AI Agents have been the rage for some time now – at least if you listen to the press releases and the endless news articles. There were of course many developments over the past few weeks; let’s look at the progress and take stock of where we are.

Agents and Enterprise Knowledge

I’ve been saying forever that the key to getting AI to work in the enterprise is by giving those agents the internal knowledge of the enterprise. If there was any doubt about that, it should be gone now, because:

Model Provider #1: OpenAI

OpenAI released “Company Knowledge” in ChatGPT, a way to connect ChatGPT for Enterprise to your internal systems that contain your organization’s knowledge. I’ve been saying this is a necessary and obvious step. It tells us that:

• OpenAI wants to secure a foothold in business, not just in consumer applications (they need the money)
• They know they can’t do that with models alone; the models need internal company knowledge, knowledge that isn’t public and therefore can’t be trained in
• The way to do that is to connect to corporate systems and information repositories, so that knowledge can be provided to the model – primarily through RAG (Retrieval Augmented Generation)

“The context you need to get work done often lives in your internal tools: docs, files, messages, emails, tickets, and project trackers. Those tools don’t always connect to each other, and the most accurate answer is often spread across them.”
– OpenAI

But OpenAI isn’t the only one. Far from it. All the big players are doing it:

Model Provider #2: Anthropic

Anthropic rolled out a Microsoft 365 connector for Claude, making it capable of the exact same thing for Microsoft repositories. So, the same can be said for Anthropic: they want enterprise revenue, they know their models need to know more than the public data they can train on, so they have to connect to enterprise repositories and use RAG (which is powered by search) to ground the models in company data.

“Claude now integrates with Microsoft 365 and provides enterprise search across your connected tools.”
– Anthropic

Model Provider #3: Google

These two announcements came shortly after Google did the same thing when they introduced Gemini Enterprise in early October. It allows Gemini to work with company knowledge through – you guessed it – RAG.

Google also expanded their Deep Research tool – previously only available on the web – to include your own personal content. For now it’s only your own Gmail, Google Docs, and Google Chat content. But you can expect this to show up in Gemini Enterprise soon, for deep research on your company’s internal content.

And Perplexity Too

All of these model providers (except Google) are in desperate need of revenue, and so is Perplexity even though they’re using models instead of training them. To get in on the enterprise game they released a report on how to use AI at work, positioning Perplexity (and the Comet browser) as a unified AI platform. Although a little self-effacing (of course) the three principles they offer are useful ways to think about how to use AI:

• Block distractions (by delegating repetitive or low-value tasks)
• Scale yourself (by multiplying your talent with the strengths of AI)
• Get results (by using your additional capacity to focus on the things that matter)

Agents in the Enterprise

Meanwhile, we continue to get information about how AI agents are faring at companies.

McKinsey released The State of AI in 2025. It includes survey responses from 1,993 participants in large global enterprises. That’s a pretty good cross-section, although the survey was in July so it’s a few months old. The main takeaways?

• Everyone is trying to deploy agentic AI.
• Some are succeeding. “23% of respondents report their organizations are scaling an agentic AI system somewhere in their enterprises.”
• Impact is still narrow. “Meaningful enterprise-wide bottom-line impact form the use of AI continues to be rare.”

Wharton released their annual report on AI industry adoption, an 85-slide report based on how leaders are using generative AI and their perceptions. Here’s what I found interesting:

• About 75% of leaders see positive returns (probably inflated, since it was about the perception among colleagues, not hard metrics).
• Cost is still not a major concern. What are the main barriers to adoption? Security, operational complexity, accuracy, customer data privacy and ethical considerations are all ranked higher than cost.
• 89% of leaders felt that generative AI enhances skills and 71% say it replaces skills. Although it’s the latter that gets more headlines, it’s not either-or, it’s yes-and.
• At the same time, 43% of those surveyed feel that gen AI use is causing skill atrophy.

If you want to go deep on this and the famous “95% of pilots fail” MIT report to understand why they may seem contradictory, I recommend this article from Alberto Romero’s blog.

Taken together, these reports tell a different story than each does separately: adoption is real, but results are mixed; both sides are, to some degree, cherry-picking timelines and definitions and targets and statistics…You can’t fully trust either report, or rather, what people share on social media about either report.
– Alberto Romero, Why You Can’t Trust Most AI Studies

---

---

We usually think AI adoption and the value it generates looks something like a bell curve, a normal distribution:

If you look past the clickbait headlines and connect the dots between McKinsey’s The State of AI in 2025 and other research over the past few months (McKinsey, MIT, BCG, Deloitte), a clear picture emerges. One that’s different from what we’re used to seeing. We see not one bell curve, but two separate groups of companies with very different results. Most companies are, in fact, part of a bell curve, with nominal benefit. But the organizations that are leading the charge for AI are seeing dramatically higher impacts. It looks something like this:

(I first saw a chart like this from GAI Insights, at their GAI World conference, which I recommend you attend next year!)

So in summary:

• It’s not a bell curve; benefits are not distributed evenly
• The average company is not seeing big returns
• But the few who are aggressive with AI are seeing outsized value
• The gap between the majority and the leaders appears to be widening

“Organizations with ambitious AI agendas are seeing the most benefit.”
– McKinsey, The State of AI in 2025

If this trend continues, then at some point, the gap between the leading companies and majority may be too great to cross. The companies that aren’t aggressive with AI won’t ever be able to catch up.

My tips so you don’t get left behind:

• Don’t wait, get started now
• Choose the right use case
  • Aligned: choose one that AI can help, something that AI is good at
  • High-Value: choose one that matters
  • Focused: choose one that is narrow in scope, with defined outcomes
• Specialize: generic AI will give generic results. Verticalize, specialize, and tailor the solution to your domain, your business, your problem.
• Be rigorous: don’t just throw AI at it. Be disciplined in your approach.
• Build on a solid foundation: you need advanced RAG on your enterprise content

And finally, don’t go it alone. Choose the right partner.

Pilots built via strategic partnerships were twice as likely to reach full deployment as those built internally. – MIT, State of AI in Business 2025