The most expensive decision in an agentic rollout is not the vendor. It is the sequence. Which work goes to the system first. Which goes later. Which never goes at all. Most teams get this wrong in the same direction: they hand over the interesting work and keep the repetitive work, because the interesting work is what the demo showed.
There is a simpler rule. Agentify the work you can already grade.
Cost is the wrong first filter. Teams rank tasks by hours spent and hand over the biggest number. But hours saved mean nothing if every output needs a long, uncertain review. The real constraint is evaluation. Work you can grade quickly scales. Work you cannot grade quickly just moves the queue from production to review.
Start where the grading is cheap.
If a reviewer can look at an output and know in thirty seconds whether it passes, that work is ready for an agent. Formatting a report against a template. First-draft briefs written to a documented structure. Tagging and routing inbound requests. Research summaries pulled from named sources. QA passes run against a checklist. In each case the pass criteria already exist, so the human approval gate does what it is built to do. Review stays fast. Rejections are specific. The system learns the standard instead of guessing at it.
This is how our own operation runs. Thirteen agents have produced 44,000+ runs in 63 days, and the gate has held because most of those runs are gradeable at a glance. An output either matches the written standard or it comes back with a reason attached. That loop is the whole mechanism.
Never agentify what you have not written down.
The work to keep human is not creative work as a category. It is unwritten judgment. If the standard for a decision lives in one person's head, an agent cannot inherit it and a reviewer cannot grade it. You will not get a bad output. You will get a plausible one, which is worse, because plausible outputs pass tired reviewers.
So the rule has a second half. Before you agentify judgment, write it down. Voice rules. Positioning boundaries. The client situations that always escalate. Some of it should stay human even after it is written: pricing calls, apologies, a change in positioning. But writing it down does two useful things. It tells you which work is actually gradeable, and it usually reveals that more of it qualifies than you thought.
The order compounds.
Sequence it right and each handover makes the next one cheaper. Review stays fast, the gate never becomes the bottleneck, and that is how teams reach 3 to 5 times their prior throughput inside 90 days without adding reviewers. Sequence it wrong and the gate clogs with outputs nobody can confidently judge. That is the shape of most stalled pilots. Not a model problem. A grading problem.
Our own agent roster was built in this order. The gradeable work went first. The judgment work was written down before anything touched it. And the posture stayed conservative throughout: read-only connectors, a human gate on everything that ships, audit-grade logs behind every run.
If you are deciding what to hand over first, that is exactly the conversation to have out loud. Book the strategy blueprint call.
