Why IR Capacity Has Not Kept Up With Data Demand

In our 50-interview study of IR teams, the pattern was almost identical across institutions: data request volume rises year over year, the IR office stays the same size, and the team is asked to absorb the gap by working faster. There is a ceiling on how fast a person can pull a file, clean a cohort definition, and ship a sourced answer. We have hit it.

The shape of the squeeze is consistent: ad-hoc data requests consume 40 to 60% of team capacity on average, and the volume has grown roughly 40% over the past five years. Team sizes have stayed flat. Half of IR offices run with one to three people. Every new dashboard, new federal reporting requirement, and new dean asking the same question as the last dean adds to the queue without adding to the capacity to answer it.

The math does not work because the question profile is heavy on repeats. The same retention figure, the same FTE breakdown, the same peer comparison arrives from different stakeholders at different times. Each one is processed as a new request even though the underlying data has not changed.

Hiring another analyst is the obvious lever, and it is the one most institutions pull first. It is also the slowest and most expensive. IR analysts with the right domain and technical mix are hard to recruit, take months to onboard, and cost a salary that dwarfs any software the IR office buys. Public institutions face hiring freezes and budget approval cycles that lengthen the timeline further. By the time a new hire is productive, the request queue has grown again.

There is also a structural ceiling: a single analyst does not scale across the dozens of question types an IR team fields. Adding a third analyst does not triple throughput; it adds capacity for a few more repeat questions, then hits the same manual-labor bottleneck.

Conversational AI fits the squeeze because the bottleneck is not insight, it is retrieval and assembly. The analyst who knows the answer still has to pull the data, clean it, calculate it, and format it. A model that understands your data and your definitions can do that retrieval-and-assembly work in seconds once it is connected.

The key is that the model has to be higher-ed native. General AI cannot tell you whether "enrolled student" means headcount or FTE this week, and it cannot pull from your SIS. Clema is built to do both, with sources shown on every figure, so the analyst reviews and ships instead of pulling and assembling.

This is the structural shift: capacity scales with question volume because the marginal cost of the next answer is near zero. That is the only model that mathematically closes the gap between flat headcount and rising demand in IR.