Most AI business cases cover two things: deployment cost and projected ROI.

The third number i.e. what costs to keep the system running in year two and three, rarely makes it into the deck.

👋🏻 I'm Leonardo Ubbiali. This week we're looking at why the manufacturers who deployed AI two or three years ago are now discovering that the deployment cost was only the first invoice.

A supplier I spoke to last week is three years into a predictive maintenance deployment across 80 machines. 

The AI still works but the annual maintenance bill is now larger than the original deployment cost.

This is not unusual.

60% of five-year AI costs come from maintenance, retraining, and scaling.

For cloud-deployed AI, inference serving alone consumes 70 to 90% of total compute costs over a model's lifetime.

A quality inspection model trained on one product configuration needs retraining when materials change, a new supplier comes in, or when tolerances shift.

In a factory running 50 SKUs, that can mean dozens of retraining cycles a year. 

Each one requires compute, data validation, testing, and sign-off before the updated model touches production again.

Nearly half of industrial AI adopters cite integration with legacy OT systems as their top adoption barrier, and every model update requires custom integration work on top of that.

Only 15% of companies forecast AI costs accurately. 56% miss by 11 to 25%, and nearly one in four miss by more than 50%

As Ana Bildea, Tech Lead AI at AlterSquare puts it: "Traditional technical debt accumulates linearly. AI technical debt compounds."

Ecolab, the chemical manufacturing company, ran into this problem at scale. 

Model deployment across their operations was taking 12 months per cycle and the maintenance overhead was consuming more engineering time than the original deployment.

They rebuilt their MLOps infrastructure from scratch and cut deployment cycles from 12 months to 30 days.

The savings came not from the AI itself but from treating model maintenance as a system, not an afterthought.

Most manufacturers ask vendors two questions:

The third question that rarely gets asked is: what does it cost to run this system in year two and year three?

Most vendors don't volunteer the answer.

The manufacturers who avoid the situation Ecolab found themselves in ask for retraining frequency estimates upfront, negotiate maintenance terms at the contract stage, and get inference costs as a separate line item before signing.

The team that deploys the system won't maintain it indefinitely, and assuming otherwise is how a seven-figure annual maintenance bill appears 18 months later.

None of that is complicated. It just has to happen before the CFO signs, not after the first maintenance invoice arrives.

The problem: Your AI business case covered deployment and projected savings. You need to model the actual three-year cost before the CFO approves it.

What you need: The vendor quote, number of machines in scope, your legacy infrastructure age, and current internal MLOps capability.

A three-year cost model with assumptions stated explicitly, the costs most commonly absent from vendor proposals, and questions to pressure-test the numbers before you commit.

Survey of 500 engineering professionals on where AI budgets go, why ROI is hard to calculate, and what drives the 36% year-over-year cost increase. The inference serving breakdown is directly usable for anyone building an AI business case.

Time to value: 20 minutes

When you approved your AI budget, did the business case include year two and year three maintenance costs, or just the deployment number?

Hit reply. I read every email.
Leo