Build vs. Buy in the Agentic Era

Overview

The first agent is easy. An AML team builds something to assemble alert context, the analyst still signs the disposition, and a thirty-minute review comes back in a few minutes. Then the fraud team asks for one too — and that request exposes a question the first agent never forced: which pieces does every new agent rebuild from scratch, and which should the institution build once and share?

This field guide is about that second question. It traces how risk decisioning got here, through rules engines, point solutions, orchestration and ML, and now agents that assemble the case itself. It lays out the five ownership decisions that determine whether building, buying standalone agents, or running on a shared platform is the right call — and what each path actually commits you to after launch.

What you'll learn

• Five decisions that come before any vendor conversation. Will the next agent start from zero? Can a finding move without a person carrying it? Can you defend the decision later? Who can change it when the rules change? What are you agreeing to own? Each comes with a side-by-side build/buy framing and a concrete test to run in a demo — the blank-case test, the reconciliation test, the reconstruction test, the policy-clock test.

• Three paths compared, including when building wins. Build in house, standalone agents, and a shared platform layer, with what you own, when each is the right fit, and the failure mode of each. Includes the five conditions that make an internal build the correct call, the hybrid path strong internal teams choose, and why bounded scope is the condition teams most often overestimate.

• What an agent really costs, and how to price it. Separating the shared operating layer from the marginal use-case work that never disappears, a line-by-line loaded estimate you can build yourself, a simple model for your finance partner, and the cost that rarely makes the spreadsheet: deployments scaled back because governance, evaluation, and maintenance were never built to production standard.