Companies are beginning to worry about rising AI costs and proving ROI with real use cases, but that anxiety isn’t leading them to pull back on AI spending. Notifications about hitting usage limits and company announcements about “thoughtful” AI model use are signs of a shift from AI experimentation and “tokenmaxxing” to greater cost scrutiny.
That shift matters because companies have already begun using AI to replace software. Earlier this year, we asked 1,000+ builders if their orgs had replaced a SaaS tool with AI, and 35% reported they had. But what happens to those tools when AI costs go up, as many increasingly expect they will?
We surveyed 101 technical leaders at enterprise (1,000+ employees) and mid-market (200–999 employees) companies and found that 43% are already over their AI budgets, but they’re still willing expand AI investments that drive value (even if only 19% would absorb the cost of a hypothetical price hike).
- Nearly half of these leaders are already over budget. 43% report spending more on AI than they planned for in 2026, compared to 34% who are roughly on track.
- Cost anxiety hasn’t slowed adoption down. 43% of leaders are pushing for broad AI adoption wherever it adds value, while 32% are selectively scaling toward high-ROI use cases.
- Vendor consolidation is coming. Leaders are bracing for higher AI costs, but few think the ROI justifies absorbing them outright. Only 19% would absorb a 20% cost increase; the most popular response (32%) is consolidating to fewer vendors.
- Most 2026 AI budgets aren’t in the millions. 52% of leaders budgeted between $100K and $1M for AI tools this year. But 24% have crossed the $1M mark, a large sum considering the unclear ROI from the tech.
- Vibe coding is adding to AI spend, but it’s not the primary driver. 52% say vibe-coded apps are a moderate-to-major contributor to AI spend.
Almost half of leaders (43%) reported overspending on AI. Just 24% are coming up under budget, and about one-third are roughly on track:
Despite the headlines of sky-high AI bills coming as a surprise, earlier this year Gartner reported that token costs are expected to fall. But the decline in per-token price won’t happen fast enough to offset the token volume per task as more businesses adopt AI.
Poor visibility into AI systems is the underlying cause of AI spending that feels out of control. Just 5% of CIOs, CISOs, and CTOs are confident in their visibility into AI-generated tools running in production, according to Retool’s State of AI Governance report. Shadow AI is more likely to incur surprise bills than AI that’s known and governed.
When it comes down to the AI projects themselves, some cost more than others. LLM chatbots are less expensive. Agentic workflows use ~5-30x more tokens per task than asking a question to an LLM, according to Gartner, and agentic use cases are assumed to be the biggest driver of AI ROI as they replace traditional SaaS workflows and reduce the employers’ need for headcount.
Early signals on how leaders would respond to a steep AI price hike prioritize implementing guardrails around access and usage over encouraging or mandating behavioral changes on the individual level. If faced with a hypothetical 20% spike in AI costs, about one-third of leaders (32%) would respond by consolidating to fewer vendors or tools, and 24% would audit and cut AI usage with unclear ties to impact.
Each of the responses leaders picked for a hypothetical 20% cost spike—consolidating vendors, auditing usage, or absorbing costs—is a tokenomics decision. Tokenomics (token + economics) has become the shorthand for the understanding of how AI tokens are acquired, consumed, and attributed to business outcomes. It involves constant rethinking, evaluating, and adjusting as models improve and change, and requires close visibility into usage.
Model routing—the practice of directing each inference request to the most appropriate model for the task instead of defaulting to an expensive model each time—is gaining ground as businesses expect to spend more on AI. The latest frontier models are powerful, but not necessarily suited for basic tasks like summarizing call notes and searching documents. Companies like Databricks and Stripe are both making bets that businesses will eventually need this level of visibility and control into how their teams are using AI, all the way down to the models they can access.
“We know which groups and teams need to develop agents, harnesses, or vibe coded apps, and we give them token allocations and licenses to execute on their work,” says one enterprise VP/CISO. “We then check and validate usage and determine if we need to cut off budget or add.”
Vendor consolidation is already playing out at scale. One survey from ADAPT found 68% of technology leaders planned to consolidate vendors in 2026. Here, the tokenomics prioritize eliminating vendor sprawl and limiting where token spend can go unaccounted for. It’s easier to manage tokens in one platform than trying to manage across multiple. Consolidating tools plus model routing is one viable path forward for businesses to stop rogue AI-related spending.
Most leaders surveyed hadn’t committed to full seven-figure budgets for AI in 2026. About half (53%) budgeted between $100K and $1M, and 24% have crossed the $1M budget mark.
Smaller budgets are easier to outspend, but not everyone will be like Uber, which famously burned through its entire 2026 AI budget in four months. The pace that AI is changing also means budgeting with imperfect information about costs and adoption.
Cost anxiety hasn’t seemed to slow adoption. Only 8% of respondents show hesitation with regards to AI adoption. The rest are either pushing maximum adoption, selectively scaling, or are actively experimenting:
“There is a corporate level fear of missing the boat on AI,” says one engineering director at an enterprise organization. “This is driving a lot of behavior that is not following normal cost controls and scrutiny. This will ultimately result in a reset and more targeted approach.”
That split mirrors a broader pattern across the industry: breadth of AI adoption is outpacing depth. WRITER’s 2026 enterprise survey found 79% of companies still face adoption challenges despite heavy investment, and the leaders actually seeing ROI are disproportionately the ones with cost visibility and governance already in place—the same dynamic behind the 7%-with-established-ROI figure from KPMG. Pushing for broad adoption before closing that gap is a bet that speed pays off faster than governance catches up.
Vibe coding—the practice of building software with natural language prompts instead of code—is one of the most visible AI use cases. As of now, it’s not a significant cost driver. But it relies on tools that come with costs that sit just out of view, and are easy to miss if AI costs are to go up. 52% say vibe-coded apps are a moderate-to-major contributor to AI spend
The “SaaSpocalypse” of early 2026—the roughly $1 trillion selloff in software stocks, driven by the prospect of vibe-coded tools replacing off-the-shelf SaaS—started when it seemed like building your own tool would be faster and cheaper than the tools in your business’s stack. An afternoon with a Claude Code Pro subscription looks like a cheaper alternative to a long procurement and evaluation process—that is, until the models get more expensive, or until builders across the company are all spending tokens on their own one-off tools.
“As a CISO, I would definitely try to veto any actual AI-gen or vibe coded apps,” says one enterprise executive. “We can’t afford mess ups from that perspective.”
A few factors could make vibe-coded apps more expensive than they appear on the surface:
- Pricing model. Base-plan, usage-based, or effort-based, often some mix. Most vibe coding tools charge for tokens consumed at build time and tokens the deployed app burns in production—two separate meters that are easy to conflate when you’re evaluating a tool by its base plan.
- Model choice. A 25-person Cursor team pays $40/seat/month base, but agentic sessions on top can add $600–$6,000/month depending on the model according to Vantage.
- Hosting once live. The Register documented $30K–38K and $127K bills from usage-based AI billing on AWS and Google Cloud alone.
- Security left to the builder. Veracode found 45% of AI-generated code ships with an OWASP Top 10 vulnerability; data breaches from shadow AI-related security incidents cost $670K more than the average for a security breach according to IBM’s 2025 Cost of a Data Breach report.
Vibe coding tools usually offer base plan pricing, usage-based pricing, effort-based pricing, or some combination. This includes tokens processed while you build and tokens your app uses. Vibe-coded apps also put the onus on the builder to secure the app, either through adding governance features themselves or through third-party security and governance tools.
The leaders in this survey aren’t pulling back on AI, even though 43% are already over budget. The looming threat of rising costs so far isn’t enough for leaders to cut back or consolidate AI investments, even if the ROI isn’t fully legible yet. But unlike the first three years of the AI boom, investment in AI is starting to come with conditions to make that spending governable and secure. Leaders want to know what’s running and who’s using it, and decide deliberately whether to cut or expand.
Most organizations don’t have that visibility. Just 5% of CIOs, CISOs, and CTOs report confidence in their visibility into AI-generated tools running in production, according to the 2026 State of AI Governance report. That gap brings surprise bills and costly security incidents. The next step for leaders will be to close that gap with the right infrastructure and platform for their vibe-coded tools.
This report is based on a survey of 101 senior technology and security leaders, conducted in partnership with Wynter in July 2026. Respondents are senior IT leaders: security directors, cybersecurity managers, VPs of IT, and similar roles with equivalent scope. 52% are C-Suite, VP, or Director-level respondents.
Companies are split between Enterprise (1,000+ employees, 70%) and Mid-Market (200–999 employees, 29%). Industries represented include healthcare, SaaS and software, financial services, IT services, professional services, government, and manufacturing.
Data analysis: Dustin Drees + Wynter
Illustration: Chris Sandlin
Content: Kasey Hickey, Maddie Latoche, Matthew Komorowski, Jordan Edelstein
