Three years ago, the honest answer to “can AI do this work?” was usually “partly.” That answer has changed, and most enterprise AI programs have not noticed.
Today’s frontier models handle the large majority of rules-based knowledge work at accuracy levels that survive audit. Document extraction, classification, reconciliation, first-draft production, triage: the intelligence layer of these tasks is solved to a degree that would have sounded like vendor talk in 2023. If your program is still running capability evaluations on this class of work, you are re-litigating a settled question.
The constraint moved. It now sits in three places, and none of them is the model.
Workflow selection. Most programs pick their first AI project by excitement rather than by readiness. The showcase project, the one the board asks about, is usually judgement-heavy, poorly defined, and impossible to baseline. It fails quietly and takes the program’s credibility with it. The right first candidate is boring: well-defined, high-volume, verifiable, and ideally already outsourced or visibly backlogged, because then the baseline is real and nobody disputes the before-and-after.
Integration. The distance between a working demo and a production system is identity, data boundaries, legacy APIs, and audit. This is unglamorous work, which is why it gets scheduled last and why pilots stall for eighteen months. Enterprises that treat integration as the main work ship. The rest accumulate proofs of concept.
The judgement line. Every workflow has a point where rules end and experience begins: which exception matters, when to escalate, which trade-off to accept. Drawing that line correctly for a specific workflow in a specific organization is the actual skill of enterprise AI deployment. Draw it too conservatively and you automate nothing meaningful. Draw it too aggressively and you discover the error in front of a customer or a regulator.
These are judgement questions, not model questions. They are answered by sitting with the work, not by reading benchmarks. That is worth remembering the next time a vendor’s first slide is about their model.