AI spend is following cloud's adoption curve at twice the speed, with a quarter of it wasted and most organizations still without a clear owner.
AI spend is climbing the same adoption curve cloud spend took roughly a decade to climb, compressed into a fraction of that time. Cloud FinOps had years to mature its tagging, allocation, and forecasting habits before anyone seriously asked what the spend was worth. With AI, leadership asks the ROI question before the rollout even finishes. Harness's own State of AI in FinOps 2026 report, based on a survey of more than 700 engineering leaders and FinOps practitioners, puts numbers on the gap: organizations estimate that roughly a quarter of all AI spend is wasted, and only a small fraction of organizations have what the panel called basic AI cost visibility.
Unit Economics, Ownership, and Attribution Separate AI Spend From Cloud Spend
The panel's framing was that three things make AI cost fundamentally different from cloud cost: unit economics, ownership, and attribution. Cloud spend scales fairly predictably with infrastructure that mostly costs the same whether it is idle or busy. AI spend varies by model, by prompt, and by task, which makes it a different kind of problem to budget for, not just a bigger one.
Unrealistic Budgets and Split Ownership Are the Most Common Pain Points
Elliott's read from customer conversations was that two problems compound each other. Budgets set at the start of the year are often disconnected from how spend actually tracks, and he described customers who blow through an annual AI budget within two months. At the same time, ownership of AI spend is spread across multiple divisions, and without one clearly accountable owner, nobody is positioned to catch that kind of overrun early. The problem is compounded further because AI usage is no longer confined to engineering. Marketing, legal, and sales are all adopting it too, which makes the spend harder to contain from a single vantage point. A CTO is often the only person positioned to see both the FinOps side and the AI usage side at once, since those functions haven't traditionally been unified.
Cloud FinOps Never Had to Answer "What Are We Getting for This"
Brogan drew on his own background as a FinOps practitioner at Wells Fargo and JPMorgan Chase to make a specific point: in that world, nobody asked what value the organization was getting for its cloud spend. The ceiling of ambition was keeping infrastructure cost flat or bringing it down a little. AI spend starts from a different question entirely, what the organization is getting for the investment, on day one of a rollout rather than years into one.
Value Has to Be Defined Before It Can Be Measured
Because "value" means something different at every organization, Brogan's recommendation was to spend deliberate time upfront defining what good looks like before trying to measure ROI. For one company that might mean more productive engineers shipping code and resolving tickets faster. For another it might mean more experimentation, more products reaching market, or faster, better customer interactions. There is no universal definition, which is exactly why skipping this step leads to measuring the wrong thing later.
Engineers Can't Factor In Costs They Can't See
Brogan cited the State of AI in FinOps report directly on this point: just over half of respondents said their engineers understand what their builds actually cost, and 45% of engineering leaders said their teams don't factor cost into how they build AI solutions at all. His read was that engineers care about cost but simply don't have the data in front of them to act on it, especially now that the volume of data produced by AI tooling has already outpaced traditional cloud billing and usage data in scale.
Finding Why a Cost Spike Happened Can Take a Full Day or More
The report's numbers on spend anomalies were, in Brogan's words, stark: a large majority of respondents had experienced an unexpected AI spend spike in the prior year, and nearly 80% said they would need a full day or longer just to trace that spike to its root cause, before any work to actually fix it begins. Of the organizations that had to resolve one of these spikes, only about a fifth could do it within hours. Roughly 40% needed a full day, and close to a third needed a full week. Visibility alone doesn't close that gap. Brogan's point was that teams need some level of automation and autonomy layered on top of visibility to actually act on what they can now see.
Harness's AI Cost Dashboard Tracks Adoption, Impact, and Cost Side by Side
Elliott's demo of the Harness AI cost management agent walked through three linked views. Adoption tracks how many developers are using AI coding tools and what share of code is AI-generated. Impact tracks the drop-off between code an AI tool generates and code that actually survives into a pull request, since a meaningful share of AI-generated code gets discarded before it ships. Cost ties a dollar figure to that same funnel, tracking what was spent generating code against what portion of that spend produced code that was actually committed. In one example from the demo, that drop-off showed AI code being discarded well over 60% of the time, a number worth investigating team by team rather than assuming it reflects AI quality alone.
Plain-Language Queries Surface Which Teams Are Actually Improving
Beyond the dashboard's fixed widgets, Elliott demoed asking the agent direct questions in plain language, such as which teams had improved the most. One example answer showed a team's AI and machine learning adoption growing from a single user to four, alongside a 78% increase in pull request velocity for that team, the kind of answer that would otherwise take a manual report to produce.
AI Can Speed Up Coding and Slow Down Review at the Same Time
Elliott's demo also surfaced a less intuitive effect: AI-assisted coding tends to produce larger pull requests, and human code review doesn't necessarily get faster just because the code arrived faster. Some pipeline stages speed up measurably while review, still mostly a human step at most organizations, can take longer for larger AI-assisted changes. That is a shift in where time is spent across the pipeline, not necessarily evidence that AI is slowing teams down overall.
A Developer Leaderboard Separates High Performers From People Who Need Training
The platform's leaderboard view lets a team pivot AI usage data by developer, repository, or application to see who is combining AI usage with strong output and who is using AI heavily without much to show for it. Elliott's framing was that this isn't about choosing whether to lift up the least productive engineers or push the most productive ones further. The same view supports both: recognizing top performers and identifying exactly who needs more training or better tooling to get there.
Anomaly Detection Connects a Cost Spike to the Specific Transaction Behind It
Elliott demoed the platform flagging a cost anomaly tied to a specific AI model, then drilling into the responsible component, a simulated flight-rebooking agent, down to individual transaction cost percentiles. Where OpenTelemetry tracing is available, the platform can show the actual exchange between an agent and its underlying model alongside business context like loyalty tier or route type, letting a team connect a cost spike directly to a customer segment or transaction pattern instead of just a dollar figure on a chart.
Budgets Can Span Multiple AI Vendors Under One Governance Layer
Because most organizations use several AI coding tools rather than a single vendor, Elliott described budgets that can be set once and enforced across tools like Claude and Cursor together, with usage warnings, hard blocks at 100% of budget, and multi-step approval workflows for raising a limit. The goal is one governance layer instead of learning each vendor's own, inconsistent budget controls. Enforcement still depends partly on what a given vendor's own platform allows. Where a vendor can't support a hard limit at the right granularity, the system falls back to softer reporting and notification instead of a true block.
Richer Billing Exports and Tracing Are Closing the Attribution Gap
On attributing AI spend down to a team or project, Elliott pointed to practical steps available today: newer, richer cloud billing exports, such as AWS's CUR 2.0, already include fields like principal ID that identify which user generated a given charge. Traditional tagging still helps, but AI vendors tend to add proper cost-allocation support retroactively, after usage has already scaled, which makes it hard to get engineering teams to go back and tag everything correctly after the fact. OpenTelemetry helps close part of that gap for production agent usage specifically, since it can surface fields like agent name with comparatively little manual instrumentation.
Customers Chose Harness's AI Cost Visibility Over Building Their Own
Brogan shared two unnamed customer examples that landed on the same conclusion from different starting points. One of the two largest grocery chains in Australia, a longstanding Harness FinOps customer, was partway through building its own in-house visibility tooling for developer AI spend when it saw an early version of Harness's AI cost management capability and abandoned its own build entirely. A top-three US retail bank, also a longstanding customer, had already built similar visibility in-house across tens of thousands of developers, but found that the real cost wasn't building the tool. It was maintaining it. The bank had three to four full-time employees dedicated just to manually keeping the homegrown system fed with data and running correctly, and chose to retire it in favor of expanding its relationship with Harness instead.
Traditional Cloud FinOps Still Delivers Value Alongside AI Cost Management
Brogan was careful to note that the AI conversation shouldn't crowd out the value still available from the kind of cloud cost management Harness has offered for years. United Airlines, a longstanding Harness customer, used autostopping and asset governance automation to cut cloud spend by 15%, an eight-figure savings with no impact on how its engineers work, and now tracks unit cost per flight as a metric it shares with its own executives and shareholders.
The Single Biggest Mistake Is Managing AI Spend Without a Named Owner
Asked for the most common budgeting mistake, Brogan's answer was direct: treating AI spend exactly like cloud spend, without adapting for how differently it behaves, and without naming a single owner accountable for it. Harness's report found that 52% of organizations have no clear, dedicated AI cost owner at all. Among the organizations that do have one, ownership is split nearly evenly across finance, DevOps, engineering, and the business side, with no single function clearly in the lead. Brogan's recommendation was to decide deliberately how AI spend gets classified, whether as infrastructure cost or as a labor-like cost, and then assign one owner who can set strategy and guardrails across the organization, rather than leaving that decision to happen by default.
