ML outsourcing

Outsourcing depends on providing a precise spec. A common issue is its underspecification: not all expectations are fully spelled out, as they seem obvious to the requester, leading to disappointment. (Incidentally, LLMs suffer from similar issues: if you leave details out, you won’t get what you want.)

Outsourcing often means “offshore” vendors, but with ML, it’s often internal. You tell the ML team what metric you want to optimize and hand over the data, and wait a few months. The underspecification and its consequences follow. The data is not what it was supposed to be. Constraints go unstated. Secondary metrics, that drive launch decisions, aren’t communicated. Deployment costs are not mentioned until later. Compliance checks and privacy limitations show up as afterthoughts. It’s a mess.

The real limitation of ML outsourcing is that the people building the solution, knee-deep in the data and algorithms, are best positioned to spot new opportunities, risks, or pivots that the original requester never imagined. Engineers (incl. ML engineers) are great problem solvers. If you use them as hourly implementers, you’re leaving a sports car parked in the driveway: all that potential, barely explored.

Give your ML folks problems to solve.