After three years of rapid innovation and experimentation, enterprise AI is entering a more grounded phase. The defining feature of this year so far is pragmatism: a shift away from demo-driven hype toward cost discipline and measurable value.
AI Companies Become More Pragmatic
The shutdown of high-cost, low-ROI initiatives like OpenAI’s AI video-generation tool Sora underscored a hard truth: Impressive outputs are not enough.
Monetization pressure has intensified. In January, OpenAI introduced ads to ChatGPT, signaling a shift toward sustainable revenue models. Providers are also abandoning flat fees for usage-based pricing: Earlier this year, Anthropic changed its pricing model to charge business customers based on the amount of AI they use. Microsoft made a similar move in June, changing Github Copilot to token-based billing, a change it called a step toward a sustainable business.
The era of growth at any cost appears to be over.
Buyers Are Pushing Back
Pragmatism is also emerging on the buyer side.
Microsoft canceled most of its Claude Code licenses over cost concerns. AT&T has begun limiting employee access to GitHub Copilot for the same reason. Walmart and Meta have capped how much AI their employees can use. Uber's COO said AI costs are getting "harder to justify." After a companywide push to get employees using AI, Tesla capped staff AI spending at $200 a week, with anything above that requiring sign-off. Even companies betting big on AI are shifting from maximizing adoption to managing it.
Organizations are also becoming more sophisticated about how they use AI. Many are routing work to cheaper models based on task complexity and restricting premium models to employees who need them most. The savings can be substantial. Ensemble Health Partners, which expects to spend up to $100 million on AI this year, moved its insurance-appeal-letter tool to a cheaper OpenAI model and projects close to $700,000 in annual savings on that workflow alone. Coinbase’s CEO Brian Armstrong recently said the company cut AI spending “nearly in half,” even as token usage grew, using measures like routing tasks to the right model, caching aggressively, and defaulting to cheaper open-weight models.
The industry's vocabulary is changing with the behavior. After encouraging employees to "tokenmaxx," some companies are now talking about "tokenminimizing" or “tokenminning,” optimizing AI usage to reduce costs without sacrificing results.
This shift toward efficiency is also influencing where investment dollars are flowing. In June, AI memory startup Engram raised $98 million on the claim that its technology can match leading models while using a fraction of the tokens, and it has already signed customers including Microsoft and Notion.
Buyer behavior is shifting in a deeper way too. In a Tegus expert transcript, a CIO and CISO at a North American IT services firm described modular, phased deployment as the preferred enterprise buying model, a change the executive has seen over the last year. Customers “want to prove value with a specific business unit or maybe a specific use case before expanding across the organization.”
The Agent Reality Check
Over the past year, autonomous agents have been positioned as the next major breakthrough in enterprise AI. Yet many organizations are discovering that moving from impressive demonstrations to reliable production systems is not so simple. As a result, enterprises are shifting their focus away from fully autonomous systems and toward narrower, more controlled forms of automation where outcomes are easier to measure and manage.
Deloitte’s 2025 Emerging Technology Trends study notes that while 38% of organizations are piloting agentic AI, only 14% have solutions that are ready to be deployed and just 11% are actively using them in production.
Security is one reason production remains challenging. Agents introduce risks enterprises haven't had to manage before. As a Chief Security Officer noted in a Tegus expert transcript, agents represent a “whole new class of security attacks that didn't exist before.”
Cost is another barrier. According to Goldman Sachs, a single agent task can require 50 times the computing power of a chatbot query. And even as model prices fell by roughly half between late 2024 and late 2025, the volume of tokens consumed grew more than fourfold, by Bain's estimate.
This does not mean agents are failing. Organizations are applying the same scrutiny to agentic AI that they are applying to every other AI investment. Vendors are responding with more focused tools. One example: In June, AlphaSense announced SuperAnalyst, an always-on AI agent purpose-built for financial and strategic workflows.
What Organizations Should Prioritize
Cost discipline is only part of the pragmatism equation. The same buyers moving to phased, prove-it-first deployments also need to get selective about the systems themselves. As AI becomes embedded in core business processes, organizations should weigh systems across five dimensions:
Managing costs deliberately. Lower cost does not automatically mean better value. Organizations should know what they are paying for and ensure AI spending is proportional to the value it creates.
Using the right evidence. As AI gains access to more documents and databases, it becomes more powerful. But access alone does not make a system smarter or more reliable. A dependable system surfaces the right evidence rather than whatever happens to be available. A fluent, confident, well-cited answer that is still wrong costs far more than the tokens used to generate it.
Making claims verifiable. A system worth relying on lets a user trace claims back to their sources and confirm them quickly, which is what separates a tool you can put in front of a client or a regulator from one you can only use privately. In high-stakes work, an answer you can’t check has limited value.
Aligning oversight with risk. The level of human review should match the stakes of the decision. A misrouted support ticket and a misstated financial disclosure carry entirely different risk profiles, and the bar for autonomy should scale with the consequences.
Maintaining architectural flexibility. Model capabilities and prices are shifting too fast to lock into one underlying model. A pragmatic organization keeps its systems model-agnostic, whether by building for flexibility or choosing providers that make it easy to switch models, so it can adopt cheaper, faster, or more specialized options as they become available.
AI’s pragmatism era is ultimately about making better decisions.
The Next Phase of AI
None of this signals the end of AI adoption. This is what maturation looks like.
The first phase of the AI boom was defined by experimentation. This one is defined by accountability. Organizations are asking tougher questions about costs and returns, and vendors are facing greater pressure to monetize.
AI is evolving from an innovation project into an operational technology, and that requires a different approach to how organizations build, buy, and manage it.
Move from research to execution around the clock with SuperAnalyst, AlphaSense's always-on AI agent that runs multi-step workflows on your behalf. Sign up for early access or start your free 2-week trial of AlphaSense today.
