AI is reshaping how platform teams operate - but it's also introducing a new class of financial challenge that traditional FinOps practices weren't built to handle. A recent survey of 130+ platform engineers reveals that while AI cost optimization ranks as a top-three priority, most teams lack clear frameworks for measuring ROI or managing spend effectively. This post breaks down a practical, three-phase framework to help you take control.

Main insights

• Traditional "shift-left" cost planning fails for AI because usage is spiky and unpredictable - you need continuous measurement and optimization instead
• Rate optimization through automated commitment management remains the biggest cost lever, yet it's often the last thing teams consider
• Measuring AI impact requires going beyond vanity metrics like lines of code to understand actual business value and behavior patterns
• The "shift right for AI" framework combines three phases: Optimize to create budget, improve ROI through behavior mapping, and reinvest for growth

Grace Gui, Senior Product Marketing Manager at ProsperOps, joined Josh Golver, a cloud FinOps specialist at ProsperOps, and David Tepper, CEO and co-founder of Ascerta, to share practical strategies for managing AI spend in a world where costs are unpredictable and leadership pressure to demonstrate value is only increasing.

You can watch the full discussion here if you missed it.

The AI cost optimization landscape: What platform teams prioritize

The platform engineering survey conducted in May-June 2026 revealed telling patterns about how teams approach cost optimization. Workload optimization and waste reduction topped the list, followed closely by AI cost optimization and enabling automation for FinOps.

What stands out is what ranked last: commitment management - the practice of using savings plans, reservations, or committed use discounts to reduce the rate you pay for compute. As Grace put it, "It remains to be the biggest lever for the rate you pay." That gap between impact and attention is significant.

The survey also surfaced a structural tension: while C-suite executives drive cost strategy with high influence, 51% of platform engineers reported feeling neutral about or actively disliking cost optimization work. Add to that the fact that nearly a third of teams don't measure internal developer platform (IDP) impact at all - and those who do rely on a patchwork of cloud-native tools, open-source solutions, and spreadsheets - and you have a recipe for misalignment at scale.

Why traditional cost planning breaks for AI

A senior platform engineer from a cloud-native enterprise SaaS company captured the frustration well: "Two years ago AI wasn't a thing, today it is. My desirables right now: traceability and unit economics. How do I get this info?"

Traditional forecasting assumes you can model usage patterns upfront, find budget by cutting elsewhere, or wait for the next budget cycle. You try to control spend before it happens. AI breaks this model entirely. Usage is spiky and can change overnight based on:

• Use case modifications
• Model upgrades
• Subtle shifts in how agents behave
• Dynamic, runtime decisions made by agents themselves

As David explained, "When people open up the tools and they start leveraging AI, it's doing its own live research. It's calling a bunch of its own tools. It's taking activities on the fly. And so you have to have governance at the edge."

This is the core argument for shifting right - moving from upfront planning to continuous measurement, optimization, and adjustment.

The shift-right for AI framework

The framework operates across three interconnected phases that feed into each other continuously.

Phase 1: Optimize to create budget

Rate optimization through automated commitment management is the fastest path to freeing up budget for AI initiatives - without requiring engineering teams to cut features or delay projects.

Josh outlined three key strategies:

Portfolio management determines the right mix of reservations, savings plans, and commitments. Because this mix shifts constantly as infrastructure evolves, automation is essential to stay optimal.

Adaptive laddering replaces large, infrequent commitment purchases with small, frequent "rungs" - potentially weekly or more often. Josh described a two-week graph where vertical lines marked individual optimization actions: exchanges, merges, splits, and adjustments. "If you stack this up over multiple weeks, months, year, this would certainly be a lot to do on an individual basis or DIY. This would certainly be a full-time job."

Cyclical optimization addresses the reality that most organizations have predictable compute patterns - workloads that scale down at night or on weekends. Traditional approaches cover only the baseline, leaving usage spikes as expensive on-demand spend. Automation can cover significantly more by intelligently leveraging hyperscaler rules.

One example from the session illustrated this well: an organization migrating half their spend to a different hyperscaler saw monthly usage drop from nearly $2 million to under $1 million over six months. Automated commitment management tracked that decline continuously, optimizing coverage the entire way down. "We never want to handcuff engineering flexibility or changes with commitments," Josh emphasized.

The benchmark to target is effective savings rate (ESR) - savings generated divided by pre-discount usage. Reaching above 50% puts you in the 90th percentile. Without automation, most teams hit a ceiling around 50% regardless of spend level or headcount dedicated to the problem.

Phase 2: Improve ROI through behavior mapping

Once you've created budget headroom, the next challenge is proving that AI investments actually deliver value. This requires moving beyond raw telemetry metrics.

David was direct about the problem: "Lines of code, commits, PRs - all of this type of stuff doesn't correlate to value. Those are just usage characteristics." He offered a pointed example: "If you ask Claude Code to write you a poem, every single line in that poem is going to be reported in the telemetry as a line of code because that's how it operates."

Ascerta's approach builds behavior maps by categorizing each action developers take with coding agents:

• High-value behaviors - debugging complex problems, accelerating architecture decisions
• Low-value behaviors - asking an agent for the current time in Uganda (easily found on Google)
• Non-work-related usage - activities that burn through expensive model calls without business benefit

The platform surfaces statistically significant patterns across your organization automatically. In the demo, David showed how one team had spent $7,500 delegating an agent to manage their email - a signal that they might need purpose-built tooling rather than expensive coding agents handling inbox management.

Measurement must also account for how modern AI systems actually work. "Most use cases use multiple models," David explained. "When you ask a question on ChatGPT, it's using at least three models under the covers. We find that about three is the industry average for a well-designed and orchestrated agent." That means measuring value at the use-case level, not the model level, and accounting for total cost of ownership across all tools and cloud resources involved.

Phase 3: Reinvest for growth

The framework becomes genuinely powerful when it runs as a continuous cycle. Savings from rate optimization fund AI initiatives. Behavior mapping identifies which initiatives deliver value and which need adjustment. Those insights drive better usage patterns, which can then be optimized further.

ProsperOps recently released unified autonomous optimization, combining commitment management with spot management, bin packing, and pod and node right-sizing in a single coordinated system. Josh explained why coordination matters: "Historically, rate and workload have essentially worked against each other in some cases if not coordinated properly. Your autoscaler scales down at night but maybe the FinOps team wasn't aware and they have a commitment on that capacity and now that commitment goes unused." A unified approach eliminates that friction.

Practical implications for platform teams

The shift-right mindset requires different processes and tooling than traditional cloud cost management. You can't pre-plan everything when agents make runtime decisions. You need systems that:

• Track AI usage across both structured agents with specific goals and unstructured everyday AI use
• Identify waste automatically - from cold cache patterns to non-work activities
• Surface actionable insights about what's working and what isn't
• Measure value based on business outcomes, not vanity metrics
• Operate asynchronously without impacting agent performance

As David noted, "It's almost a misnomer to assume that there is a set of metrics that's always going to be true that you use to measure success and impact. It's changing all the time as the agents change, as the models change."

The good news: you don't have to choose between innovation and cost control. Automate rate optimization to free up budget without engineering effort. Measure AI behavior patterns to scale what works and fix what doesn't. Treat this as a continuous cycle rather than a one-time project, and you build the foundation for sustainable AI transformation.

​

If you enjoyed this, find more great insights and events from our Platform Engineering Community.

If you want to dive deeper, explore our instructor-led Platform Engineering Certified Professional course and connect with peers from large-scale enterprises who are driving platform engineering initiatives.

​

Key takeaways

• Automate rate optimization first - Commitment management through savings plans and reservations remains your biggest cost lever, yet most teams tackle it last. Automation can achieve 50%+ effective savings rates without restricting engineering flexibility, creating immediate budget for AI initiatives without waiting for the next planning cycle.
• Measure AI value, not just usage - Raw metrics like lines of code or commit counts don't correlate to business value. You must track behavior patterns to identify high-value activities worth scaling versus low-value or non-work usage that drains budget without return.
• Adopt continuous optimization over upfront planning - AI's unpredictable, spiky nature breaks traditional forecasting. Success requires shifting right - measuring, optimizing, and adjusting continuously as agents make runtime decisions and models evolve.
• Align teams on standardized ROI metrics - Finance, product, platform teams, and leadership need a shared view of AI impact tied to business outcomes. That alignment enables faster decision-making and more effective resource allocation across AI initiatives.