Are Your Incentives Quietly Training Teams to Game the AI Agent's Metrics?

Quick answer

Any single AI agent metric attached directly to a bonus, promotion criterion, or hard target will eventually get optimized in ways that technically hit the number while quietly undermining what it was meant to represent. The most reliable defense is structural, not moral: pair every optimizable metric with at least one counter-metric that would catch the most obvious way to game it, rotate or randomly audit a sample of the underlying interactions rather than trusting the aggregate number alone, and avoid tying compensation to a single narrow metric in isolation. Teams are rarely acting in bad faith when this happens; they are responding rationally to whatever the incentive structure actually rewards, which is exactly why the fix has to be in the incentive design, not in trusting people to resist the pressure.

This is a design problem, not a trust problem

The instinct when a metric looks suspiciously good is to question whether someone is cheating. That framing usually misses the real issue. If containment rate is tied to a team's bonus, and containment rate can technically be improved by making the agent more reluctant to escalate difficult cases, most teams will drift toward that behavior over time even without anyone consciously deciding to game the system. Each individual choice, tune the escalation threshold slightly higher, close out an ambiguous conversation as resolved rather than escalated, feels reasonable in isolation. The aggregate effect is a metric that looks better while the underlying experience, the thing the metric was supposed to represent, gets worse.

This is the standard shape of Goodhart's law: a measure that becomes a target stops being a reliable measure. It is worth designing around explicitly, precisely because reporting the metric at all, the practice covered in what AI agent metrics belong in front of the board or CEO, creates the visibility and pressure that makes gaming more likely, not less, once compensation is attached to it.

Pair every optimizable metric with a counter-metric

The single most effective structural defense is to never let a metric stand alone as a target. For every metric you attach incentives to, identify the most obvious way a team could improve the number without improving the underlying reality, and track that specifically as a paired counter-metric:

• Containment rate paired with post-resolution reopen rate or customer-reported satisfaction on "contained" conversations. If containment climbs while reopen rate or satisfaction on those same conversations also climbs (in the wrong direction), that is the signature of gaming, not genuine improvement. This connects directly to the measurement discipline in measuring and improving AI agent containment rate, which already warns against optimizing containment in isolation.
• Response speed paired with a quality or accuracy sample. A team incentivized purely on speed will find ways to answer faster, and some of those ways involve giving thinner, less careful answers.
• Cost per resolution paired with escalation and complaint volume. Cutting cost by routing more cases to a cheaper model tier, the tradeoff covered in deciding which conversations get routed to a cheaper model vs a more capable one, looks great on a cost dashboard while quietly degrading outcomes if quality is not tracked alongside it.

Sample and audit, do not trust the aggregate alone

Aggregate metrics can be gamed at the margin in ways that are invisible without looking at actual underlying conversations. Build a standing practice of pulling a random, statistically meaningful sample of interactions each period and having a human review them against the metric's intended meaning, not just its reported number. This is the same discipline covered in ongoing human QA transcript sampling, applied specifically to metrics that carry incentive weight rather than to quality monitoring in general. A metric that only gets checked through its own dashboard, with no independent sampling, is the easiest one to quietly game undetected.

Design the incentive, not just the metric

A few structural incentive choices reduce gaming pressure before it starts:

1. Avoid single-metric bonuses. Tie compensation or recognition to a small basket of paired metrics rather than one number, so that improving the basket requires genuine improvement rather than optimizing one dimension at the expense of another.
2. Reward trend and consistency over a single period's snapshot. A metric that is easy to spike briefly (right before a review period) but hard to sustain is a strong gaming signal; rewarding sustained performance over several periods reduces the payoff from short-term manipulation.
3. Make the counter-metrics visible to the same audience as the primary metric. If leadership only ever sees containment rate and never sees the paired reopen rate, the incentive to report the flattering number without the context that would explain it remains strong.

FAQ

Is metrics gaming usually intentional? Rarely in a deliberate, dishonest sense. Most gaming emerges from many small, individually reasonable decisions responding rationally to whatever the incentive structure rewards, which is why fixing the incentive structure works better than appealing to individual judgment.

How many counter-metrics does each primary metric need? Usually one is enough if it is well chosen, specifically targeting the most obvious way the primary metric could be gamed. More than two or three paired metrics tends to dilute focus rather than add protection.

Should we stop reporting a metric if it turns out to be gameable? Not necessarily. Most useful metrics are gameable to some degree; the fix is pairing them with a counter-metric and independent sampling, not abandoning metrics that are otherwise informative.

Does this only apply to metrics tied to bonuses? It applies most strongly there, but any metric a team knows is being watched by leadership, even without direct compensation attached, can create similar, if weaker, gaming pressure.