Nobody knows what AI really costs

Enterprise AI spending is spreading from software licences and embedded features into consumption charges faster than traditional budgets can track it. The challenge is no longer simply controlling the bill but understanding which parts of a fragmented AI estate are creating enough value to justify what they cost.

The first generation of enterprise software was comparatively easy to count. A company bought a licence, assigned it to an employee and could usually work out what the application cost over a year. Software-as-a-service complicated that picture, but artificial intelligence is changing the economics again because the cost can sit inside products the business already owns or appear through separate tools and variable usage.

That makes AI sprawl difficult to recognise using application count. Zylo’s 2026 SaaS Management Index, which combines 24 months of anonymised platform data with a survey of 218 IT leaders at organisations with more than 1,000 employees, found that the average software portfolio was essentially flat even as annual SaaS spending rose by almost 8 per cent. AI features and changing patterns of consumption can make the estate substantially more expensive without adding many applications.

Cutting duplicate software will still matter, but the greater exposure may sit inside applications that survived consolidation. Premium AI tiers and variable model consumption can turn a familiar SaaS contract into an unpredictable expense, while internal development can disperse additional costs into budgets that are not labelled AI. The sprawl is no longer visible only as more logos on an architecture diagram.

Sprawl without more software

Zylo’s data shows how quickly the cost profile is changing. Average spending on AI-native applications more than doubled year on year, while the increase among organisations with more than 10,000 employees reached 393 per cent. At the same time, the overall number of applications barely moved. Enterprises can rationalise their portfolios and still spend considerably more because the software left behind is being monetised differently.

Embedded capability makes the problem harder to isolate. An organisation may not have deliberately acquired another AI product; a supplier can introduce intelligence into an existing platform and attach it to a more expensive edition or a usage meter. Zylo found that 77 per cent of surveyed IT leaders had encountered unexpected costs after a contract was signed, reflecting a move away from static seat counts towards pricing tied more closely to consumption.

BetterCloud reaches the same issue from a different dataset. Its State of SaaS 2026, based on a survey of 525 SaaS-first organisations, found that the average business now deploys 27 AI-powered SaaS applications, roughly 22 per cent of its portfolio. More significantly for budgeting, 36 per cent are billed through usage-based or token-consumption plans rather than straightforward per-seat pricing.

A successful deployment can therefore become harder to forecast as more people use it. Conventional licences create waste when seats sit idle, but consumption models create a different problem because greater adoption can drive the bill upwards throughout the contract. The organisation needs to understand not only whether the capability is used, but what that activity costs and whether it produces an outcome worth paying for.

The bill follows the work

Agents make the relationship between activity and expenditure more important. A person may see one completed task while the system behind it makes several model calls or invokes other services before producing the result. BetterCloud warns that unmonitored applications and agents can drain budgets through redundant accounts and uncontrolled token consumption, so the cost of automation cannot always be inferred from the number of people using the software.

Zylo describes the emerging model as closer to cloud consumption than traditional licensing. Value is increasingly measured through tokens and credits tied to compute, requiring organisations to monitor utilisation while the system is running rather than waiting for an annual renewal. Its report predicts that AI could become an expensive “invisible worker” as operational tasks move from predictable labour costs towards software consumption that is harder to attribute to a business outcome.

Consumption pricing is not inherently undesirable. Charging for activity can expose whether a tool is genuinely being used. The difficulty comes when enterprises cannot connect that activity with its result. A department can generate thousands of model interactions without finance knowing whether the additional consumption shortened a process or simply produced more machine activity.

J.R. Storment, Executive Director of the FinOps Foundation, compares the pattern with the early expansion of public cloud: “That same thing is happening with AI right now,” he says. “People are signing up with their credit cards.” Companies eventually brought cloud consumption under stronger financial management when decentralised experimentation reached a scale that could no longer be ignored. AI appears to be moving through a similar cycle.

Shadow AI creates a shadow budget

Employee purchasing makes the problem particularly visible. Zylo found expense-based SaaS spending increased sharply year on year, with ChatGPT becoming the most expensed application and the OpenAI API also near the top of the list. Small groups can acquire capability before central procurement knows it exists, creating a financial blind spot alongside concerns about data access and governance.

Low switching costs also encourage teams to test overlapping products while new capabilities appear quickly. Zylo found generative AI had already become one of the most redundant application functions in the enterprise, with organisations averaging almost seven tools in the category. Several small subscriptions can look harmless individually while duplicating capability across departments.

Flexera’s 2026 State of ITAM Report, based on a worldwide survey of 512 professionals with IT asset management responsibilities, suggests organisations are struggling to see the resulting estate. Only 31 per cent report visibility into AI software, while 59 per cent say wasted AI software spending increased over the previous year. Spending is rising in the area where organisations have the least confidence that they know what they own.

The mismatch complicates optimisation. Flexera found almost half of surveyed organisations claimed to be tracking usage and rightsizing AI contracts, yet the lower figure for overall software visibility raises an obvious question about what sits outside those exercises. Savings programmes can be rigorous within the known estate while unmanaged tools or embedded features continue generating spend elsewhere.

Cost control moves into operations

Traditional procurement concentrates on purchase and renewal. Consumption changes that rhythm because a commercially sensible contract can become expensive through the way it is used months later. Zylo argues that FinOps and IT asset management are consequently moving closer together, bringing the experience of managing variable consumption into contact with the people who understand software entitlements and commercial rights.

Ross Milne, Lead ITAM Consultant at Capgemini Invent UK, captures the shift: “As more SaaS publishers move away from traditional user-based annual subscriptions into a usage-based model, the crossover between FinOps and ITAM becomes increasingly important. Nobody can optimise a variable AI bill using contract data alone, just as consumption telemetry without an understanding of commercial rights provides only half the picture.”

BetterCloud’s findings suggest visibility has to reach beyond sanctioned applications. Only around a quarter of the organisations in its research had continuous automated discovery across apps and agents. A business cannot manage consumption or redundant accounts it does not know exist, particularly where premium capabilities can be enabled without appearing as a completely new application.

The objective should not be to force every experimental tool through a long purchasing cycle. That would address cost by slowing the experimentation from which useful applications emerge. A better discipline is to make experimentation visible early, establish ownership as usage grows and connect consumption to the business process benefiting from it. Tools that prove their value can then scale deliberately, while duplication and low-value activity can be challenged before they harden into another permanent budget line.

This is why AI sprawl is becoming a different problem from the SaaS sprawl that preceded it. The estate may not expand dramatically in application count while expenditure becomes more fragmented underneath. A feature embedded into an existing platform can add cost without adding another application, and an agent can increase consumption without adding another employee.

The financial question must move beyond how much the organisation spends on AI. Leaders need to know where that spending originates and which outcome it supports. The danger is not simply that artificial intelligence becomes expensive; it is that costs spread far enough through the enterprise that nobody can tell which part of the bill is buying value.