Last week, a16z published a sharp piece called "Workday's Last Workday?" The argument is simple and uncomfortable.
Workday reported $400 million in AI ARR. Sounds impressive. But when you look at how that number was built, things get strange. Customers are purchasing something called "Flex Credits." Workday records it as AI revenue. The question — which agent does what, for whom, in which workflow — goes unanswered. Both sides hit their KPIs. Nothing moves to production.
This isn't just a Workday problem.
Three weeks ago I was reviewing activation data for one of our SaaS products. A customer had enabled our AI-powered reminder feature. It was part of their paid plan. They'd turned it on six weeks earlier. They hadn't used it once.
When I followed up, the response was almost word for word: "Someone told me to turn it on, so I did. I'm not really sure what it does."
Sold. Activated. Unused.
Is that adoption? No. But where does it show up in our metrics?
That's the gap between AI ARR and AI adoption.
ARR is a finance number. It says "this customer has a plan that includes AI features" or "this customer activated the AI bundle." Perfectly valid for accounting. Not useful for product decisions.
Adoption is different. It means: a real user completed a real workflow using this feature, more than once, and got something out of it. And most PMs don't measure how far apart those two numbers actually are.
What fraction of Workday's $400M maps to genuine, repeated, value-generating usage? Unknown. Workday might not know either.
Why is this a PM problem?
Because when the numbers look good, we stop asking.
AI feature sold — ✓ AI feature actively generating value — unknown.
Feature acquisition and feature engagement aren't the same thing. We tend to measure the first and skip the second. In AI products, this gap is particularly costly because two things break simultaneously when a feature isn't used.
First, users don't see value. No value means churn risk. But the ARR dashboard looks healthy so we move on.
Second, the model stops improving. AI systems learn from real usage signals. A dormant AI feature is a frozen AI feature. Expecting it to get better on its own is wishful thinking.
"AI washing" is a real pattern. Companies announce AI roadmaps, bundle AI features into plans, report AI ARR. Which customers are actually doing what, how often? Unclear.
The right PM question isn't "How many customers purchased our AI feature?"
It's "How many customers used this feature at least three times in the last 30 days to complete a real work task?"
That second question feels uncomfortable. Because the answer is almost always lower than you expect.
But not asking it is worse. Because then you can't see your users, your product, or your own roadmap clearly.
Workday isn't alone. Salesforce, SAP, most major B2B SaaS players are sitting in some version of this paradox. They're reporting AI revenue. Customers are buying. The actual usage data tells a different story.
Who can see both sides at once?
The PM.
Because the PM has access to both the ARR dashboard and the usage data. Putting them side by side makes the gap visible.
Making the gap visible is part of the job.