What the current generation of agent tooling changes for an operations team, what it leaves exactly as hard as it was, and which figures still need sourcing before publication.
2 min read
What a production AI budget contains that a pilot budget does not, why the second workflow costs less than the first, and which cost figures still need sourcing.
2 min read
Draft — not for publication. The argument below is written from first principles and stands on its own. Every figure is an unfilled placeholder and must be sourced before this goes live.
The tooling around language models has moved from "call a model and paste the output somewhere" to systems that hold state across steps, call tools, and retry when a step fails. That is a plumbing change rather than an intelligence change, and plumbing changes are exactly the ones that move a technology from demo to operations.
For a mid-market operations team the practical consequence is narrow and specific: a workflow that needs six steps, two lookups and a judgement call in the middle is now something a small team can build and maintain, where two years ago it needed either a bespoke integration project or a person doing it by hand.
Three things are exactly as hard as they were.
The data still has to exist. An agent that needs your order history needs an interface to your order history. Most stalled projects stall here, not on the model.
Someone still has to own the exceptions. Agentic tooling makes the happy path cheap and leaves the exception path where it was: with a person, who now sees fewer cases but harder ones.
The measurement problem is untouched. If you could not say what the manual process cost, you still cannot say what you saved. Better tooling does not supply a baseline.
The build side got cheaper for narrow, well-understood workflows and did not get cheaper for anything that touches a system without an API. Buying still wins where the workflow is generic across companies; building wins where the judgement being automated is specific to yours. Cost comparison, once sourced: [PLACEHOLDER — verify before publishing]
Cross-vendor tool protocols, and whether portable agent definitions survive contact with production. Adoption figures deliberately omitted here rather than guessed: [PLACEHOLDER — verify before publishing]
Nothing in the current tooling generation removes the need for a human decision point where money, people or compliance are involved. It moves the cost of building the surrounding workflow down far enough that the decision point is worth building around — which is a real change, and a smaller one than the marketing suggests.