AI is starting to do the work. Most enterprises are not wired for it yet.

02 What the leaders already get

The companies that got this right have the numbers to show for it.

These are results companies have published themselves: on their engineering blogs, in earnings calls and in filings.

The clearest view of where this goes comes from inside an AI lab. Anthropic’s filing to list its shares, and its own research on its workforce, describe a company that now runs on agents.

Sources: Anthropic PBC, Form S-1 (October 2026); Anthropic, “How employment is changing at Anthropic” (October 2026).

Agent work overtook human work in about three months. Output per person rose in every function, not only engineering.

The work itself changed too. Engineers who adopted agents wrote far more code with no measurable drop in quality. The share of junior technical hires fell from about 28% to 9%, and hiring scorecards now ask about AI tools three to four times as often. Skills are moving to people who can direct agents and judge their output. Helping your teams make that move is part of our job.

03 What AI can already do

The intelligence has arrived. It keeps doubling.

In 2023 the best AI model could reliably finish a software task that takes a skilled person about four minutes. By early 2026 the best model finished tasks that take a person twelve hours, about half the time. The length of task AI can complete has doubled about every four months.

It is not only code. OpenAI’s GDPval benchmark gives models real assignments from 44 occupations in the nine largest sectors of the US economy, each about seven hours of expert work. Industry experts judge the results blind. In under two years, the share of tasks where AI matched or beat the expert went from about one in eight to more than four in five.

What AI could do and what it is used for are still far apart. Anthropic compared the share of each job family’s tasks that current models could handle with the share it actually sees in use. In business and finance, models could handle almost every task. They are used for under a third.

Where it is used, it already covers most of some jobs. In Anthropic’s data, AI already appears in three quarters of a computer programmer’s tasks and seven in ten of a customer service representative’s. With AI, the average task in those conversations was done about 80% faster.

Controlled studies agree that the gains are real, largest for less experienced people, and not automatic. Experienced developers working in code they knew well were slower with AI in one study, and more code does not always mean more shipped software.

That is the pattern we see in practice. The value comes from redesigning the work around what AI does well and checking what it produces, not from handing people a tool.

04 The gap

Most enterprises are nowhere close.

Outside the frontier, the picture is very different. Every two weeks, the US Census Bureau asks firms whether they used AI. Even after it widened the question to cover any business function, only about one US firm in four says yes.

The gap between companies is widening. OpenAI’s enterprise data shows the top tenth of firms producing about eight times as much AI output per employee as typical firms. Boston Consulting Group finds that only 5% of companies get value from AI at scale, while 60% get hardly any. Those few leaders pull ahead on every financial measure.

And spend is still tiny. Among Fortune 500 companies that pay Anthropic more than $100,000 a year, that spend is on average less than 0.1% of their total expenses. The problem is not access to the model. It is everything between the model and the work: your data, your permissions, your processes and your people.

05 Where the value is

The money is in the work, not the software.

For every dollar companies spend on software, they spend about six on services: people doing work, in house or outsourced. That is where AI agents create the most value, because they can take over the work itself instead of selling another tool to the people who do it.

The pools are large. The work that is already outsourced and mostly rule-based, such as accounting, claims and revenue cycle, is where agents can start today. The McKinsey Global Institute estimates that today’s technology could in principle automate about 57% of US work hours.

06 The risk

The same intelligence works for attackers.

Look again at the drawing at the top of this page. Every block is something an attacker can reach, and over 90% of codebases contain components four or more years out of date. Attackers now have AI to find the weak blocks.

They are much faster. Mandiant tracks the average time from a patch being released to attackers exploiting the flaw. It fell from 63 days to 5 days in five years. For 2025, Google estimates minus 7 days: on average, exploitation now starts before the patch exists.

And there is more to find. After a steady 2025, the number of vulnerabilities disclosed each month doubled in 2026. Exploitation in the wild rose sharply with it, and most of those flaws had no patch when they were attacked.

The flaws AI finds are also the dangerous kind. Half of AI-discovered vulnerabilities let an attacker run code on the target, against about a quarter of all others.

October 2026 showed what this looks like. Two critical flaws in PHP, the language behind about 69% of websites whose server language is known, were found with the help of an AI model. Exploit attempts started the day the patches shipped.

The losses are already real. Cybercrime losses reported to the FBI have grown five-fold since 2020, past $20 billion a year, and card chargebacks are forecast to keep rising. Meanwhile, the agents companies deploy are a new target: in IBM’s study, 97% of organizations that had an AI-related breach lacked proper AI access controls.

07 How we help

We close the gap with you.

The leaders did not get there with a better model. They connected AI to their data, rebuilt processes around it, secured it, and trained their people. That is the work we do for you, delivered by the same forward-deployed engineers on the same foundation.

Also live: OpenNash CX, an AI customer support team for phone, text, email and web chat.

08 How we work

Build for what will not change.

“I very rarely get asked ‘What’s not going to change in the next five to ten years?’ That second question is actually the more important of the two.”

Jeff Bezos, 2012

Models will keep getting better. The work will still have to be right, the answer will still live inside your company, and someone will still own the result. So we work like this:

Further reading