A recent TechRadar Pro piece by Flexera's Chief Product Officer argues that the AI spending spree is over, and that the next phase of enterprise AI will be defined by visibility, governance, and financial accountability over usage. Get real visibility into the stack, build governance, renegotiate contracts, align finance and procurement, and use AI to make cost optimization continuous instead of reactive.
It's a well-reasoned piece, and every one of those five steps is operationally correct. None of them answer the question that actually determines whether AI adoption succeeds.
The measurement is still consumption. Tokens, credits, GPU-hours, contract terms — the entire framework treats AI as a metered utility to be governed more tightly, the same way an enterprise governs cloud spend or SaaS seat sprawl. That's a legitimate problem. Wasted spend is real, and 59% of organizations reporting increased AI waste year over year is a number worth taking seriously.
But cost governance answers "are we overpaying for what we're using." It does not answer "what can this organization now do that it couldn't do before."
AI adoption is a catalyst for capability, not a line item to be metered. The distinction matters because it changes what you optimize for. An organization that optimizes consumption will drive down tokens, seats, and GPU-hours — and can succeed at that while capability stays flat or even declines, because the incentive is to use less, not to achieve more. An organization that optimizes capability asks a different question: has cycle time actually compressed, has decision latency actually dropped, has engineer time actually been redirected to higher-value work, can the business now do something it structurally could not do a year ago.
Those are outcome metrics, not spend metrics. They live in a different measurement stack entirely.
This is the same category error IT operations made with ticket reduction. For years, IT organizations measured automation success by tickets closed or deflected — a consumption-adjacent proxy that rewards handling more volume, not eliminating the underlying failure demand. Two bodies of work I've written about elsewhere only make sense under a capability lens, and both go invisible under a consumption lens. Anomaly-detection-based incident avoidance identifies and resolves root cause before a ticket is ever generated. Operational intelligence built on historical data goes a step earlier: it mines the patterns behind recurring ticket categories and eliminates the underlying workload itself, so the ticket never has a reason to exist in the first place. Neither shows up as a win if you're counting tickets closed or measuring throughput — both only register as wins if you're measuring peak business performance impact. The FinOps lens on AI spend is making the identical mistake one layer up: it will faithfully report that consumption is under control while missing entirely whether the business got more capable.
Governance and capability aren't in tension — but capability has to be the primary lens. None of this is an argument against visibility, contracts, or cross-functional alignment; an organization that can't see its AI spend can't make good decisions about anything. But when cost visibility becomes the primary measure of AI success, organizations end up optimizing the wrong variable — and repeat the same ROI disappointment that hit the first wave of automation, where deflection numbers looked great and the business didn't actually get faster, cheaper, or more capable.
The next phase of enterprise AI shouldn't be defined by how well organizations govern what they consume. It should be defined by whether they can prove what that consumption made them capable of.