Industry

Two things we learned from a North American e-commerce project

·7 min read

Last year, we built a decision platform for cross-ocean replenishment and fulfilment risk across two North American warehouses.

On the surface, it is a supply chain system: suppliers in China, two warehouses in LA and Houston, multi-channel sales, retailer POs, ocean booking, customs clearance, freight rates, cross-region fulfilment and transfers.

But once we started taking it apart, we found it wasn't really a "forecasting" problem, and it wasn't "a dashboard with AI" either. What it deals with is this: when information is incomplete, risks keep shifting and every action affects the next one, what should we decide today?

The experience made us more certain of one thing: using AI to write a system is no longer the hardest part. The hard part is whether the team can take in the scenarios, constraints and consequences in an industry that nobody has spelled out yet.

I want to write down the two things from this project that struck us most.

First: the model can't live after the fact, it has to live in the moment

Many numbers in a supply chain look intuitive.

When the goods will arrive. How much stock is left. What the quote is. Whether a shipment was inspected.

But when you actually make a decision, what matters most is often not the final answer but: on that day, at that moment, what could we actually see?

For example, a later-updated ETA can't be treated as information that was known at the time. The final invoice can't be used to pretend we knew the cost from the start. Inspection results can't show up early in a backtest. These look like small data engineering details, but they directly decide whether the model is fooling itself.

If a model uses information that only became known later, of course it looks accurate. But that isn't predictive power. It is just hindsight.

So in this system we paid close attention to when events happened, when data first became visible, rule versions, and the state snapshot at every decision timestamp. Not because these things are technical, but because the real world doesn't let you go back and make decisions with the answers in hand.

We see this as a turning point where AI starts to have real value: not just computing data faster, but being able to state the uncertainty clearly when the information available right now is incomplete.

Not just an ETA, but an arrival window and its risk. Not just whether to transfer stock, but what could happen under each scenario and how likely it is. Not just which option is cheapest, but stockouts, service commitments, customer relationships and downstream opportunity costs laid out side by side.

This isn't only a model capability. Behind it you need an understanding of the problem itself.

Second: people are not a sign-off before AI goes live, they are part of how the system learns

A typical question in this project: should Houston's stock be used to save orders on the West Coast?

Looking at a single order, the answer might be simple. The West Coast is out of stock, Houston has stock, so ship across regions.

But look one layer further and things change.

Sending this stock to the West Coast might save a high-margin order, but it might also leave Houston's own region short at the next demand peak. In between there are cross-region shipping costs, promised dates, retailer penalties, warehouse capacity and long-term customer value.

A mature system shouldn't just output the two words "recommend transfer".

At the very least it has to make clear:

  • Under what scenario this action is worth taking.
  • What you lose by not doing it, and what risks doing it carries to the next stop.
  • Whether it hits a hard constraint that no cost saving can offset.
  • If a person decides to change the recommendation, what they changed and why.

That is how we understand human-in-the-loop.

People are not there to click "approve" before the model goes live. Human approvals, changes, rejections and deferrals are all important signals in themselves. They may reflect customer relationships, compliance conditions, conditions on the ground, or business judgement that hasn't been written into the rules yet.

But human judgement can't just stay in chat logs either.

Which candidate options the model proposed at the time, what the person finally chose, what was actually executed, and what really happened weeks or months later all have to be kept. Only then can the system build a decision loop that can be replayed, instead of guessing from scratch every time.

We are increasingly sure that if AI is going to work in complex operations, what it has to learn is not "did a person click approve", but why the person made that choice in that scenario, and what that choice led to in the end.

In complex scenarios, the hard part was never wiring up the model

Many things can be turned into a demo very quickly now.

Pulling data, connecting a model, building an interface and having it generate recommendations: the bar for all of this really is coming down. That is a good thing.

But once you actually get into an industry, the difficulty doesn't disappear. It just shifts from "can it be done" to "do we know how it should be done".

Which data only became known afterwards and can't be used for training. Which rules are hard constraints that the model can't trade away for higher revenue. Which averages hide failures for key customers, specific routes or specific warehouses. Which actions look like they save one region but really just move the risk somewhere else.

None of these questions can be answered by a model name alone.

It takes customer insight, and it takes problem insight. Put more directly, it needs a team willing to walk into the details with the business, and bit by bit pull out, define and record the judgements that were scattered across the experience of different roles.

This is a lot of what we have been doing this year.

We don't start with an AI solution and then look for problems to fit it to. We first get a clear view of where each industry is really stuck, then decide how models, rules, data and people should be combined.

Some situations suit automation. Some should only produce recommendations. In some, the most important thing is for the system to know: there isn't enough data right now, stop and hand it to a person.

When decisions start drawing on more than a hundred data sources at once

In the end the project connected over a hundred data sources, several thousand fields and more than a thousand key factors. Demand, inventory, retailer POs, ocean booking, cross-ocean milestones, customs clearance, freight rates, external weather signals, and what was actually executed after human review all had to line up by SKU, warehouse, channel and time.

It sounds like data integration, but the difficulty was never just getting the data in. Every piece of information changes the judgement on the next action: replenish first, change the warehouse allocation, ship direct across regions, or do nothing. The system has to put inventory turnover, stockout risk, service commitments, logistics cost and contribution margin on the same decision ledger in real time, so the team doesn't need a chain of manual relays and spreadsheet calculations to know how a given shipment should move.

As a result, both inventory turnover and gross margin improved significantly. The project paid back in under a year, and at the time of writing the ROI is over 300%.

But this result was never ours alone. It came from an owner willing to invest for long-term value, and from an internal team that gave its full cooperation so that data, processes and on-the-ground judgement could be laid out and worked on together. Rarer still, when the small mistakes that are inevitable early in a rollout appeared, the team chose to fix and recalibrate together instead of giving up the first time things weren't perfect. These conditions, which seem to sit outside the system, are what let the project end with such strong results.

Closing

Looking back at this two-warehouse supply chain project, what we delivered was more than a design baseline for models and rules.

More importantly, we confirmed again something we have always believed: AI's maturity is not about how clever it looks, but whether it understands what kind of uncertainty it is facing, whether it can state constraints and consequences clearly, and whether it can work with human experience to make complex decisions more steadily.

What really creates value in an industry is not just putting AI into a system. It is putting in the scenarios and judgements the system never saw before, too.

Need Professional Assistance?

Our team are senior engineers and former executives with 20+ years in commercial product development — from apps serving hundreds of millions of users to B2B SaaS platforms serving tens of thousands of merchants. In recent years the focus has been enterprise AI adoption, from mapping the opportunity through to tuning what is live. Whether that means:

  • Mapping AI opportunities and sequencing them
  • Preparing data flows and integrating existing systems
  • Building forecasting and optimisation models
  • Monitoring and tuning after go-live
  • Test strategy and process review
  • Building and adopting test automation

We can provide professional advice and concrete solutions for teams of different scales and types. If you're facing similar challenges, feel free tocontact our consulting team.