A field guide to the seven decisions that separate an AI experiment from a system your organization can operate.
An AI pilot that reaches the production conversation has usually already shown that the model can produce a useful answer. What stalls the transition is a different category of problem.
A pilot is a demonstration; a production system is an operational commitment. That commitment requires decisions about where the workload runs, how data is handled, who is accountable at 2 a.m., what happens when the model needs to change, and what the whole thing costs at real volume. Those decisions tend to arrive together, often in a security review, and often after someone has told an executive that the hard part is finished.
This guide organizes them into seven questions, in roughly the order production forces them.
1. What is actually moving into production?
One risk at this stage is scope that was never defined. "We're putting AI into customer support" is a direction rather than a workload. A workload has a task, an input, an acceptable output, a volume, a latency expectation and an owner.
The definition matters because every subsequent decision depends on it. Model selection, hardware sizing, placement and the commercial model are all downstream of the task profile, and a workload defined only as "AI" will generate an infrastructure design that fits nothing in particular.
The test: can you write one sentence describing what must run, for whom, how often, and what a good result looks like? If not, that's the first piece of work, and it isn't an infrastructure problem.
Go deeper
2. What shape is the workload?
Training and inference are genuinely different, but the usual shorthand (training is bursty and latency-tolerant, inference is continuous and latency-sensitive) describes some deployments and misleads about plenty of others.
Training is often batch-oriented and distributed across many devices with synchronized communication between them. Whether it arrives in bursts, though, depends on your organization rather than on the technique: a research group with a steady queue of experiments runs training as a continuous load. Fine-tuning spans an especially wide range, and a low-rank adaptation on a single device has little operationally in common with a full fine-tune spread across a cluster.
