The First 200 Agents

Public research already counts how often enterprise AI agents fail. This paper looks at the problem from inside the build. It draws on 200 agents built between January 2024 and November 2026 for 13 organisations across the EMEA region and Central Asia, in financial services, telecom, shipping, manufacturing and healthcare, and in the author’s own practice. By November 2026, 188 remained in use; none of the 12 that ended failed technically. From these cases come three testable propositions: most of an agent is business logic, supported by a blind two-rater tally (kappa 0.81); without unit ownership an agent falls out of use; and people, not models, decide the economics. The paper turns the third into the Agent P&L, a per-task costing method with six cost lines, an exact-bound rule for earning less human review, and a test that separates cash savings from released staff time. It adds an agent ladder tying autonomy to the cost of an error, nine failure patterns, a two-track procedure for the next agent, and technical, contractual and governance appendices aligned with the EU AI Act and regional regulation.