Stop Guessing, Start Asking: What ML Demand Forecasting Actually Fixes

I was talking with a merchant last month who told me their team spends every Monday morning manually adjusting reorder points in their ERP. Three people, four hours, every week. They have years of sales data sitting in that system. They have seasonal patterns you could set a clock to. And yet the reorder points are still someone’s best guess, updated by hand.

This is the gap I keep seeing. The data is there. The tools are there. The connection between the two is missing.

Machine learning for demand forecasting isn’t new. What’s changed in the last year is that the barrier to actually using it has dropped enough that mid-market teams can get started without a data science hire. The platforms have gotten better. The integrations with ERPs and inventory systems have gotten simpler. And the results, when done right, are worth paying attention to.

What the numbers actually look like

The U.S. National Retail Federation reported that online returns hit 19.3% of sales in 2025, totaling $849.9 billion across retail. A big chunk of those returns trace back to stocking the wrong things in the wrong quantities. You buy too much of SKU A, not enough of SKU B, and you’re stuck with markdowns on one side and stockouts on the other.

A Forbes piece earlier this month highlighted research showing that ERP transaction data, the stuff already sitting in your system, can predict planning run times and infrastructure needs before execution begins. The point isn’t that AI is doing something magical. It’s that the data you already have is more useful than most teams realize.

McKinsey has previously estimated that AI-driven demand forecasting can reduce forecasting errors by 30-50% and cut lost sales from out-of-stock items by up to 65%. Those are big numbers. But I’d take them with a grain of salt until you see them in your own catalog, with your own lead times and your own customer behavior. Published ROI claims in this space are often vendor-led or borrowed from contexts that don’t match B2B distribution.

Where ML actually helps (and where it doesn’t)

Machine learning shines at the stuff humans are bad at: processing thousands of SKUs simultaneously, detecting subtle seasonal shifts, factoring in lead time variability across dozens of suppliers, and adjusting reorder points continuously instead of once a quarter.

What ML doesn’t do well is understand context. It doesn’t know that your biggest customer just renegotiated their contract and will be ordering differently next quarter. It doesn’t know that a competitor just entered your market. It doesn’t know that your sales team just closed a deal that will shift demand for a whole product category.

This is why the best implementations I’ve seen pair ML predictions with human review. The system runs the numbers and surfaces recommendations. A planner with real customer knowledge reviews the exceptions and overrides when needed. The system learns from those overrides.

Elogic published a roundup this month on AI in B2B ecommerce that made a point I agree with: the biggest gap right now isn’t awareness or experimentation. It’s the move from isolated pilots to measurable business impact. Too many teams run a demand forecasting pilot, get decent results, and then never scale it because the process change feels harder than the technology.

A practical starting point

If you’re thinking about this, here’s what I’d suggest:

Start with your top 20% of SKUs by revenue. Those are the items where forecasting errors cost you the most. Get the data clean first. Garbage in, garbage out still applies, and it applies harder with ML than with a spreadsheet because the system will confidently optimize on bad data. Then run a parallel test. Let the ML system make predictions alongside your current process for 6-8 weeks. Compare the results. See where the machine beats your team and where your team beats the machine.

The goal isn’t to replace the Monday morning reorder meeting. The goal is to make it 30 minutes instead of four hours, and to have the decisions be better because the baseline numbers are more accurate.

What I’m watching

CNBC ran a piece this week on AI startups tackling retail returns. Simeon Siegel at Guggenheim made a comment that stuck with me: returns are a “silent killer” of margins, and the best AI applications are the ones that address the root cause, which is getting the right product to the right person in the first place.

That’s demand forecasting at its core. Not a flashy AI demo. Just a better answer to the question: how many of these should we stock?

The teams winning with this aren’t the ones with the biggest AI budgets. They’re the ones who started with a specific problem, measured the results, and let the data change how they work.

What’s your current process for adjusting reorder points? Is it still manual, or have you started bringing in automated demand signals? I’d like to hear what’s working and what isn’t.