Your AI Bill Is About to Stop Making Sense. What CFOs and PMs Need to Know.
Somewhere inside your company right now, an employee just put a $20/month ChatGPT subscription on a personal credit card. Your procurement team doesn't know. Your IT department can't see it. And the invoice that lands on someone's desk next quarter won't look like any software bill your finance team has seen before.
This isn't a hypothetical. According to Zylo's 2026 SaaS Management Index, released this month, ChatGPT is now the most expensed application in enterprise software. Expense-based SaaS (tools bought by employees outside procurement) surged 267% year-over-year. And AI-native application spending at organizations with 10,000+ employees jumped 393% in a single year.
The numbers get worse from there: 78% of IT leaders reported unexpected charges tied to AI or consumption-based pricing in the past 12 months. 61% were forced to cut other projects to cover the overruns. Business units now control 81% of SaaS spend, while IT directly manages just 15%.
If you manage a budget, approve software purchases, or run a team that uses AI tools, this is your problem. And the root cause isn't overspending. It's that AI pricing works nothing like the software billing model your organization was built to handle.
Why AI Bills Don't Work Like SaaS Bills
Traditional SaaS pricing is simple: you pay per seat, per month. Ten users of Slack cost ten times what one user costs. Finance teams can forecast this. Procurement can negotiate volume discounts. Renewals are predictable.
AI pricing breaks all three assumptions.
When your team uses an AI tool, costs scale with consumption, not headcount. The billing units are tokens, API calls, inference cycles, or "credits" (every vendor has invented its own currency). Salesforce charges Agentforce customers in Flex Credits at $0.10 per action. Microsoft prices Copilot Studio in packs of 25,000 Copilot Credits at $200/month. Adobe uses generative credits. None of these units map neatly to a per-user cost.
The deeper problem: usage varies wildly depending on what people actually do. A marketing team experimenting with AI-generated copy might burn through 10x more tokens in a brainstorming week than during a quiet month. A single prompt engineering change can double or halve your costs overnight. An engineer who starts running agentic workflows can consume in one afternoon what a casual user consumes in a year.
As Greg Gorman, VP of Product Management at North, put it: "Gone are the days where you can have a great product and a great service, and your invoices aren't any good."
AI invoices arrive as dense ledgers of token counts, model tiers, and throughput metrics. They look more like utility bills than software subscriptions. And most finance teams have no framework for mapping those charges back to specific business activities.
A Forecasting Framework That Actually Works
If you're about to scale an AI use case, don't approve the budget until you've answered four questions:
1. What's the unit of work? Before anything else, define what "one unit of work" means for your specific use case. Is it one customer service resolution? One document summarized? One code review? This is the denominator that makes everything else calculable.
2. What does one unit cost today? Run your use case for two weeks at small scale and measure actual consumption. Don't rely on vendor pricing calculators; they're built around ideal scenarios. Track the real tokens, credits, or API calls per unit of work across different users and different days.
3. What's the usage ceiling? Estimate your maximum monthly volume and multiply by the per-unit cost. Then add 40%. Consumption-based pricing consistently surprises on the upside because usage grows faster than organizations expect, especially once a tool proves useful and adoption spreads organically.
4. What triggers a cost review? Set explicit thresholds. If monthly spend exceeds your forecast by 20%, that should trigger an automatic review, not wait for someone to notice at quarter-end. Most cloud providers and AI vendors offer usage alerts; turn them on from day one.
Five Governance Questions for Every AI Vendor Contract
Procurement teams need a new checklist. Before signing or renewing any AI-powered tool, ask:
- What is the billing unit, and how does it correlate to business activity? If the vendor can't explain how token counts translate to tasks your team actually performs, that's a red flag.
- Is there a spend cap or circuit breaker? Can you set hard limits that pause the service rather than generate overages? If not, negotiate one into the contract.
- What happens when the vendor changes model tiers or pricing mid-contract? An increasing number of AI vendors combine subscriptions with usage-based charges, and the ratio can shift without warning. Get change-notification terms in writing.
- Can you access raw usage data via API? You need to pipe consumption data into your own dashboards. If the vendor only offers a proprietary portal, you can't correlate spending with business outcomes.
- Who owns the audit trail? When finance asks why last month's bill was 3x higher than expected, you need granular logs that tie costs to specific teams, projects, and workflows.
Building a Monitoring Layer
The 40% buffer in your forecast buys time, but it's not a strategy. What you actually need is a monitoring layer that sits between your AI tools and your finance reporting.
At minimum, this means three things: a centralized inventory of every AI tool in use (including the ones employees are expensing), usage dashboards that update weekly rather than at invoice time, and automated alerts when any tool crosses its spending threshold.
Zylo's data shows that the average organization spends $55.7 million annually on SaaS, with 36% of licenses sitting unused. Large enterprises add an average of 21 new applications per month. If you can't see it, you can't manage it, and right now most organizations can't see their AI spending until the bill arrives.
Ben Pippenger, Zylo's co-founder, called it directly: "AI is quickly becoming the most expensive 'invisible worker' in the organization." This is exactly why model pricing matters more than it used to: Doubao 2.0's tiered pricing structure shows what a 90% cost reduction looks like in practice for teams running agent loops at scale.
The Bigger Question: Is Total Spend Going Up, or Just Moving?
Here's the strategic question that most cost-management articles skip. As AI replaces tasks that used to require per-seat SaaS licenses, does the total software budget grow, shrink, or stay flat?
The honest answer: it depends on how fast you consolidate.
If an AI coding assistant lets you cut five developer tool seats, that's a savings. But if that same assistant runs on consumption-based pricing and gets adopted by 50 engineers who each use it differently, you may spend more than you saved. The math only works if someone is actively managing the trade-off.
Gartner forecasts worldwide AI spending will hit $2.52 trillion in 2026, a 44% increase year-over-year. That's not replacing existing IT budgets. That's net new spending, driven largely by infrastructure costs that most enterprises haven't fully accounted for yet.
The organizations that will manage this well aren't the ones with the best AI tools. They're the ones that treat AI spending as a new cost category with its own forecasting model, its own governance framework, and its own monitoring infrastructure, rather than trying to shove consumption-based billing into a spreadsheet built for seat licenses.
Start with visibility. Audit every AI tool your organization uses today, including the ones on employee credit cards. Set per-tool and per-team spending thresholds before the next budget cycle. And stop signing AI contracts that don't include circuit breakers.
The bill is coming. The only question is whether you'll understand it when it arrives.
