when a team tells me they want to use ai for a task, the first thing i ask is: what would happen if you just automated it? not with a model. with a script, a workflow, a trigger.
half the time, the answer is: that would actually solve it. the rest of the time, the answer reveals something even more useful — that the process is not stable enough for automation either, let alone intelligence.
i have audited projects where teams spent six months fine-tuning a model to classify customer intent when a two-week automation project would have routed 85% of tickets correctly with a decision tree.
- 01 repeatability does this task follow a stable pattern? if not, neither automation nor ai will stick.
- 02 volume is there enough volume to justify the complexity? ai overhead is real.
- 03 tolerance what is the cost of being wrong? deterministic systems fail predictably. models fail creatively.
none of this is anti-ai. it is pro-results. the sequence matters: order, automate, augment. most teams start at augment and wonder why nothing holds.
automation before intelligence is not a compromise. it is the discipline that makes intelligence worth it later.
start with automation. if it is not enough, then — and only then — reach for the model.