AI Spend Is Easy to Count. AI Leverage Is Harder to Prove.

AI spending has become visible quickly. Finance teams can now see license growth, model usage, inference charges, and tool proliferation with far more precision than they could a year ago. What remains difficult is the more important question: what is that spend buying?

In many organizations, AI adoption is being narrated through activity metrics. Seats provisioned. Prompts run. Assist completions accepted. Those measures have value, but they are not evidence of leverage. They show access and engagement, not whether engineering performance has actually improved.

Spend is not leverage.

A tool can make drafting easier while moving more work into review, validation, or correction. The investment question is whether the complete workflow improves. Measure the accepted result and the effort required to reach it, alongside the cost of the tool.

METR’s July 2025 randomized study illustrates why assumptions need testing. Its 16 experienced open-source developers took 19% longer on tasks when allowed to use early-2025 AI tools, despite expecting a speedup. That result applies to the studied developers, repositories, and tools; it is not a general verdict on AI assistance.

In its February 2026 follow-up, METR said selection effects and time-measurement problems made its newer experiment an unreliable estimate of current AI productivity effects. Our practical conclusion is to evaluate your own workflows, keep the measurement limits visible, and revisit the baseline as tools change.

A company can therefore increase AI spend and still fail to create AI leverage. That happens when engineering output becomes noisier, when review queues expand, when rework rises, or when teams generate more code without improving throughput.

The scorecard that matters.

The fix is not to ignore spend. It is to pair spend with a fuller operating scorecard. In practice, leadership should measure AI economics across four dimensions.

• Spend: licenses, inference costs, and workflow-specific tooling overhead.
• Adoption: where AI is being used, by whom, and in which parts of the development lifecycle.
• Effectiveness: whether AI-assisted work correlates with better throughput, lower review burden, or less rework.
• Risk: whether AI usage is increasing duplication, weak testing, policy exceptions, or architecture drift.

These four dimensions matter because they prevent a common reporting mistake. Leaders often see spend and adoption because those measures are easiest to gather. They rarely see effectiveness and risk with equal clarity, even though those are the dimensions that determine whether the investment should scale.

What leverage looks like.

Real AI leverage rarely appears as a single dramatic number. It appears as a pattern: faster review cycles without lower code quality, better throughput without higher remediation exposure, and wider developer support without collapsing standards or ownership.

That means AI leverage should be understood at the team and workflow level, not just at the enterprise total. One team may be using AI to reduce repetitive work and documentation overhead with strong gains. Another may be increasing review churn by flooding the system with poorly structured code. Aggregate spend hides that difference.

The right question is not “How much are we spending on AI?” It is “Which engineering outcomes improve when AI is present, and which ones get worse?”

Why finance and engineering need the same view.

CTOs need the answer because AI is changing engineering management. CFOs need the answer because AI is becoming a material line item whose value cannot be justified by enthusiasm alone. Both functions need a shared frame that distinguishes productive adoption from usage that adds cost without improving delivery.

The companies that handle this well will not be the ones with the highest rollout percentage. They will be the ones that can identify where AI reduces review load, rework, or delay, where it is neutral, and where it quietly introduces future cost.

What the system must measure.

The management challenge is not to track AI tools in isolation, but to relate AI-assisted engineering behavior to cost, throughput, review efficiency, and risk across the full software delivery system.

The AI Engineering Readiness Assessment connects individual coaching and codebase readiness outputs to that shared view of cost, flow, quality, and risk.

That is the distinction Binomial is designed to make legible. Reporting AI activity is easy; managing AI economics is harder, and materially more valuable.