Even Uber says the free ride on AI is ending. CFOs need to measure ROI before the bill hardens.

In December 2025, Uber gave roughly 5,000 engineers access to Claude Code. Adoption was measured and ramped steadily, and by April 2026, the company had reportedly exhausted its annual AI coding budget four months into the calendar year. Goodhart's Law states that when a measure becomes a target, it ceases to become a good measure — and we're seeing that play out across numerous companies right now. AI adoption as a metric isn't wrong when adoption itself is fundamental to transformation, but tokenmaxxing and leaderboards like those at Uber and Meta will be looked upon by history books as evidence of AI hysteria. Amid the public acknowledgment of the overrun at Uber, CTO Praveen Neppalli Naga commented to The Information that the company would need to entirely revisit and interrogate its assumptions around AI investment and value.

Estimates put AI coding costs somewhere between $500 and $2,000 per engineer per month, which means an organization can burn through a budget line that seemed reasonable in January before Q2 closes. In Uber's case, it wasn't a small team running an unsanctioned experiment; it was one of the most technically sophisticated companies in the world, with mature engineering discipline, and their leadership still got surprised by the variable costs embedded in daily workflows.

The lesson that tends to get drawn from this story is overly simplistic: implement cost controls, set limits, monitor usage, implement circuit breakers. Uber did or is doing those things; there are really smart people who work there! Their situation is a canary in the coal mine for all of us.

The issue is complex. We have access to a revolutionary technology that can be leveraged to produce revolutionary gains, so not using it is irresponsible. It's also not free; there is no way to right-size the investment because there is no way to value the return. This is the hardest budgeting exercise in our lifetime. It won't be as simple as spend on AI and the gains will come, and it also won't be as straightforward as simply capping spending without relying on data and insight which describes the value of that spend.

The stakes are also high. Not spending enough means your competitors will beat you handily on product, price, and profit. Spend too much, and lose your shirt.

When the music stops

The pricing environment that enabled Uber's situation wasn't accidental. Frontier model companies have spent the better part of four years subsidizing enterprise AI adoption — pricing below sustainable margins to build usage habits, win deployment commitments, and accumulate enterprise relationships. That strategy largely worked, as AI moved into daily enterprise workflows and organizations built tooling dependencies on top of models whose true cost was partially obscured in the process.

But that dynamic is shifting, and faster than most CFOs and budget-holders realize. OpenAI and Anthropic are among the most closely watched IPO candidates in technology, and public-market investors apply margin scrutiny that venture-stage capital does not. Anthropic has moved its enterprise customers to usage-based billing on top of a monthly base fee, and GitHub moved Copilot plans for organizations and enterprises toward usage-based billing with AI credits becoming the core billing unit.

Finance teams should assume that today's AI bill is not the permanent floor — it is the introductory rate. Whatever their current AI budget calculus is, it likely won't survive a collision with 2027 prices. Importantly, rising prices don't just inflate the bill — they raise the burden of proof. At 1x subsidized pricing, fuzzy AI ROI is tolerable. At 2–3x, AI goes from CapEx to OpEx.

Infrastructure commitments should require infrastructure-grade measurement

JPMorgan offers the clearest example of what it looks like when an organization takes AI seriously as a financial category. The bank has reportedly reclassified roughly $2 billion of annual AI spending as core infrastructure within a ~$19.8 billion technology budget — the same planning category as data centers, payments infrastructure, and cybersecurity.

More instructive than the scale is what JPMorgan acknowledged alongside it: Jamie Dimon has outright said that measuring AI's precise impact isn't straightforward. The institution committed to infrastructure-level investment without accounting-grade measurement in place. They relied on directional evidence, productivity signals across 150,000 employees, and linked operating savings, but lacked the workflow-level value attribution that would make the case airtight. Even with their analytical depth and considerable resources, they still can't close the loop. They believe the bet is worth it and I do, too.

JPMorgan can afford to take bets you can't, and they can afford to lose in ways you probably cannot, too. Organizations we speak with are navigating the same gap with considerably less rigor. They have deployed AI broadly, accumulated adoption metrics, and conducted internal surveys showing employees find the tools useful. When this interim answer accommodates a need, it's a deadly trap. That evidence base won't support the same kind of budget conversation when pricing normalizes and embedded AI commitments come under scrutiny.

A lot of executives I speak with shake their heads and wonder what bets are even worth making in a market moving so quickly. I have a strong opinion. No company should commit to one model provider and build institutional context around them specifically or any AI point solution that comes with a long duration contract.

The defensible bet to be making is on your enterprise data, because that is the input everything else requires. The best model isn't useful without the right data structured and applied intelligently. Not enough people are taking this bet seriously. The investments that firms made to move their data to modeled warehouses in Snowflake and Databricks does not solve this problem; really, how could a pre-AI infrastructure investment conceivably solve this?

Cost visibility is the easier problem — not the most important one

Recent research found that 80% of enterprises miss AI infrastructure forecasts by more than 25%, while nearly a quarter miss by 50% or more — with meaningful gross margin erosion often running six percentage points or more.

Cost visibility and value attribution are different problems. Both merit attention, but the second is harder to quantify, and necessary to create and reinforce the virtuous feedback cycles that create value from AI investment. At the end of the day, executives and investors want internal AI to improve revenue per employee through improved output — velocity, quality, productivity. They want operating leverage. Most executives I speak with want more revenue without being forced to hire more employees as a result, and that optionality is everything to a leader. Revenue per employee can be gamed, however. The layoffs at Meta and Block are exactly this. Despite the headlines, AI had nothing to do with them; they were entirely performative.

So how do we know that AI itself is actually responsible for an improvement in revenue per employee? What is required is attribution. In advertising, it used to be said that half of advertising is wasted — the problem is that we don't know which half. In advertising we largely solved this with multi-channel attribution, and now that truism is no longer true.

AI is the same, and Parable can attribute how AI tools and agents do or do not create positive operating leverage on the work of an enterprise team. At the end of the day, people spend time and AI spends tokens. You have to be able to attribute how each unit is spent, contextually, to understand if you are creating a better working environment with more leverage.

At Parable we call the process of creating this dynamic measurement loop value harvest — where saved capacity or cost is captured then converted into other generative activities which in turn are also measured.

What finance leaders should do now

The enterprise response to rising AI costs has split into two positions, and both miss the point. The first: AI has transformative potential, so cost doesn't matter yet. The second: AI is expensive, so we will clamp down on usage until ROI is perfectly measurable.

A more practical posture is to separate AI systems into categories — experiments driving innovation, governed by strict stage gates and kill criteria; and foundational systems that deserve infrastructure-style planning and measurement, ongoing optimization loops, and clear ownership.

For CFOs and other executives looking at budget, that means four concrete priorities:

• Track AI costs more granularly, at the workflow or product level — not just at the vendor level. This must go beyond discrete processes where ROI is easy to calculate like ticket resolution. That alone won't equal transformation.
• Measure value with directional operational metrics and outcomes where AI is clearly in the critical path: engineering throughput, service productivity, cycle time, customer retention, revenue growth, and compounding efficiencies across teams — not individual productivity.
• Assume usage-based pricing will expand rather than contract, and build scenario ranges rather than single-point forecasts.
• Decide explicitly which AI systems are experiments and which are infrastructure — they should both be held accountable but should not be governed the same way.

Much of the signal needed for measurement already exists across the tools organizations use every day: workflow activity, time allocation, communication patterns, system interactions.

What's missing is the intelligence layer that structures that data, surfaces relationships across it, and produces the kind of before-and-after evidence that makes an infrastructure case credible. Parable turns pattern detection into margin protection: we connect the data that exists to the operational outcomes that matter, so finance has something more defensible than adoption surveys when the budget conversation gets harder in the boardroom.

There is something clarifying about the subsidy ending. When AI was cheap enough to run indefinitely, there was no forcing function to distinguish genuine infrastructure investment from well-intentioned experimentation. CFOs who respond to the repricing by working with their transformation leads to architect real measurement frameworks — who use the pricing shift as a reason to instrument value, not just cost — will end up with AI budgets and strategy they can defend in any environment, plus business discipline that sets them up for success in the AI future.