Why We Rebuilt a Client's ERP Around an AI Layer

In February, we finished the biggest project Arvension has shipped to date. We rebuilt an ERP system for a mid-sized manufacturer. Not a migration. Not an upgrade. A fundamental rebuild with AI as the core engine instead of an add-on.

The client came to us with a problem that looked straightforward: they were losing $2 million a year to inventory mismanagement. Their current ERP—a SAP installation from 2015—was technically sound. The database had no corruption, the workflows were well-designed, the team knew how to use it. But the system couldn't see what the team could see.

Here's what I mean by that.

The Problem We Inherited

A human buyer at this company looks at a requisition for 500 units of Component X. They think: "We don't need 500. It's February, slowdown is coming, we have 200 in stock already." But the ERP doesn't think that way. The ERP sees that stock is below the reorder point (set to 750). So it approves the order. Nobody's being negligent—the system is just following rules written two years ago during a different market condition.

Multiply that by 100 requisitions a day. Multiply that by the fact that nobody's looking at all 100 approvals, so half of them sit in people's inboxes for weeks. Then multiply it by seasonal swings, supplier quality issues, and the fact that demand signals are coming from Shopify, not just from the ERP.

When I say their inventory was mismanaged, I mean they had $6 million in capital sitting on shelves that they didn't need, and parts shortages in the warehouse for components that would have arrived if someone had caught a pattern.

The team was good. The ERP was fine. But there was no bridge between human understanding and system execution.

Why a Typical Fix Doesn't Work

Our first instinct was to do what everyone else does: add an AI forecasting layer on top of SAP. Feed it historical demand, train a model, surface predictions.

We prototyped it. It worked okay. Accuracy was decent. But we ran into the same problem that every enterprise runs into: the AI understood the data, but the system didn't use that understanding.

You generate a forecast saying demand is going to drop 30%. The ERP still processes reorders at the same rate because the reorder rules haven't changed. Now you need a human to manually adjust the reorder points. You've just moved the problem. You haven't solved it.

So we pitched the client on something different. What if we rebuilt the approval workflow so that AI understanding was baked into the decision-making, not bolted on top of it?

The Architecture We Built

We kept the core SAP database and core financial modules. But we stripped out the approval workflow and replaced it with what I'll call an "AI-assisted ERP layer."

Here's how it works:

When a requisition comes in, instead of hitting approval rules, it hits our AI model. The model ingests: historical purchase data, current inventory, demand signals from all connected systems, supplier reliability metrics, cash flow position, seasonal patterns from the past three years.

The model produces three outputs:

First, a recommendation (approve, reject, modify quantity).

Second, a confidence score and the reasoning. This is critical—we always show the human why the system decided something.

Third, an action trigger. If the model is highly confident (95%+), it auto-approves. If it's between 60-90%, it routes to a human buyer with the reasoning. If it's below 60%, it goes to a manager because something unusual is happening.

The human approval is still there. But now the human isn't drowning in approvals. They're only seeing exceptions. And when they do review something, they have context.

What Changed

In the first 60 days:

• Average approval time dropped from 4 days to 4 hours.
• Auto-approval rate hit 68% (the ones that didn't need human review).
• Manual rejections of the model's recommendations went from 12% in week one to 3% by week eight (the model was learning).
• Excess inventory dropped by $1.8 million.

More interestingly, the team started seeing patterns they couldn't see before. The AI was catching seasonal swings they thought they knew about but had actually gotten numb to. It was flagging suppliers who had subtly gotten less reliable. It was catching demand signals from channels that nobody had connected to the ERP yet.

The CFO told me: "For the first time, I can see what's actually driving our inventory cost."

Why This Isn't Just a Tech Project

Here's what matters: this rebuild required the client to change how they think about their ERP. Not just use a new system. Think differently about it.

Before, the ERP was the source of truth. You queried it, you got answers.

After, the ERP is the engine, and the AI is the sense-maker. The system is reasoning about your business, not just recording it.

That shift is cultural. It took three months to get the team comfortable with the idea that an AI was making purchasing decisions. But then the team realized: the AI was making better decisions than they were, not because it's smarter, but because it wasn't subject to recency bias, fatigue, or missing context.

The Investment

This project cost the client $340,000. They make that back in excess inventory savings in about 2.5 months. After that, it's pure cash.

But I want to be honest: this isn't the kind of project that works if you're not ready for it. You need clean data. You need systems talking to each other. You need a team willing to let go of some control.

Most clients aren't ready yet. But the ones who are? They're going to run circles around their competition this year.

The next wave of ERP isn't about adding AI features. It's about rethinking what an ERP system even is. And the clients who figure that out first won't need to catch up later.