Devendra Rajput, Technical Architect Lead at Accenture.
Ask most enterprise technology leaders how their cloud cost program is doing, and you’ll likely hear the same answer: “We’ve done the hard work. We rightsized instances, negotiated committed-use discounts, scheduled non-production environments to shut down at night and built dashboards that give finance and engineering a shared view of the bill. The low-hanging fruit is gone.”
And yet the waste is climbing again. Flexera’s 2026 State of the Cloud report put estimated wasted cloud spend at 29%, the first increase in five years. That should tell every leader something uncomfortable: the playbook that got us this far has quietly reached its ceiling. The next gains won’t come from a better dashboard or a sharper discount. They’ll likely come from autonomy.
After two decades running large, mission-critical cloud and database estates, I’ve watched this evolve through three distinct waves.
The Three Waves Of Optimization
The first wave was manual: humans read the bill, spotted the anomaly and filed the ticket. It worked, but it was slow and impossible to scale. The second wave was rules-based automation: autoscaling, scheduled shutdowns, threshold alerts and reservation recommendations. This is where most mature organizations sit today, but rules encode yesterday’s patterns. A static rule can’t reason about a workload it has never seen, and it breaks the moment the environment shifts.
The evidence that this wave has plateaued is hiding in our own utilization numbers, where compute routinely runs well below capacity. Datadog found that 83% of container costs go to idle resources, even in well-instrumented environments. This isn’t waste born of ignorance; these teams already have the dashboards. The waste persists because there’s a gap between knowing and doing, measured in human latency no dashboard can close. The third wave is agentic, and it exists precisely to close that gap.
What Agentic AI Actually Changes
The difference between the second and third wave is the difference between rules and reasoning. A rule fires when a threshold is crossed. An agent observes continuously, diagnoses why the cost is what it is, weighs the trade-offs of acting and then either executes a remediation within defined guardrails or escalates it to the right owner at the speed and scale of the workloads it governs.
This matters most as AI becomes a major new driver of cloud cost. The FinOps Foundation reports that 98% of teams now manage AI spend, up from 31% two years ago, and AI workloads are the least amenable to static optimization. Their cost scales with the complexity of a decision, not predictable traffic: a single request can fan out into model calls, retrieval, tool use and retries. You cannot write a fixed rule for a cost curve that reshapes itself hour to hour. That makes AI particularly well suited to help optimize the AI-driven spend it is helping create.
Design For Efficiency: Spend Intelligence Only Where It Pays
Here’s the irony: an agentic cost program can become a cost problem of its own. Piping every metric and log line through a large model can rebuild the waste it was meant to remove. The discipline is one we already preach to everyone else: spend intelligence only where it changes the outcome.
Collection and normalization are solved, deterministic problems. Billing exports, tagging, infrastructure-as-code state and telemetry can be pulled by a plain automation framework on a fixed schedule, with no reasoning required. Keeping AI out of this layer is one of the most important levers for controlling the cost of the system itself.
Reasoning is where AI earns its keep: diagnosing why a bill moved, correlating a spike to a deployment and weighing trade-offs. The trick is invoking it selectively, on an anomaly or scheduled review, not on every data point. Anticipation is where intelligence can pay for itself again: forecasting spend and projecting where a workload’s cost curve is heading moves the program from reactive to proactive. Action closes the loop back in deterministic automation: once approved, execution is scripted and auditable, with humans signing off on anything high risk.
Designed this way, intelligence can concentrate where it compounds value, while the plumbing that feeds it stays cheap. You end up applying FinOps to the FinOps program itself, ensuring the cost of optimization remains proportional to the value it creates.
From Cheaper To Smarter
The frontier isn’t only about spending less; it’s about changing what “optimized” means. For years we measured success in dollars per instance, per hour. Agentic optimization pushes toward a better question: What is the cost per useful outcome, whether that is per ticket resolved, per document processed or per transaction completed?
A workload that costs more but delivers disproportionately more value isn’t a problem to cut; it’s an investment to scale. When optimization is continuous and outcome-aware, cloud spend can stop being a number the CFO wants smaller and become a unit of economics the business can deliberately improve.
How Leaders Should Approach The Frontier
Reaching this frontier is a leadership exercise more than a technical one. Three principles have served me well.
First, optimize value, not just spend. If your only metric is total dollars, autonomous tooling will help you cut faster in the wrong places.
Second, earn autonomy incrementally: let agents act on low-risk decisions first, with humans approving the rest, widening the boundary only as the system proves trustworthy. Autonomy without guardrails and audit trails isn’t optimization; it’s exposure.
Third, treat your best judgment as an asset to encode. Your best FinOps practitioners’ instincts are the raw material for this next wave, and the organizations that pull ahead will capture that expertise as policy an agent can act on, rather than leaving it locked in a few people’s heads.
The frontier of cloud cost optimization was never really about finding more savings to squeeze out; those have been visible on our dashboards for years. The real frontier is removing the human latency between seeing a problem and solving it, and that’s exactly what agentic AI can help make possible. The leaders who cross it first aren’t just the ones I believe will spend less; they’ll build organizations that spend well by default.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?