Containment is a vanity metric: agent escalations can be the real work.

The opening trap. A 96% containment agent looks like the obvious hire. Druid AI argues the opposite. In their banking example, an agent that resolves a mortgage payoff question and routes the customer to a senior banker is doing more valuable work than an agent that contains everything. That handoff is the future revenue signal. Containment, the default dashboard metric, rewards agents for keeping humans out of conversations even when a human is the right outcome.

The framework. Druid built a 26-component scorecard split between configuration and runtime, with governance across all of it. Each component earns zero if missing, one if measurable, two if managed over time. The weights deliberately favor business impact: governance gets 39%, mandate 26%, platform 12%, output 9%, with skills and tools lower. In the example, the 80% containment agent scores 83 overall while the 96% agent scores worse because it lacks governance, explainability, and linked success criteria.

The benchmark. The intern story is not hypothetical. Druid says it comes from a 2026 AI adoption benchmark built from 15 months of anonymized production data. Containment runs 80% in financial services, 87% in healthcare, 93% in HR and IT, and 99% in higher education. The speakers are clear: higher education is not better. Finance deliberately designs more handoffs because a mortgage question is a refinance event. A single containment number is meaningless without knowing what the agent was designed to do.

The management move. The strongest operational advice is to re-evaluate agents weekly on live traffic against a human baseline. Druid shows a dashboard where a flat 92% containment line stays green while a weighted score slides from 83 to 60, the drift that a vanity metric hides. The ROI question for the CFO should not be which jobs disappear. It should be what the agent took off the plate and what the freed time let the business do.

The weak evidence. This is a vendor pitch. Druid sells conversation AI and the talk ends with a QR code to download the report and a booth invitation. The benchmark methodology, sample size, and confidence intervals are not shown. The 39% governance weight is presented as a given, not a derived result. Teams should take the critique of containment seriously and treat the specific scorecard as a starting point, not a verified standard.

The bottom line. Engineering leaders should stop treating containment as quality. An escalation can be a feature when it routes a customer to a human who can act. Build a scorecard around business outcomes, governance, and drift, update it weekly, and judge agents against a human baseline. The vendor's weights matter less than the habit of looking beyond the green dashboard.