The title changed to Chief Data and AI Officer. The instrumentation didn't.
Your mandate grew to cover probabilistic systems. Your measurement stack still assumes deterministic ones.
The role is quietly being rewritten. Chief Data Officer became Chief Data and AI Officer across a lot of org charts in the last two years, and the survey data now shows AI governance appearing as a critical priority in its own right for the first time.
The weak signal: the mandate expanded faster than the instrumentation under it.
The forces
Data quality was always measurable — completeness, freshness, conformance, lineage. Hard work, well-understood shape. Every mature data function has a dashboard for it.
Model behaviour is not measurable by any of those. A model can consume perfectly governed data and still fabricate a citation, cave under pushback, or state a fact that expired last quarter.
Old way, new way
Becoming obsolete: treating AI assurance as an extension of data quality. Clean inputs are necessary and nowhere near sufficient.
Becoming valuable: a behaviour measurement capability that sits alongside the data one — named modes, rates with intervals, controlled comparison, provenance to the raw output.
The timeline
Immediate: AI governance gets bolted onto the data governance function because that is where governance lives, and it inherits tooling designed for data at rest.
Near-term: the first board question that the data-quality dashboard cannot answer. Usually "how often" or "is it getting worse".
Dominant: behaviour measurement becomes a named capability with an owner, the way data quality did a decade ago. The CDAIOs who build it early will have a year of baseline when everyone else is starting.
How to prepare
One named failure mode, one rate, one interval, measured monthly on the system that matters most.
Why: The value is not the first number. It is having a second one three months later, because a trend is the thing a board actually wants and you cannot backfill it.
What would prove this wrong: if model providers ship behaviour guarantees strong enough that measuring it yourself becomes redundant. I would not plan for that, but I would watch for it.
Does your function own a number that describes how your AI behaves — or only numbers describing what it was fed?
Every figure here describes something measured and committed. See the measurements · read the method