When You Finally See Which Products Actually Make Money

Ask a cost accountant which products truly make money and you'll often get a pause. Not because they don't know their business — because traditional costing smears overhead across products with broad-brush allocations that hide the truth. The high-volume product looks less profitable than it is; the fiddly, resource-hungry one looks better than it is. Whole product strategies get built on that distortion.

Activity-based costing was the answer, in theory. In practice it collapsed under its own weight — tracing every activity to every product was so laborious that most teams did it once, produced a binder, and never updated it. The methodology was right; the manual effort made it impractical to keep alive.

Where AI makes ABC practical

The reason ABC was painful was the allocation work: figuring out how much of each activity and resource each product consumed. AI is well-suited to exactly that. Models can learn cost-driver relationships from operational data — machine hours, setup counts, labor time, energy, handling — and allocate costs to products and activities far more granularly than a manual scheme, and keep it current as the data flows rather than freezing it in a one-time study.

What you get is margin visibility at the level decisions actually get made — by product, by SKU, by customer, by channel. And it's frequently uncomfortable. The SKU everyone loves turns out to lose money once you account for the setups and handling it demands; the boring product quietly carries the P&L. That discomfort is the value. You can't fix a margin problem you can't see, and broad-brush costing keeps it invisible.

From there the actions get concrete: reprice or rationalize the true loss-makers, protect and grow the real earners, renegotiate with customers whose service demands you'd been subsidizing without knowing it. These are decisions with real money attached, and they were being made on distorted numbers.

The cautions

Cost allocation involves choices, and the model reflects the drivers you give it. If you pick the wrong cost drivers, you'll get precise-looking numbers built on a flawed premise — precision isn't the same as accuracy. So the driver selection deserves genuine scrutiny from people who understand the operation, not just the data science.

Source data quality is the ceiling again. This needs decent operational data — production records, machine logs, labor tracking — and if that's thin or unreliable, the costing rests on sand. Sometimes the first real project is instrumenting the shop floor, not building the model.

And treat the output as decision support, not gospel. The allocations are more accurate, not perfect, and the point is better decisions, not a false sense of five-decimal certainty. Keep humans reading the results with judgment.

One closing thought: the value here isn't really the costing engine. It's the conversations it forces — about pricing, about which products earn their place, about which customers are worth the trouble. Better cost data just means those conversations finally start from the truth.