 2026, Jason Gowans, Chief Digital & Technology Officer of Levi Strauss & Co, put a clear number on the table: order processing went from five days to ten minutes. Behind that shift are AI agents that read over 200 order formats and write directly into the ERP.

The useful part of the story lies in how it happened.

The original idea: agents inside the system, upstream of the process

Many companies put AI in front of the process, like a new window in an old queue. Levi's chose the opposite end of the chain.

The agents work where the order becomes system data: inside the ERP. A wholesale customer sends an email, a PDF or a spreadsheet, each with its own structure. Previously, a person would read, interpret and manually re-enter everything.

Counting the formats helps understand the scale of the problem: over 200 different ways of saying the same thing, how many items, what size, for which store.

The design choice is clear. Instead of asking customers to adapt to a single form, the company taught the machine to understand each customer's format. This is an architectural decision, before it's a technology one: the value lies in the point where the agent is connected, more than in the model adopted.

From five days to ten minutes: the before and after

The figure reported on September 10, 2026 by diginomica^[1] is stark: five days before, ten minutes after.

It's worth reading carefully. This is operational efficiency, a different category from revenue growth. Those who mix the two end up promising the board one story and delivering another.

The value of a shorter order cycle shows up elsewhere: the product reaches the store while demand still exists. In retail, the right merchandise at the wrong time is worth much less, because it ends up on clearance.

There's a second effect, less visible. People who spent days re-entering order lines can now review the data, correct discrepancies and talk with customers. Time freed up only stays valuable when it's reassigned to something useful.

Stitch: the app born from an internal hackathon

The second story concerns the stores. It's called Stitch and it's the app store staff use on the sales floor.

Its origins matter: it came from an internal hackathon. People who understood the problem built the tool to solve it, rather than waiting for a turnkey solution from an outside vendor.

The measured result is eight more points of customer satisfaction. An app born from a hackathon that moves an experience metric by eight points says something specific about the organizational model that produced it.

People who work in stores know where the day breaks down: the missing size, the slow return, the customer stuck at checkout. That knowledge usually stays locked in. A hackathon pulls it out and turns it into code, then the real work begins, rolling it out across thousands of locations with training, support and maintenance.

Product search: two points of conversion

The third thread touches app and website: product search and matching. The reported result is a 2% increase in conversion rates.

Two percentage points sound small in the abstract. Applied to a global brand's traffic, they become a significant number, which is why internal search remains one of the most profitable levers in digital commerce.

The logic is simple. A customer searching for "black straight jeans size 32" has already decided to buy; the system's job is to remove friction between that intent and the cart.

Matching adds a layer of strategy. The company aims to sell the entire wardrobe, so showing the right shirt next to the right jeans shifts the weight of the transaction. The source reports the conversion figure; other effects on average order value remain outside the published narrative.

The question a board should ask: what does a Copilot return?

The most interesting moment of the interview arrives on uncomfortable ground: the return on assistants distributed to people, Copilots.

Measuring it is difficult. An agent processing orders offers a clean before and after; an assistant helping someone write an email or prepare a meeting leaves scattered traces across a thousand different tasks.

Gowans responds with a comparison: email and Excel. Who would try today to calculate the ROI of Excel? Yet remove it from a company and the day stops.

The answer is honest and still deserves scrutiny. The comparison holds for tools that become infrastructure; the difficulty lies in knowing in advance which ones truly will, while the bill arrives every month. A board does well to ask the number, and equally well to accept that for part of the spending, the answer arrives later.

The scale that makes the case serious

These results matter in the context where they live. The group operates in approximately 120 countries, with about 3,300 stores and 50 million Red Tab loyalty program members.

A demo works anywhere. A rollout across that surface is another discipline: languages, currencies, tax systems, distribution contracts, local purchasing habits.

The ERP program is multi-year. Europe reaches completion in spring 2027, so today's narrative describes a construction site still open.

This is the piece missing from nearly every enthusiastic AI presentation. The agents on orders are only as good as the system that hosts them: ERP modernization and data foundation is the base, everything else is built on top. Those who skip the foundation get elegant pilots that stay pilots for years.

What you can take away

The case offers four readings, depending on where you sit.

• Founders and CEOs of SMBs: look at the data entry point, where customers send you information in their format.
• CTOs and Heads of Product: the agent writes into the system of record, so the real constraint is integration, before the model.
• Board and investors: ask for before and after on a metric with a denominator, like five days versus ten minutes.
• Managers and team leads: the internal hackathon is a product tool, beyond just a company culture event.

The common thread is one: AI has made it economical to manage variety. For decades, companies imposed standards on customers because interpreting a thousand formats cost too much in labor hours. That constraint has loosened, and every process built around it deserves a review.

There's also a lesson on sequencing. First the data and the system, then agents on high-volume processes, then tools for those in the store. The order of moves explains much of the result.

The open question

There remains the question that applies to any organization: at what points in your process are you still asking people to adapt to the machine's format?

An honest answer usually produces a short and embarrassing list. That list is your backlog.

The next verification arrives in spring 2027, when Europe closes the ERP cycle. Until then, these numbers should be read for what they are: real measurements on a real perimeter, within an ongoing program.