Someone just checked the AI bill…

As recent dignomica articles have shown, there’s s a tokenomics crisis going on out there as some AI investment realism in the form of unexpectedly high bills hits home, leaving users questioning value derived and many vendors scrabbling to justify economic models.

More evidence of the issue this week in the form of a new study into AI and cloud cost management specialist Mavvrik, State of AI Cost Governance 2026. It notes:

ROI is math (Value-Cost/Cost): It needs a denominator. With an incomplete cost picture, that number of missing. Before anyone can answer whether AI delivers a return, they need confidence in what it costs.

In partnership with Benchmarkit, Mavvrik surveyed nearly 400 organizations in April and May this year, 47% of them in tech, and 23% in financial services. It found a troubling picture of emergency cost freezes, abandoned plans, forced re-pricing of customer products, and crises that demand board-level escalation.

In one particularly stark finding, one in four respondents have delayed or cancelled an AI initiative due to soaring costs. As the report puts it:

Bill shock is mainstream.

Upside

The good news is that 98% of organizations now track their AI costs – with many alarmed at the findings, it seems – while 95% have formal AI budgets. Despite this, only 11% can forecast their spending to plus or minus 10% accuracy, down from 15% last year. So, 89% of enterprises are unable to predict their AI costs at all.

That widening gulf between tracking and forecasting is the report’s core concern – and, lest we forget, it is also Mavvrik’s stock in trade. So, what is the problem? The report explains:

Tracking is only as useful as what it covers, and the fastest-growing categories – agents, GPU infrastructure, and on-prem – sit largely outside the frameworks doing the reporting. You cannot predict what you cannot see.

It adds:

AI spend has pulled ahead of the discipline meant to govern it, and the distance has widened on every front…with visibility falling off sharply as spend expands into agents, GPUs, developer tooling, hosted models, data platforms, and private infrastructure.

It’s an important point, and we should set it beside the findings of other recent reports, such as that infamous 2025 MIT NANDA research which found 95% of AI projects failing to demonstrate a measurable ROI, and a recent Economist Enterprise report which revealed that agentic AIs are causing problems in 98% of organizations.

Indeed, the Mavvrik report confirms the Economist’s estimate of agents’ near ubiquity: 98% of respondents run agentic workloads, it says, but only 36% include them in cost reporting. That’s an astonishing figure, and those costs presumably don’t include the broken business processes identified by Economist Enterprise.

Downside

This creates an uncomfortable conclusion. While most organizations have formal AI budgets and say they track all the costs, the reality is different – most have no idea what tokenomics is costing them in the real world. Meanwhile, the partial view they do have persuades many to cancel projects, freeze spending, and escalate problems to the board.

This gives the lie to the naïve but socially amplified belief that AI ushers in a brave new world of easy wins, lower costs, and productivity growth. Indeed, it is an expression of those two paradoxes that govern the AI age, once we strip away the hype – The Productivity (or Solow) Paradox, whereby as more investment is made in IT, worker productivity may go down instead of up, and Jevons Paradox, whereby tech that improves efficiency lead to a rise, rather than a fall, in total consumption. Between them, they tell us that introducing a new technology at scale undermines the benefits it creates by causing entirely new problems and costs.

A key challenge for decision makers is that those problems reside in different areas of responsibility and, therefore, of corporate governance. For example, when it comes to AI in products, customers drive the cost in terms of gross margin, with the key metrics being cost-per-inference and cost-per-customer. When those costs are larger than anticipated, the result is margin erosion and forced re-pricing – in short, your customer finds dealing with you is becoming more and more expensive.

But when it comes to AI in workflows, however, employees drive the cost, in terms of headcount decisions and the organization’s assumptions about productivity gains. The latter are largely rooted in vendor promises and (as we have seen from other reports), may never appear, or may be counter-balanced by complexity, lack of skills, emergent business problems, and un-addressed errors in AI outputs.

In this case, the key metrics are the costs per workflow and per seat, but hidden costs relate to un-validated ROI, un-budgeted expenses, and the un-predictable impact of errors and hallucinations.

Recent examples of the latter include the 1,900-plus cases of lawyers presenting fake AI precedent in court – a figure that is rising by over 100 a month – or the services giants that have withdrawn or redacted enterprise reports, due to hallucinated citations  – see diginomica, passim.

Indeed, cases like this reveal another business challenge – AI’s supposed productivity gains may only exist, in transactional terms, if the output is never checked. Doubtless all those lawyers and services professionals believed they had saved time and money by using LLMs (Large Language Models) to generate critical work, yet failure to check it cost them dear.

Where cost lies

But when it comes to agentic AI, the Mavvrik report warns that the key cost driver is task complexity, which is even harder to calculate, let alone to forecast.

Economist Enterprise warns us that agents’ hidden costs relate to broken business processes – where the damage may be reputational as well as financial. Meanwhile, Mavvrik cites re-try loops – the cost of repeated attempts to perform a task – plus orchestration overheads. And as we have seen, the report also tells us that while 98% of enterprises are using AI agents, only a third include them in their cost reporting.

Step back from all this, and we can see that many AI costs are not only invisible or difficult to forecast, but also hard to attribute to the right part of the organization or the appropriate page of the budget. The report explains:

Product AI and workflow AI look the same on an invoice. [But] one scales with customers and hits gross margin [the report cites 49% of respondents repricing an AI product in response to rising costs], while the other scales with employees and accumulates as unvalidated OpEx.

Where the AI sits in a customer-facing product, it should be treated as COGS (Cost of Goods Sold), but when it sits within internal workflows, it is operational expenditure that scales with employee activity. In the latter case, the knock-on effect is invisible spend accumulating against productivity assumptions that no one has tested, says Mavvrik. Put another way, you may have some idea of who spent the money, but not why the program cost so much, or what its measurable ROI might be.

Lumping all these costs together produces a number that may be accurate in aggregate, but will be “useless” in decision-making, adds the report.

My take

Taking all the findings together gives us a picture that bears no resemblance to the easy wins, savings, and productivity gains that so many vendors promised in the earliest days of the AI hype cycle. Cost-forecasting is currently getting worse, not better, which means that more and more AI costs are unexpected, or unseen by the business until it is too late.

This is why nearly two-thirds of organizations (62%) report that a “cost surprise” has forced a material change to a business decision, with one in four saying an AI program was abandoned as a result.

Despite this, 44% of organizations do believe they are “advanced” in AI cost management. But if that is true, why are 89% of them unable to forecast their expenditure accurately? And why are 81% seeing a moderate to high gross margin impact (a negative one)?

But perhaps the most troubling finding is that the fastest growing categories of AI adoption, such as agentic systems, are the least visible to decision-makers in cost terms. Mavvrik tells us that while 98% of respondents are running AI agents, 15% are unable to attribute the costs internally.

And when it comes to vibe-coding, that other phenomenon which promises easy wins and productivity gains, 98% of respondents use AI copilots, but nearly 40% of those tools exceed the expected cost. Indeed, the report warns that, in many cases, developer tool costs are “invisible to finance.”

So, what’s the core message of this bleak, but useful document? Cloud-era visibility was a property of the platform, says Mavvrik, but AI cost visibility must be built.