This is the part that changes the numbers. Applications built onto the processes carrying your volume, one at a time, each with the gates where a person still decides and a record of everything the agents did. Every one that goes live makes the next one cheaper.
A chat window beside someone's work is a tool they have to remember to use. An agentic application takes a process and runs it: it is triggered by the same events the old process was, it reaches the systems it needs, it stops at the decisions that carry cost, and it leaves a record. The difference shows up in whether anything changed six months later.
Precisely enough to survive literal execution. This is the part most projects skip and the reason most of them stall.
The repetitive work, at the rate it arrives, on the systems it already lives in. Your data stays where it is.
Thresholds, exceptions and anything with a customer or a regulator on the other side. Designed in, not bolted on.
What was asked, what came out, which model, who approved. The account you need when someone asks why.
We build in the domains where a wrong answer costs money or standing, because that is where the gates and the record are worth what they cost.
Specification to build to review, with gated merges and a record of what shipped and why.
People and agents on one board: chained dependencies, gates where consequence lives, and an operating review written from the record.
Test generation and execution at volume, evaluations at the gates, and evidence kept as work happens.
The high-volume back office: exceptions, triage, reconciliation, and the reporting that hangs off it.
The first application pays for the ground. Everything after it is built on ground that already exists, which is why the second is faster than the first and the fifth is faster again.
Permissioned routes to your systems, built once. Most of the cost of a first agent is getting to the data, and you only pay it once.
After the first build your team can write a process down precisely enough to hand to an agent. That skill is the actual asset.
Every application adds to one record of how your work goes wrong and what you did about it. Nobody else can rebuild that.
When execution costs almost nothing, the companies that pull ahead are the ones whose people can say precisely what they want and stand behind what comes out. Our working paper calls that a principal company: people decide what good looks like, agents do the volume, orchestrators govern them, and a record proves both.
Most companies will get there eventually, because the economics push everyone the same way. The advantage belongs to whoever arrives first, with the ground laid and the practice already in their own team.
Start with one application and a number. Build the next one on ground you already own.