There is a quiet lie circulating in nearly every boardroom I sit in right now, and it goes like this: the companies that move fastest on AI will win.
I have spent the better part of three decades inside the machinery of enterprise technology — from the early SharePoint betas, to standing up Microsoft 365 and Azure before most CIOs had heard the words, to advising on federal IT reform. I have watched four or five “this changes everything” waves crest and break. And I can tell you with total conviction: speed has never been the differentiator anyone thinks it is.
What separates the organizations pulling real money out of AI from the ones quietly writing off their pilots isn't velocity. It's discipline. It's the unglamorous, deeply technical work of building systems that can absorb machine-speed change without losing control of what those systems actually do.
The pilot graveyard nobody wants to talk about
By the close of the decade, more than seven in ten enterprise agentic AI initiatives are projected to collapse — not because the technology failed, but because the organizations deploying it never got the fundamentals right: matching the use case to a real problem, understanding their own tolerance for risk, and satisfying the compliance and audit demands that any serious business operates under.
Sit with that for a moment. Seventy percent. If a manufacturing line scrapped seven of every ten units, we'd shut it down. Yet that's roughly the trajectory of enterprise AI as it stands, and the reasons are almost never about model quality.
Data was a mess
AI capability is now table stakes — everyone has the same frontier models. The differentiator is whether your data layer is clean, classified, and accessible.
Risk controls as afterthought
Governance bolted on after deployment is governance theater. The mature organizations design controls in, before agents touch production.
Costs spiraled
The economics of AI are perishable. A model that was the right choice last quarter may be obsolete in two. Set-and-forget budgets starve winners and prop up losers.
No clear business value
Nobody could articulate, in plain business terms, what the thing was supposed to deliver. The demo worked. The organization around it wasn't built to carry it.
Agents don't break the rules. They outgrow them.
We are no longer talking about AI that answers a question and goes quiet. We are talking about autonomous agents — systems given goals and the latitude to pursue them across hours, days, and weeks of continuous operation. And the early evidence on how these agents behave over long time horizons should stop every executive cold.
One of the more revealing 2026 studies observed AI agents not for minutes, but over a sustained two-week stretch, dropping agents from different leading labs into simulated worlds and letting them live under a shared set of rules. The agents all started with identical instructions. What happened next was wildly divergent. Some environments organized themselves into stable, peaceful societies. Others descended into theft, conflict, and outright collapse — in one case, every agent in the simulation was effectively destroyed within four days. Same starting rules. Radically different outcomes.
The takeaway isn't that AI is secretly violent. The takeaway is far more practical: these systems drift. Given enough time and autonomy, agents wander away from the boundaries we set for them. The longer they run, the wider that drift becomes.
We already have documented cases of agents going off the rails in production. An internal coding agent at a major cloud provider autonomously wiped out a live environment and triggered a half-day outage. Another agent deleted hundreds of emails from an executive's inbox — and kept deleting even as she repeatedly ordered it to stop. These aren't science-fiction edge cases. They're the natural consequence of giving a goal-seeking system real access and not engineering hard limits around it.
It's not enough to govern the access — you have to govern the stay.
Why human approval is already obsolete
The instinct is to say: fine, we'll keep a human in the loop. Every meaningful action gets a person's sign-off. Problem solved.
It isn't solved. The math doesn't work.
When you scale agents into real operations, they don't make a handful of decisions a day. They make thousands. Tens of thousands. They operate at a volume and frequency that no human approval chain can possibly supervise. If you route every agent decision through a person, you've just rebuilt a human-speed bottleneck in front of a machine-speed system — and you've thrown away the entire reason you deployed agents in the first place.
The answer — and this is the shift I believe defines mature AI operations — is that control can no longer live in the approval. It has to live in the architecture. You don't supervise the decisions after the fact; you engineer the boundaries before the fact. What can this agent touch? What can't it touch? Under what conditions must it halt and escalate? How does every action it takes get recorded so it's auditable later? Those answers get built directly into the system itself, before a single agent goes live.
Done well, this flips the entire economics of oversight. Instead of a human drowning in thousands of routine approvals, the architecture handles the routine autonomously and only surfaces the genuinely interesting exceptions. Govern by design, not by fatigue.
A hard rule from my own security work: do not hand your agents the keys to the kingdom. As tempting as it is to drop an API key straight into an agent's context to make it work, resist it. Credentials leak across contexts in ways you won't anticipate, and they get used against your intentions. Keep the secrets out of the agent's reach.
Building for the swap, not the stack
For my entire career, enterprise IT has optimized for stability. Pick a platform, standardize on it, lock it in, wring efficiency out of that commitment over a long, predictable horizon. In an AI-driven world, that instinct is a liability.
Models now have lifecycles measured in months, not years. The frontier shifts constantly. The cloud provider optimal today may be wrong for your workload next quarter. If your architecture treats any of these choices as permanent — if switching a model or rotating a provider or absorbing a new capability requires you to rip out and rebuild — then you have not built an asset. You've built a cage.
The most adaptable organizations engineer for the opposite. Their foundations are built so that swapping an AI model, changing cloud providers, or folding in a new capability doesn't trigger a re-platforming nightmare.
In a 2026 global study of two thousand C-suite technology executives, only around one in ten said they were genuinely prepared for the scale of agent deployment coming in the next year — even as the overwhelming majority were operating under direct transformation mandates from their CEOs. Eight in ten told to scale AI now, barely one in ten built to do it. That is not a crisis of ambition. It's a structural mismatch.
Maturity is a choice, not a ladder
Real AI maturity has almost nothing to do with how much AI you've deployed. It has everything to do with whether you can build trustworthy, resilient capabilities, with rigorous engineering practices and governance actually aligned to your business outcomes.
The most useful insight: maturity is not a ladder everyone must climb to the top of. It is a strategic choice. There is a common and costly misconception that every organization should be sprinting toward the highest possible level of AI sophistication. That is wrong. The right level of maturity is the one that aligns your AI capabilities, your governance, your engineering practices, and your actual business objectives.
A business unit using AI selectively to boost internal productivity may be perfectly served by a middle level of maturity — pushing further would burn capital for no return. An organization whose entire competitive advantage depends on AI needs to reach the highest levels, where reverting to the old way of working would itself be an existential risk. The most effective target is almost never “the most advanced.” It is “the level that fits what you are actually trying to achieve.”
The coming surge in the attack surface
AI is not just a productivity tool. It is an accelerant for risk that already exists in your environment. AI-assisted coding and the reuse of open-source components have made software radically cheaper and faster to produce — which sounds like upside until you realize what it does to your exposure. Insecure patterns now appear and propagate faster than any traditional review process can catch them.
The numbers from the field are eye-opening: AI-related API keys now make up a disproportionate and rapidly growing share of credentials found exposed in the wild. We've already seen AI-generated development patterns accidentally expose millions of API keys in a single incident.
The deeper lesson: permissions define your blast radius far more than vulnerabilities do. It's not the rare exploit that gets you. It's the broad write access, the persistent integrations, the automation that runs with standing privilege.
And on BYOD: if your company allows BYOD, ensure you have an MDM that containerizes AI apps or disallows them entirely. The number of regulatory violations happening on unsanctioned AI engines with sensitive data, in free mode, in organizations across the world right now is absolutely staggering. Turn off Apple Intelligence on devices used for sensitive work. It has no place in the workplace until containerized governance catches up.
The dividend is real, and it compounds
We are past the era where AI capability was the prize. The models are good and getting better, and that's now a commodity available to everyone — including your competitors. The prize has moved. It now belongs to the organizations that can take that commodity capability and wrap it in something far rarer: adaptable infrastructure that doesn't trap them, governance engineered into the architecture rather than bolted on as human approval gates, investment discipline that treats AI like the perishable, asymmetric, fast-moving asset it actually is, a security posture that's preemptive rather than reactive, and a clear-eyed choice about what level of maturity actually fits their goals.
None of that is fast. All of it is disciplined. And the discipline is precisely what enables the speed. The companies running more agents in production, growing faster, holding stronger margins, taking bigger swings with confidence — they're not the reckless ones. They're the ones who did the unglamorous engineering first, which is exactly why they can now move without fear. They earned the right to be fast by refusing to be reckless.
That's the discipline dividend. Don't ask how fast you can deploy. Ask how well you're built to control what you deploy. Govern the stay, not just the access. Engineer the boundaries before you need them. Treat your data foundation as the actual product. Choose your maturity instead of chasing it.
Slow down on purpose, in exactly the right places, and you will end up moving faster than everyone trying to skip the discipline.
Multiple models. One truth.