If It’s So Good, Why the Hesitation?
AI is everywhere right now. It writes code, summarizes meetings, analyzes data, drafts content, automates workflows, and increasingly takes action on our behalf. It even helps people write blog articles, which I’m told is becoming something of a trend. So, if AI is really that good, why are so many enterprises still hesitant to use it everywhere?
 |
| AI comes at a significant cost |
Large organizations tend to be cautious for a reason. They have more systems, more data, more employees, more customers, and often more regulation to deal with. A tool that saves one person a few hours a week can look very different when you start thinking about what happens when thousands of people use it every day.
That is where the conversation around AI gets more interesting.
The Downside Still Counts
This article actually started with a tweet from someone who was amazed that there are still enterprises with thousands of developers writing code "by hand" because AI access is restricted or unavailable. I understood the reaction, but it also struck me as a perfect example of how different the equation looks inside a large company.
One reason large companies hesitate is simple: the downside still matters.
I’ve written before that AI can dramatically raise the bar for what people are able to accomplish without removing the need for human judgment. That same idea applies here: a developer can write code faster; an analyst can get to an answer more quickly; a support team can automate more of the work that used to require manual effort. Those gains are real, but somebody still has to understand when the output is wrong, incomplete, or unsafe.
That matters a lot more in a regulated environment. If AI-generated code introduces a flaw that leads to a HIPAA, PCI-DSS, GDPR, or similar compliance problem, the cost is not limited to fixing the code. There can be remediation work, legal exposure, penalties, customer impact, lost revenue, and damage to the company’s reputation. The original productivity gain can become irrelevant very quickly.
This is also where organizational readiness comes into play. How much of that risk an organization actually takes on depends in large part on how prepared it is to use AI responsibly. A recent CIO Dive article pointed to Forrester research showing that only 16% of information workers had a high understanding of AI tools, while 43% were at risk of misunderstanding them. Fewer than half knew when they should question AI output.
That is a meaningful problem. If people do not know when to trust AI, when to challenge it, what data they should expose, or where human review is still required, then pushing adoption faster simply increases operational risk.
There is a lot of pressure right now to move quickly because everyone else seems to be moving quickly, and here FOMO can become very dangerous. An organization that adopts AI faster than it can put the right controls, governance, and operating practices around it may end up creating more exposure than value.
For some companies, hesitation is not resistance to AI. It is a sign that they are still trying to make sure they are ready for it.
The Bill Still Has to Be Paid
Even when an organization is comfortable with the risk, the economics can still get complicated very quickly.
There are some spectacular examples of AI lowering costs. Yahoo Finance reported recently that Curative replaced a $600,000-a-year Salesforce CRM contract with an internally developed system that was built largely through AI-assisted 'vibe coding' in about two months. The company is also trying to cut roughly 80% of its SaaS spending this year and redirect some of that money toward AI.
That sounds like exactly the kind of story that makes the case for using AI everywhere. The problem is that Curative’s AI spending has also been growing at an extraordinary rate. Its CEO said the company’s Anthropic usage costs had increased sixfold every month over a period of six or seven months, reaching millions of dollars per month. He still believes the economics work for Curative, but he also acknowledged that the rate of spending eventually has to slow down.
Curative is hardly alone. Uber reportedly burned through its entire 2026 AI coding budget by April. TechCrunch also reported companies finding themselves several times over their annual AI usage budgets, with much of that usage measured in units called tokens, only a few months into the year, while Priceline saw a routine renewal for an AI coding tool, Cursor, come back four to five times more expensive.
The problem gets harder as AI becomes more capable. Better models encourage people to use them more often, and increasingly autonomous agents can consume far more tokens than a simple question-and-answer interaction. The aforementioned CIO Dive article cited Gartner research showing that some complex workloads in which AI agents perform multiple steps or take actions on their own (this is referred to as agentic workloads) can consume three to five times as many tokens per query.
This creates an unusual economic problem. AI can absolutely save money. It can eliminate software licenses, automate work, and let people accomplish things that previously required much larger teams. At the same time, greater success can produce greater consumption, and greater consumption can produce a much larger bill than anyone originally expected.
So the question eventually changes from "What can we do with AI?" to "How much AI can we afford to use?"
Is Any of This Actually Paying Off?
Even if an organization can manage the risk and absorb the cost, there is still one more question to answer: is the AI actually creating enough value to justify the investment?
That sounds obvious, but it is becoming a real problem. Companies have spent the last few years experimenting with AI assistants, more autonomous agents, internal tools, and new AI features because the technology was moving too quickly to ignore. In many cases, the first goal was simply to get started. Now the conversation is shifting toward what those investments are actually producing.
Business Insider recently reported that EY created an AI Value Realization Office for exactly that reason. Its job is to look across AI initiatives, track usage and spending, decide what should scale, and make sure the company can connect AI investment to measurable business outcomes.
That is a pretty strong signal. EY is not questioning whether AI is useful. It is asking a more mature question about where the value is actually showing up. And the question gets even more interesting when you look beyond an individual company.
Fidelity, for example, issued a warning - as reported by The Street - about AI agents in financial services. More automation, more transactions, and more activity do not necessarily mean the company deploying the most AI captures the most value. In some cases, the value may accrue to the firms that already control the scarce things around the technology, such as trusted data, distribution, liquidity, customer relationships, or regulatory standing. In other words, making something easier or cheaper with AI does not necessarily mean the company using the AI is the one that captures the economic benefit.
That is an important distinction. It is easy to measure how much AI is being used. It is harder to measure whether that usage is improving revenue, reducing cost, helping the same workforce accomplish more, or creating some other meaningful business outcome.
At some point, "we are using more AI" stops being a strategy. The real question is whether the business is better because of it.
So Why the Hesitation?
The contradiction is not really much of a contradiction after all.
AI can be genuinely transformational and still be something enterprises approach carefully. The larger the organization, the more variables there are to consider before a promising experiment becomes something used broadly across the business.
That comes down to three basic questions.
- Can we control it?
- Can we afford it?
- Is it actually worth it?
If the answer to all three is yes, then the case for scaling AI becomes much easier to make. If even one of them is uncertain, hesitation starts to look a lot more reasonable. That does not make AI less important. It just means that adoption is a business decision, not a popularity contest.