The first thing you learn when you build AI agents yourself (not just use them, but actually build them), is what they still can’t do.
The second thing you learn is how few executives know the difference.
I run Syndio, a 140-person pay-decision intelligence company serving nearly 400 enterprise customers, including half the Fortune 100. Two years ago, I was using AI the way most executives do: organizing information, drafting emails. It helped at the margins and missed what mattered. The drafts were polished. They sounded nothing like me. More importantly, the AI reasoned nothing like me.
In my early innings with AI, I made all the rookie mistakes. And unfortunately, my mistakes became magnified because I was also modeling for my team. I used AI to create the agendas for my executive-leadership meetings. Without context and direction, those agendas became word salads and AI slop. I knew it could do better.
So I enrolled in a six-week course for executives taught by Nufar Gaspar, a former Intel executive. It didn’t teach us to use AI. It taught us to build systems that could reason alongside us, that remember context, challenge assumptions, and adapt to how we think, not how a product manager imagined we might.
I took the class on the weekends and in the evenings. This is the part that most executives miss. You can’t get good at AI in between meetings. It truly requires the kind of unstructured thinking that most of us wish we had more of. When I was early in my CEO journey at Syndio, a seasoned Microsoft executive named Dean Hachamovitch gave me sage advice. “Protect your thinking time,” he said. I’ve returned to those words often over the years, but they feel especially relevant today. In an era of AI, the scarcest resource isn’t information, it’s the uninterrupted space to think.
The agent build process taught me that most executives are solving the wrong problem.
The assumption is that AI’s value is productivity: go faster, delegate drafting, automate the routine. That’s real, but it’s the least interesting thing these systems can do. The right level isn’t speed. It’s judgment.
Today I use three custom-built agents every day. They are not chatbots. They are systems trained around my workflows, my decisions, my communication style, and the institutional knowledge of my business. The distinction matters more than it sounds.
I first understood why when one of my employees used AI to draft an email on my behalf.
It was competent enough. But the voice was wrong. The framing missed the customer’s actual concern.
My own agent, however, compared the draft against a writing rubric built from years of my sent emails: no em dashes, ever. No “it’s not x, it’s y.” Reflect the other person’s idea back before pitching. Customize the sign-off based on the relationship.
The agent restructured the message entirely. It referenced a concern from a conversation six weeks earlier, cut several paragraphs of unnecessary setup, and changed the opening to focus on what the customer actually cared about.
An AI caught what another AI had missed. The reason: mine understood how I communicate. The other one understood how people generally communicate.
Meet My New Staff #
The most useful agent I’ve built is a strategic advisor, less a tool than a thinking partner. Before major decisions, before board updates, before I bring anything to my leadership team, I go here first. The agent has context about our business, competitors, roadmap, and past strategic decisions. Its value isn’t that it retrieves information. It’s that it pushes back.
I adapted a prompting framework called “grill me,” originally created by developer Matt Pocock. The agent interrogates the logic behind a decision one question at a time: What evidence supports this? What assumptions are you making? What would an investor challenge here?
The Rise of the AI Memory Layer #
The second agent functions like a chief of staff.
Every morning, it sorts my inbox into four buckets: urgent, needs response, FYI, and ignore. Before meetings, it pulls context from past conversations. It drafts follow-up emails in my voice, assembles agendas, surfaces buried Slack messages, and prepares Monday-morning briefings.
What surprised me was how much the quality depended on what I fed it. Every customer conversation, strategic decision, and deal note gets logged in a way my agents can retrieve later. Ingesting transcripts and emails is only part of it. Giving the agent qualitative observations after a conversation, the tone, body language, off-the-cuff reactions, makes it more accurate than systems-of-record data alone. That’s not something I understood going in. It’s something I learned by getting it wrong first.
Board preparation changed most visibly. I built profiles for each board member from past meeting transcripts, public interviews, investment theses, and prior conversation notes. Before a meeting, I’ve already pressure-tested the conversation from each person’s likely perspective: One wants data. One watches leading indicators. One evaluates everything through long-term positioning. I don’t walk in reacting anymore. I walk in having already had the argument.
AI Isn’t Replacing My Judgment. It’s Strengthening It. #
The biggest impact isn’t productivity. It’s judgment.
These systems save time. But more importantly, they catch weak logic before it becomes a public mistake, free up cognitive load, and ensure I rarely walk into an important conversation without full context.
The cumulative effect: I operate with more context, more clarity, and better pattern recognition than I could alone. That’s a different category of advantage than productivity.
The New Executive Skill #
I’ve committed to sending 20 employees through the same program, not to turn them into engineers, but because companies now need people who can recognize a new capability, understand what it makes possible, and let go of how things used to work. Every week they meet to share what they’re building, what’s failing, and what they’re learning.
The most consistent observation: people become less intimidated once they’re building. The technology stops feeling abstract. It starts revealing its own limits, which is the thing you actually need to know.
The underlying models are improving faster than most executive timelines assume. Costs are falling. Capabilities that felt experimental six months ago are now usable.
A senior engineering candidate told me recently that I was the first CEO he’d ever interviewed who was actively building AI agents herself. Most executives, he said, talked about outcomes and urgency but didn’t understand the mechanics well enough to know where the real friction lived.
That disconnect is a liability. The tools are accessible enough that any executive willing to invest a few weekends can build something genuinely useful, not a demo, not a proof of concept, but a system that understands your business, remembers context, and operates at the level of judgment, not just execution.I started with a blank file and a six-week class. The first thing you learn is what AI still can’t do. The second is how much it already can. Too many executives are betting their companies on assumptions about both.
The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.