Containment rate
Metricself-service rate
The share of cases an AI agent closes end to end without a human stepping in. Implementa uses this formulation — adapted from contact-centre practice — as the headline metric for whether an agent is actually doing work or just standing in front of the queue.
Containment is the complement of escalation rate: if 70% of cases are contained, 30% escalate to a human. It's a deceptive metric on its own, which is why Implementa always publishes it paired with a quality metric (correct resolution, reopen rate or satisfaction). An agent can contain 95% of cases simply by closing them badly: nobody stepped in, but nobody solved anything either. Read properly, containment answers the economic question in AI Operations — how much human work stops happening — while the quality metric answers the risk question. Pushing containment up at the expense of quality is the fastest way to make a deployment look like a win for one quarter and become a problem the next.
How it differs from
- Resolution rate
- Containment measures that nobody had to step in. Resolution measures that the problem was actually solved. They can diverge widely, and the gap between them is where quality debt lives.
- AI adoption rate
- Adoption measures how many people use the tool. Containment measures how much work the tool finishes unaided. You can have high adoption and laughable containment.
Examples
- A support agent closing 7 of every 10 tier-1 tickets without escalating: 70% containment
- Measuring containment by case type and discovering the escalating 30% concentrates in two specific categories
FAQ
- What counts as a good containment rate?
- There is no universal number, and be suspicious of anyone offering one: it depends on case type, how much risk the process tolerates and how much autonomy you gave the agent. What's useful is your own series — containment at launch, containment today, and what happened to quality along the way.
- Can you optimise containment without breaking things?
- Yes, if you optimise per category rather than in aggregate. Look at where the agent escalates, decide case by case whether that escalation is a context failure, a permissions failure or a correct decision that should stay human, and only attack the first group.