The metric trap: containment is the new AHT

Earlier this week, I tried to do something pretty straightforward—cancel a subscription to avoid an automatic renewal. I had six months left, fully paid, so I assumed it would be a quick chat with a customer support bot, confirm the cancellation, and move on with my day. Instead, I found myself stuck in a virtual Groundhog Day.

The chatbot immediately recognized something wasn’t right—its own system showed my subscription ending next month instead of six months later. It even acknowledged “this doesn’t look right” in its response. But here’s the problem: rather than escalating me to a human who could fix it, the bot cycled me through the same “how cancellation should work and that it wasn’t working right” explanation over and over, like it was forbidden to hand off the conversation.

The bot never admitted defeat. It never volunteered an escape route. I had to explicitly demand a transfer to secure the help I needed.

My immediate question was not about the technology’s capability, but about its incentive structure. Why would the bot fight so hard to keep a problem it knew it couldn't solve? The likely answer is that the organization driving the bot was focused on the wrong metric.

The seduction of containment

In the world of customer service and CX automation, containment rate is the single most common metric. It is easy to measure, easy to report, and directly correlates with reduced operational cost. But it’s all wrong.

Containment measures whether the customer was prevented from reaching a more expensive channel, typically a human agent.

Vendors and internal CX teams love this metric because its absence is easily quantifiable:

• Did the bot need to transfer the customer? (Failure)
• Did the customer abandon the conversation? (Failure)

It is much simpler to measure whether the desired outcome did not occur (no transfer, no abandonment) than to measure whether the desired outcome did occur (true resolution, confirmed satisfaction). This ease of measurement creates a strategic blind spot.

When success is defined by what doesn't happen (no transfer), the system, whether human or AI, is incentivized to prioritize that metric, even if it leads to customer frustration.

The historical parallel: The ghost of AHT

We have seen this exact dynamic destroy customer experience before, and the core lesson applies directly to modern AI.

For decades, contact centers were ruled by average handle time (AHT). The goal was to keep calls brief and efficient. Managers were bonused on achieving a low AHT.

The result was predictable: agents rushed customers, disconnected complex calls, and avoided necessary research or escalation. AHT was an easily measurable metric that achieved a goal (speed), but it actively drove the wrong behavior and sacrificed quality. More importantly, increased repeat conversations and customer dissatisfaction, which can lead to attrition and detraction, quickly outweighed any cost savings from reduced AHT.

Today, containment rate is the new AHT. It focuses AI on the wrong behavior, prioritizing the internal cost metric over the customer's need for a quick resolution. The bot I encountered was behaving like a human agent desperate to avoid a long AHT. It was trying to secure the "containment" win, even if it meant looping through policy explanations that did nothing to resolve the core issue.

Why escalation is not failure

In many CX strategies, escalation to a human is seen as a breakdown in automation. But that view is outdated. Escalation is not failure – it’s adaptation.

By treating escalation as “costly failure” instead of “value-driven collaboration,” some pricing models train bots to resist passing the baton. The result? Frustrated customers, unresolved issues, and ironically, more resource cost in the long run.

The Cobra Effect: A warning to CX leaders

The ultimate danger of focusing on an easy, but imperfect, metric is known as the Cobra Effect.

The classic story goes that during British rule in India, a bounty was offered for every dead cobra to reduce the snake population. Locals quickly realized they could simply breed cobras to kill them and collect the bounty. When the government ended the program, the newly worthless snakes were released, resulting in more cobras than when the program began.

The lesson for customer experience is chilling: Much like AHT, when you measure and reward containment, you may be generating "contained" conversations that deliver zero customer value.

Your automated interaction may create friction, loop endlessly, or offer partial, vague answers that encourage the customer to abandon the conversation out of sheer exhaustion. These abandoned interactions are often recorded as a "failure" by the vendor (no outcome paid), but the customer leaves angry, leading to future churn and increased upstream costs.

Outcome-based pricing and misaligned metrics vs. usage-based pricing and CX prioritization

This metric problem is often amplified when organizations utilize outcome-based pricing for their AI agent contracts.

Outcome-based pricing is inherently fair—you pay for success. However, if the contract defines "success" simply as the absence of a transfer or abandonment, the vendor (and the AI they build) is financially compelled to prioritize that simple measurement over the complex reality of resolution.

The pricing model itself becomes a constraint on strategic CX design. It rewards the aggressive gatekeeper rather than the efficient problem solver.

Instead, when you pay for usage, the vendor has no financial fear of the transfer. This allows you, the buyer and the steward of the customer relationship, to dictate the rules of engagement.

You can program the bot to be efficient and humble: "If the system error is a billing anomaly involving prepaid dates, immediately apologize and route to a tier-2 agent."

In this model, if the bot spends two minutes looping before a transfer, you, the buyer, pay for those two wasted minutes, plus the agent time. Your incentive is crystal clear: make the bot fail fast and transfer quicker. The only way to save money is to make the bot smarter or faster at escalating.

The path to true resolution

To escape the metric trap, CX leaders must shift their focus to resolution metrics that are difficult for the AI to game.

1. Measure resolution, not deflection: Implement contextual post-interaction surveys such as those provided by CallMiner Outreach or use conversation intelligence to verify the customer received the expected outcome. Did they achieve their goal? You can use that same conversational insight to identify opportunities to improve the bot.
2. Focus on next issue avoidance: Measure the rate at which customers return with the same issue within 24 or 48 hours. A contained but unsolved conversation is always a costly failure.
3. Incentivize smart escalation: Design your AI and your metrics to reward the bot for failing fast. The bot should be programmed to immediately transfer when a conversation deviates from its scope or when negative sentiment is detected. The cost savings come from reducing the time spent in the loop, not avoiding the transfer itself.

The goal of automation is not to prevent customers from talking to you; it is to solve their problems efficiently. By prioritizing metrics that measure true value—resolution and satisfaction—over the tempting simplicity of containment, we can ensure our AI is built to serve the customer, not just the spreadsheet.