What containment counts
Containment is the share of conversations that ended without a human being becoming involved. Definitions vary in the details, particularly around what counts as a conversation and what counts as involvement, and those details move the number more than most people realise.
The intent behind it is reasonable. If an automated layer is doing useful work, fewer conversations should require a person, and the share that do not is a rough proxy for that. As a direction of travel within one business over time, it carries some information.
The problem is that the metric measures the absence of an event rather than the presence of a good outcome. Absence of escalation is compatible with success and it is equally compatible with the visitor giving up. Nothing in the number distinguishes the two, and the two feel identical from the reporting side.
The trivial way to make it perfect
Remove the route to a person. Containment goes to a hundred percent. Every conversation ends without human involvement, because human involvement was not available.
Nobody does this deliberately in one step, which is why it is worth naming. It happens by degrees, through changes that each look like an improvement in isolation. Move the handover offer to the third reply instead of the first. Add a qualifying question before the form. Make the assistant try once more before offering a person. Reword the handover so it sounds like a last resort. Each of these raises containment and each of them makes the product worse.
This is the specific danger of a metric that can be improved by degrading the thing it is supposed to measure. Nobody has to be dishonest for it to happen. A team optimising in good faith against a number will find these changes, and the number will reward them for every one.
The three ways it goes quietly wrong
Abandonment reads as containment. A visitor who asks a question, gets a poor answer, and closes the tab has been contained. The metric records a success. In many implementations this is the single largest contributor to a high containment score, and it is invisible unless you go looking.
Displacement reads as containment. The visitor gives up on the widget and phones instead, or emails, or turns up in person. The conversation was contained and the contact happened anyway, on a more expensive channel, without a link back to the widget conversation that caused it. Businesses that measure only their chat containment can report a rising number while their phone volume rises with it.
Confidently wrong answers read as containment. A visitor who receives an incorrect answer, believes it, and leaves satisfied has produced a perfect containment event and a future problem. This is the failure mode that makes containment actively dangerous rather than merely uninformative, because the metric rewards it identically to a correct answer.
The pairings that make it honest
Never report containment alone. Report it as a pair with at least one number that moves in the opposite direction when the assistant gets worse, so that gaming one visibly damages the other.
Pair one, and the most important: handover availability. Something that confirms the route to a person is genuinely offered, such as the share of conversations where a handover was presented in the first reply. If containment rises while that falls, containment rose because you hid the exit.
Pair two: what happened in the other channels. Total contacts across every route, not chat volume. If chat containment rises and total contacts are flat, work has genuinely moved. If chat containment rises and phone volume rises too, you have moved contacts rather than removed them.
Pair three: answer correctness on a sample. Any regular check on whether the contained answers were actually right, however small the sample, breaks the equivalence between confidently wrong and correct. Without this, containment cannot distinguish the two, and the two are the whole question.
When a vendor quotes it at you
A containment figure in a pitch tells you almost nothing without three follow up questions, and the answers to those questions are more informative than the number.
First, what counts as a conversation. Whether a visitor who opens the widget and types nothing is in the denominator changes the figure substantially, and there is no standard answer. Second, what counts as escalation: only a completed handover form, or any offer of one, or any conversation where the visitor asked for a person and did not get one. Third, at what point in the conversation is a person first offered, because that single design choice sets the ceiling on the number.
A vendor who can answer all three specifically is describing a real measurement. A vendor who cannot is quoting a number produced by a definition chosen to make the number large. The question is not whether they are being dishonest, it is whether the figure means anything, and the answers tell you.
Reporting it internally without embarrassing yourself
If containment has to appear in a report because somebody senior asks for it, put it next to its pairs and give it a caption. One sentence saying what it does not measure is enough to stop it being read as a quality score, and captions on dashboards are read more often than anybody expects.
Be especially careful about targets. A containment target hands your team an instruction to reduce human availability, whether or not that is what anybody intended, and teams are good at hitting targets. If you must have a target, set it on the pair rather than on containment alone: containment at or above some level while handover availability stays at or above another.
Better still, replace it. The decisions containment is supposed to inform are better served by handovers per topic and unmatched question volume, both of which point at an action. Containment points at a feeling.