Tool Sprawl and the Meter: Where Enterprise AI Budgets Actually Leak

Ask a CFO what the company spends on AI and you'll usually get a number that's wrong in both directions. Too low, because it misses the add-ons buried in SaaS renewals and the subscriptions sitting on department cards. Too confident, because a growing share of that spend is metered, and metered spend isn't known until the invoice lands.

The data now backs up what budget owners have been feeling. Enterprise AI budgets leak in two distinct places: the number of tools, and the way those tools charge. Either one alone is manageable. Together they compound, and this month the second leak got noticeably wider.

Leak one: the stack nobody designed

Zapier surveyed 550 enterprise leaders at companies with 1,000 or more employees in late 2025. The findings read like an audit nobody asked for: 28% of enterprises now run more than 10 different AI applications, and only 35% of leaders say their AI tools go through proper approval channels. Three in four report at least one negative outcome from disconnected AI. And 30% say it plainly: they're wasting money on redundant AI software.

None of this happened by strategy. Marketing bought a writing tool, support bought a resolution bot, engineering bought two coding assistants, and half the SaaS stack switched on AI features and raised the renewal price. Zylo's 2026 SaaS Management Index found ChatGPT is now the single most-expensed application in corporate spend data, ahead of iCloud and Canva, and AI-native apps make up 16% of the 50 most-expensed applications. Expensed, as in: bought on a card, outside procurement, invisible to whoever supposedly manages the stack.

Each tool in that pile carries costs beyond its subscription: a contract to negotiate, a security review to run, a renewal to track, an integration to maintain, and a login your people forget about. The overlap is the expensive part. Two tools doing the same job means one of them is pure waste, and at 10-plus tools, overlap stops being the exception.

Leak two: the meters multiplied in June

The second leak used to be a quirk of API pricing. As of this month it's the default direction of the industry. Consider the last six weeks:

On June 1, GitHub moved every Copilot plan onto usage-based AI credits, pegged at a penny per credit, and the Pro plan lost its unlimited premium requests. Cursor restructured its team seats into metered usage pools days later. Anthropic moved agent workloads onto metered credits effective June 15. And OpenAI confirmed that the free period for ChatGPT Workspace Agents ends July 6, when credit pricing begins. A typical agent run will cost 5 to 25 credits, and OpenAI doesn't publish a public dollar price per credit. The teams that spent spring wiring agents into daily operations got 26 days to figure out what those agents will cost.

GitHub's chief product officer explained the shift honestly: a quick chat question and a multi-hour autonomous coding session were costing users the same amount, and absorbing the inference bill behind heavy usage stopped being sustainable. Fair enough. But notice what that means for the buyer. The work you automate most aggressively is exactly the work that now runs up the meter, and per High Alpha's benchmarks, 31% of AI vendors already layer usage charges on top of subscriptions. Salesforce charges $2 per Agentforce conversation. Intercom charges 99 cents per AI resolution. Microsoft prices Copilot for Security at $4 per hour.

What the two leaks do together

Here's where it gets expensive. Sprawl multiplies the number of meters, and metering makes each tool's cost unpredictable. Put the two together and you get a dozen tools, each with its own pricing logic, its own usage curve, and its own ability to reprice mid-year.

The aggregate numbers show the result. Zylo's index puts average enterprise spending on AI-native applications at $1.2 million per organization, up 108% year over year, and 78% of IT leaders report unexpected charges they didn't budget. A separate survey by Benchmarkit and Mavvrik, covered by CIO, found 85% of organizations miss their AI cost forecasts by more than 10%, nearly a quarter miss by 50% or more, and the misses run almost exclusively in one direction: under. More than eight in ten companies said AI costs ate over 6% of gross margin.

There's also a quieter cost that doesn't show up on invoices. When teams know the meter is running, they ration. The Zylo report calls it throttled adoption: usage-based pricing pushes teams to limit the very AI use that was supposed to produce the return. You bought the tool to use it, then the pricing teaches everyone to use it less.

Plugging both leaks at once

The two leaks have the same root: capability got bought one sliver at a time, on other people's infrastructure, at other people's prices. So the fix has two halves that work better together.

Consolidation closes the first leak. Fewer tools means fewer contracts, fewer overlapping subscriptions, fewer security reviews, and real negotiating weight with the vendors that remain. The practical method, including how to map overlap by the job each tool does, is in our companion piece on how to consolidate AI vendors. Consolidation works pre-emptively too. An organization adopting AI now can pick one platform that covers document work, agents, and search from the start, and simply never grow the pile.

The second leak closes when the heaviest workloads move off the meter entirely. A locally hosted AI platform runs on infrastructure inside your environment at a fixed cost, priced by capacity rather than by token. One platform carries the document analysis, the internal search, and the agents your teams build, so it replaces several line items at once. And because the cost is fixed, heavy use costs the same as light use. Scaling up becomes a planning decision instead of a surprise on an invoice, and nobody has to ration the tool that was bought to be used. The honest caveat: whether that pencils out depends on your scale. A company doing occasional AI work won't beat metered pricing with fixed infrastructure. A company where AI runs all day, across departments, almost certainly will, and those are exactly the companies the meters were built for.

Brent Fisher

Co-Founder & Head of Go-to-Market, Cognetryx

Brent writes on private AI deployment, compliance architecture, and the operational gap between enterprise AI adoption and institutional readiness. Cognetryx builds private, locally hosted AI platforms for enterprises.