The pattern is consistent enough to set a watch by. A team doubles headcount over two quarters to keep up with volume. Training runs on schedule, the new cohorts certify, the queue stops backing up. And somewhere in month four, CSAT slides three points and nobody can name a single thing that went wrong.

Nothing did. Scaling support quality is not a training problem, it is a composition problem — and composition changes the moment new people arrive, whether or not anyone trained them well.

What breaks first when scaling support quality?

The mix does. A team of twenty with an average tenure of eighteen months has a deep bench of people who have seen the weird cases. Add twenty more and the average tenure drops to nine months overnight, without a single existing agent getting worse at their job.

That matters because experience is not decorative. The clearest evidence comes from a study by Brynjolfsson, Li and Raymond published in the Quarterly Journal of Economics in 2025, covering 5,172 customer-support agents at a software firm. Introducing an AI assistant raised resolutions per hour by 15% on average — but the gains concentrated almost entirely among less experienced and lower-skilled workers, while the most experienced agents saw small speed gains and small quality declines.

Read that from the other direction and it is a measurement of the experience gap itself. The tool moved novices a long way because novices had a long way to go. Tenure was doing work that nothing else was doing, and adding people dilutes it by arithmetic.

So why do the metrics stay green while it happens?

Because averages hide composition, and support dashboards are built almost entirely on averages.

Handle time, CSAT, first-contact resolution: each is a mean across a population whose makeup just shifted. A stable average during a hiring wave usually means the tenured half improved enough to mask the new half. That is not stability, it is offsetting, and it ends when the tenured half gets tired or leaves.

QA sampling has the same blind spot. Reviewing five calls per agent per month sounds even-handed and is not — it distributes attention by headcount rather than by risk, so a cohort that needs ten times the scrutiny gets exactly the same five.

The ambient conditions make it worse. Verint’s State of Agent Experience 2026, published in April 2026 and based on a survey of 1,000 contact center agents conducted in late 2025, found that 45% of calls require agents to search for answers during the interaction and 57% require them to gather context when an issue escalates. Those are the tasks a new agent is worst at and slowest to complete, and they are on nearly half of all calls.

If it is composition, why does more training not fix it?

Training fixes the part of the gap that is factual. It does not fix the part that is judgment, and judgment is what the ambiguous cases require.

A new agent can pass every product assessment and still not know that this particular complaint, phrased this particular way, is the one that turns into a chargeback if handled literally. That knowledge is not withheld from the curriculum out of negligence. It is not curriculum-shaped. It exists as a stock of encountered cases, and stocks accumulate at a fixed rate no matter how much money is spent.

Which is why the same 31% of agents in the Verint research saying they were likely to leave within six months is not a separate problem from quality — it is the same problem, running in the opposite direction. Scaling adds people at the bottom of the curve while attrition removes them from the top.

What actually slows the dilution?

Three levers govern most of the outcome, and scaling support quality without at least one of them is close to hoping.

• Stagger the intake. Two cohorts of ten, twelve weeks apart, produce a very different tenure distribution than one cohort of twenty, and cost the same. The second cohort also gets trained by a floor that is no longer in crisis.
• Weight QA by tenure, not by headcount. Move the review budget toward the newest cohort and toward ambiguous case types specifically. It costs nothing except a decision about where analyst hours go.
• Protect the tenured bench from being consumed by the ramp. Senior agents pulled into shadowing, escalations and buddy shifts during a growth wave stop producing the results that were holding the average up. Their capacity has to be planned as capacity, not borrowed quietly.

Geography sits underneath all three, because the labor market determines whether the curve resets every year or compounds. Companies that pair a rapid ramp with nearshore call centers in Tijuana, Guadalajara or Monterrey are usually buying tenure stability rather than a lower rate — support is a durable career track in those cities in a way it often is not in a US metro.

A team that holds its people for three years never has the mix problem in the acute form. This is the same reason ramp timelines deserve honest planning rather than optimistic ones: the ramp is not the cost, the dilution during it is.

The uncomfortable version of all this is that some quality loss during fast growth is not preventable. It is a consequence of arithmetic. What is preventable is not seeing it, not planning for it, and discovering it in a quarterly review when the customers who left have already left. Teams working through a structured 90-day scaling sequence at least know which weeks to watch.

FAQ: Scaling Support Quality